Technological advances have lead to process-driven automated equipment that is increasingly complex. A tool system to accomplish a specific goal or perform a specific, highly technical process can commonly incorporate multiple functional elements to accomplish the goal or successfully execute the process, and various sensors that collect data to monitor the operation of the equipment. Such automated equipment can generate a large volume of data. Data can include substantial information related to a product or service performed as a part of the specific task, but it can also comprise sizable log information related to the execution of the process itself.
While modern electronic storage technologies can afford retaining constantly increasing quantities of data, utilization of the accumulated data remains far from optimal. Examination and interpretation of collected information generally requires human intervention, and while advances in computing power such as multiple-core processors, massively parallel platforms and processor grids, as well as advances in computing paradigms like object-oriented programming, modular code reuse, web based applications and more recently quantum computing, the processing of the collected data remains to be a non-autonomous, static programmatic enterprise wherein the data is operated upon. More importantly, in non-autonomous data processing, the data fails to drive the analysis process itself. As a consequence of such data processing paradigm, much of the rich relationships that can be present among data generated in automated equipment during a highly technical process can be unnoticed unless a specific analysis is designed and focused on a specific type of relationship. More importantly, emergent phenomena that can originate from multiple correlations among disparate data generated by disparate units in the equipment, and that can determine optimal performance of a complex automated tool or machine, can remain unnoticed.
Therefore, there is a need for automated equipment that is autonomous and can analyze data of a specific process, and on assets produced according to the specific process, consistently with a paradigm that is based on relationships among the data, and wherein the analysis of the data can be driven or affected by the data that surrounds the process or the associated asset themselves through learning, much like in the fashion that the human brain operates—understanding of information associated with an processes or asset is affected by the information itself, generally leading to learning and the ensuing revision of analysis goal(s), and analysis instrument(s) and approach(es) in order to improve the understanding of the information and the quality of an associated asset.
The following presents a simplified summary of the innovation in order to provide a basic understanding of some aspects of the invention. This summary is not an extensive overview of the invention. It is intended to neither identify key or critical elements of the invention nor delineate the scope of the invention. Its sole purpose is to present some concepts of the invention in a simplified form as a prelude to the more detailed description that is presented later.
An autonomous biologically based learning tool system and a method that the tool system employs for learning are provided. The autonomous biologically based learning tool system includes (i) one or more tool systems that are either individual systems or hierarchically deployed group and conglomerated systems, which perform a specific task, e.g., a semiconductor manufacturing task, or process, such as oxide etching or ion implantation, and generates data that reflects the process and a tool performance, (ii) an interaction manager that receives data and packages the data for further utilization, and (iii) an autonomous learning system based on biological principles of learning; the learning implemented through spread activation of concepts in a set of semantic networks. The autonomous learning system comprises a functional structure that can be defined recursively from a group of three functional blocks: a memory platform, a processing platform, and a knowledge communication network, through which information is communicated among the memory and processing platforms, as well as the tool system and an external actor (e.g., a computer or a human agent). Memory platform includes a hierarchy of memories, including an episodic memory to receive data impressions and associated learning instructions, a short term memory that is employed for knowledge development, and a long term memory that stores knowledge, casting the knowledge into semantic networks. Functional units in the processing platform operate on the information stored in the memory platform, facilitating learning. Such building blocks and associated functionality are inspired by the biological structure and behavior of the human brain.
Learning is accomplished through concept activation in the defined semantic networks, with activation thresholds dictated through combination of priorities associated with each concept. Priorities depend on the type of concept that is manipulated; namely, a procedural concept possesses a priority based on activation and inhibition energies.
Individual, group or conglomerate autonomous tool systems exploit the knowledge that is generated and accumulated in the autonomous learning system, which leads to multiple improvements in the autonomous biologically based learning tool as well as on assets fabricated by the various tool systems: (a) increased independence leading to lesser actor intervention (e.g., human direction and supervision) as time progresses, (b) increased production performance of outputs (e.g., output assets at least partially finished) and ensuing higher quality of the outputs, (c) data assets that convey actionable information to actors (e.g.; status of autonomous system degradation; better identification of root causes of failures; prediction of a set of system time-to-failure for individual parts, tools, tool groups and conglomerated tool, as well as associated time scales such as mean time between failures and mean time to repair), and (c) enhanced performance over time—improved products or services are delivered at a faster rate, with fewer resources consumed, and are produced with reduced tool down time.
To the accomplishment of the foregoing and related ends, the following description and the annexed drawings set forth in detail certain illustrative aspects of the claimed subject matter. These aspects are indicative, however, of but a few of the various ways in which the principles of the claimed subject matter may be employed and the claimed subject matter is intended to include all such aspects and their equivalents. Other advantages and novel features of the claimed subject matter will become apparent from the following detailed description of the claimed subject matter when considered in conjunction with the drawings.
The subject innovation is now described with reference to the drawings, wherein like reference numerals are used to refer to like elements throughout. In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present invention. It may be evident, however, that the present invention may be practiced without these specific details. In other instances, well-known structures and devices are shown in block diagram form in order to facilitate describing the present innovation.
As used in the subject specification, the terms “object,” “module,” “interface,” “component,” “system,” “platform,” “engine,” “unit,” “store,” and the like are intended to refer to a computer-related entity or an entity related to an operational machine with a specific functionality, the entity can be either hardware, a combination of hardware and software, software, or software in execution. For example, a component may be, but is not limited to being, a process running on a processor, a processor, an object, an executable, a thread of execution, a program, and/or a computer. By way of illustration, both an application running on a server and the server can be a component. One or more components may reside within a process and/or thread of execution and a component may be localized on one computer and/or distributed between two or more computers. Also, these components can execute from various computer readable media having various data structures stored thereon. The components may communicate via local and/or remote processes such as in accordance with a signal having one or more data packets (e.g., data from one component interacting with another component in a local system, distributed system, and/or across a network such as the Internet with other systems via the signal).
Moreover, the term “or” is intended to mean an inclusive “or” rather than an exclusive “or”. That is, unless specified otherwise, or clear from context, “X employs A or B” is intended to mean any of the natural inclusive permutations. That is, if X employs A; X employs B; or X employs both A and B, then “X employs A or B” is satisfied under any of the foregoing instances. In addition, the articles “a” and “an” as used in this application and the appended claims should generally be construed to mean “one or more” unless specified otherwise or clear from context to be directed to a singular form.
Referring to the drawings,
Input 130 can be regarded as extrinsic data or information, which can include (1) sounds, e.g., voice commands, environment noises or voices, alarms; (2) images captured by a static or mobile earth-based camera, or an airborne (e.g., plane, satellite) camera, wherein cameras can operate in multiple intervals of the radiation spectrum; (3) biometric indicators; (4) tokens such as batches of manufactured products, samples of materials; data which can include instructions, records, results of measurements; and so on. Output 140 can be substantially the same in nature as input 130, and it can be regarded as intrinsic data. Input and output 140 can be received and conveyed, respectively, by input and output interfaces, e.g., cameras, input pads, media docks (e.g., USB ports, IR wireless inputs), that can reside in adaptive inference component 110. As indicated above, input 130 and output 140 can be a portion of a context for adaptive inference engine 110. Additionally, adaptive inference component 110 can request input 130 as a result of pursuing a goal.
Components in autonomous biologically based system 100 can be defined recursively, which can confer the autonomous system 100 a substantial degree of competent learning complexity with basic elementary components.
Each link 115, 155, or 165 can include a communication interface that can facilitate manipulation of data or information to be transmitted or received; can utilize databases for data storage and data mining; and can receive and convey information from and to an actor. Wired embodiments of links 115, 155, or 165 can include a twisted-pair line, a T1/E1 phone line, an AC line, an optical fiber line, and corresponding circuitry, whereas wireless embodiments can comprise an ultra-mobile wide band link, a long-term evolution link, or an IEEE 802.11 link, and associated electronics. Regarding data store 150, although it is illustrated as a single element, it can be a distributed data warehouse, wherein set of data memories are deployed in disparate physical or logical locations
In example system 100, the adaptive inference engine 110 and the goal component 320 are illustrated as separate components, however, it should be appreciated that one of such components can reside within the other.
Goal component 120 can belong to one or more disciplines (e.g., a scientific discipline, a commercial discipline, an artistic discipline, a cultural discipline, and so on) or enterprise sectors (e.g., a market sector, an industry sector, a research sector, energy sector, public policy sector, and so on). Additionally, as goals can typically be multidisciplinary and focus on multiple markets, a goal component can establish multiple disparate goals within one or more particular disciplines or sectors. To pursue a goal, a goal component can comprise a functional component and a monitor component. Specific operations to accomplish a goal are effected through the functional component(s), whereas conditions of variables related to the accomplishment of the goal are determined by the monitor component. Additionally, the functional component(s) can determine a space of goals that can be accomplished by the goal component 120. A space of goals comprises substantially all goals that can be attained with a specific functionality. It should be appreciated that, for such specific functionality afforded by a functional component, a contextual adaptation of a specific goal can adapt a first goal to a second goal within a space of goals. An initial goal within a space of goals can be determined by one or more actors; wherein an actor can be a machine or a human agent (e.g., an end user). It should be noted that an initial goal can be a generic, high-level objective, as the adaptation inference engine 110 can drive goal component 120 towards a complex detailed objective through goal drifting. Goals, goal components and goal adaptation are illustrated next.
In another illustration of goal, goal component and goal adaptation, a goal can be to “purchase a DVD of movie A at store B,” the goal component 120 can be a vehicle with a navigation system that comprises an adaptive inference engine 110. (It should be noted that in this illustration the adaptive inference engine 110 resides in the goal component 120.) An actor (e.g., a vehicle operator) can enter or select the location of store B and goal component can generate directions to accomplish the goal. In the instance that the adaptive inference engine 110 receives input 130 that store B has ceased to carry in inventory movie A (e.g., an RFID reader has updated an inventory database and an update message has been broadcasted to component 110) while the actor is traveling to the store, adaptive inference engine 110 can (i) request additional input 330 to identify a store C with movie A in stock, (ii) evaluate the resources available to the actor to reach store C, and (iii) assess the level of interest of the actor in accomplishing the goal. Based on the modified context developed through input 130 as illustrated in (i)-(iii), goal component can receive an indication to adapt the goal “to purchase a DVD of movie A at store C.”
It should be appreciated that adaptive inference engine 110 can establish sub-goals associated with a goal determined by goal component 120. A sub-goal can facilitate accomplishing the goal by enabling adaptive inference engine to accomplish complementary task or to learn concepts associated with the goal.
As a summary, autonomous biologically based system 100 is a goal-driven system with contextual goal-adaptation. It should be appreciated that goal adaptation based on received context introduces an additional layer of adaptation to the analysis of input information to generate actionable information output 140. The capabilities of (a) adapting the process of information or data analysis and (b) adapting an initial goal based on context render the system massively adaptive or autonomous.
Autonomous biologically based learning system 360 includes a memory platform 365 that stores received information 358 (e.g., data, variables and associated relationships, causal graphs, templates, and so on) which can be communicated via a knowledge network 375 to a processing platform 385 that can operate on the received information, and can communicate back a processed information through the knowledge network 375 to the memory platform 365. The constituent components of autonomous learning system 360 can generally resemble biological aspects of the brain, in which a memory is networked with processing components to manipulate information and generate knowledge. Additionally, knowledge network 375 can receive information from, and convey information to, interaction component 330, which can communicate the information to tool system 310, or an actor 390 via interaction manager 345. As information 358 is received, stored, processed and conveyed by the autonomous learning system 360, multiples improvements can be effected in tool system 310 and actors that rely on it. Namely, improvements include (a) the autonomous learning system 360 and tool system 310 become increasingly independent as time progresses, and require lesser actor intervention (e.g., human direction and supervision), (b) the autonomous system improves the quality of its outputs to actors (for example, better identification of root causes of failures, or prediction of system failure before occurrence thereof), and (c) the autonomous learning system 360 improves its performance over time—the autonomous system 360 delivers improved results at a faster rate and with fewer resources consumed.
Memory platform 365 comprises a hierarchy of functional memory components, which can be configured to store knowledge (e.g., information 358) received during initialization or configuration of tool system 310 (e.g., a priori knowledge). A priori knowledge can be conveyed as information input 358 through the interaction component 330. In addition, memory platform 365 can store (a) training data (e.g., information input 358) employed to train the autonomous learning system 360 after initialization/configuration of tool system 310, and (b) knowledge generated by the autonomous learning system 360; the knowledge can be conveyed to tool system 310 or actor 390 through interaction component 330, via interaction manager 345.
Information input 358 (e.g., data) supplied by an actor 390, e.g., a human agent, can comprise data identifying a variable associated with a process, a relationship between two or more variables, a causal graph (e.g., a dependency graph), or an episode information. Such information can facilitate to guide the autonomous biologically based system 360 in a learning process. Additionally, in one aspect, such information input 358 can be deemed important by actor 390, and the importance can be related to the relevance of the information to a specific process performed by tool system 310. For instance, an operator (e.g., actor 390 is a human agent) of an oxide etch system can determine that etch rate is critical to the outcome of the manufacturing process; thus, etch rate can be an attribute communicated to autonomous learning system 360. In another aspect, information input 358 supplied by actor 390 can be a hint, whereby an indication to learn a particular relationship among process variables is made. As an example, hint can convey a suggestion to learn the behavior of pressure in a deposition chamber in tool system 310, within a specific deposition step, as a function of chamber volume, exhaust pressure and incoming gas flow. As another example, a hint can indicate to learn a detailed temporal relationship for a chamber pressure. Such example hints can activate one or more functional processing units in the autonomous learning system that can learn the functional dependence of pressure on multiple process variables. Moreover, such hints can activate one or more functional units that can apply and compare a learnt functionality with respect to model or empirical functionalities available to actor 390.
A tool system 310, e.g., a semiconductor manufacturing tool, can be complex and therefore disparate actors can specialize in manipulating and operating the tool system through disparate types of specific, complete or incomplete knowledge. As an example, a human agent, e.g., a tool engineer can know that different gases have different molecular weight and thus can produce different pressures, whereas a process/tool engineer can know how to convert a pressure reading resulting from a first gas to an equivalent pressure resulting from a second gas; an elementary example of such knowledge can be to convert a pressure reading from a unit (e.g., Pa) to another (e.g., lb/in2, or PSI). An additional type of general, more complex knowledge present in the autonomous biologically based learning system can be functional relationships between properties of a tool system (e.g., volume of a chamber) and measurements performed in the tool system (e.g., measured pressure in the chamber). For example, etch-engineers know that the etch rate is dependent on the temperature in the etch chamber. To allow for the diversity of knowledge and the fact that such knowledge can be incomplete, an actor (e.g., a human agent such as an end-user) can guide an autonomous learning system 360 through multiple degrees of conveyed knowledge: (i) No knowledge specified. Actor delivers no guidance for the autonomous learning system. (ii) Basic knowledge. Actor can convey a valid relationship between properties of a tool system and measurements in the tool system; for instance, actor conveys a relationship (e.g., relationship(κE,T)) between an etch rate (κE) and process temperature (T) without further detail. (iii) Basic knowledge with identified output. Further to a relationship between a tool system property and a tool system measurement, actor can provide specific output for a dependent variable in a relationship (e.g., relationship(output(κE), T). (iv) Partial knowledge about a relationship. Actor knows the structure of a mathematical equation among a tool system property and a measurement, as well as relevant dependent and independent variables (e.g., κE=k1e−k2/T without concrete values for k1 or k2). The actor, however, can fail to know a precise value of one for more associated constants of the relationship. (v) Complete knowledge. Actor possesses a complete mathematical description of a functional relationship. It should be noted that such guidance can be incrementally provided over time, as the autonomous learning system 360 evolves and attempts to learn tool functional relationships autonomously.
Knowledge network 375 is a knowledge bus that communicates information (e.g., data) or transfers power according to an established priority. The priority can be established by a pair of information source and information destination components or platforms. Additionally, priority can be based on the information being transmitted (e.g., this information must be dispatched in real-time). It should be noted that priorities can be dynamic instead of static and change as a function of learning development in the autonomous learning system 360, and in view of one or more demands in the one or more components present in the autonomous biologically based learning tool 300—e.g., a problem situation can be recognized and a communication can be warranted and effected in response. Communication, and power transfer, via knowledge network 375 can be effected over a wired link (e.g., a twisted pair link, a T1/E1 phone line, an AC line, an optical fiber line) or a wireless link (e.g., UMB, LTE, IEEE 802.11), and can occur among components (not shown) within a functional platform (e.g., memory platform 365 and processing platform 385) or among components in disparate platforms (e.g., a component in memory platform of self-awareness communicating with another sub-component of self-awareness) or the communication can be between components (e.g., a component of awareness communicates with a component in conceptualization).
Processing platform 385 comprises functional processing units that operate on information: Input information of a specific type (e.g., specific data types such as a number, a sequence, a time sequence, a function, a class, a causal graph, and so on) is received or retrieved and a computation is performed by a processing unit to generate output information of a specific type. Output information can be conveyed to one or more components in memory platform 365 via knowledge network 375. In an aspect, the functional processing units can read and modify data structures, or data type instance, stored in memory platform 335, and can deposit new data structures therein. In another aspect, functional processing units can provide adjustments to various numeric attributes like suitability, importance, activation/inhibition energy, and communication priority. Each functional processing unit has a dynamic priority, which determines a hierarchy for operating on information; higher priority units operate on data earlier than lower priority units. In case a functional processing unit that has operated on specific information fails to generate new knowledge (e.g., learn), like generating a ranking number or ranking function that distinguishes a bad run from a good run associated with operation of a tool system 310, the priority associated with the functional processing unit can be lowered. Conversely, if new knowledge is generated, the processing unit's priority is increased.
It should be appreciated that processing platform 385, through prioritized functional processing units, emulates a human tendency to attempt a first operation in a specific situation (e.g., a specific data type), if the operation generates new knowledge, the operation is exploited in a subsequent substantially identical situation. Conversely, when the first operation fails to produce new knowledge, a tendency to employ the first operation to handle the situation is reduced and a second operation is utilized (e.g., spread activation). If the second operation fails to generate new knowledge, its priority is reduced, and a third operation is employed. Processing platform 385 continues to employ an operation until new knowledge is generated, and another operation(s) acquire higher priority.
In an aspect, actor 390 can provide process recipe parameters, instructions (e.g., a temperature profile for an annealing cycle of an ion implanted wafer, a shutter open/close sequence in a vapor deposition of a semiconductor, an energy of an ion beam in an ion implantation process, or an electric field magnitude in a sputtering deposition), as well as initialization parameters for the autonomous learning system 360. In another aspect, an actor can supply data associated with maintenance of tool system 310. In yet another aspect, actor 390 can generate and provide results of a computer simulation of the process performed by tool system 310. Results generated in such a simulation can be employed as training data to train the autonomous biologically based learning system. Additionally, a simulation or an end-user can deliver optimization data associated with a process to tool system 370.
Autonomous learning system 360 can be trained through one or more training cycles, each training cycle can be utilized to develop the autonomous biologically based learning tool 300 to (i) be able to perform a larger number of functions without external intervention; (ii) provide better response such as improved accuracy, or correctness, when diagnosing root cause of manufacturing system health root causes; and (iii) increase performance such as faster response time, reduced memory consumption, or improved quality of product. Training data can be supplied to the autonomous learning system via adaptor component 335, in case training data is collected from data 328 associated with a process calibration or standard run in tool system 310—such data can be deemed to be internal—or through interaction manager 345. When training data is retrieved from database(s) 365 (e.g., data related to external measurements conducted through an external probe, or records of repair intervention in tool system 310); such training data can be deemed external. When training data is supplied by an actor, data is conveyed through interaction manager 345 and can be deemed external. A training cycle based on internal or external training data facilitates autonomous learning system 360 to learn an expected behavior of tool system 310.
As indicated above, functional component 315 can comprise multiple functional tool components (not shown) associated with the tool specific semiconductor manufacturing capabilities and that enable the tool to be used to (a) manufacture semiconductor substrates (e.g., wafers, flat panels, liquid crystal displays (LCDs), and so forth), (b) conduct epitaxial vapor depositiontion or non-epitaxial vapor deposition, (c) facilitate ion implantation or gas cluster ion infusion, (d) perform a plasma or non-plasma (dry or wet) an oxide etch treatment, (e) implement a lithographic process (e.g., photo-lithography, e-beam lithography, etc.), and so on. The tool system 310 can also be embodied in a furnace; an exposure tool for operation in a controlled electrochemical environment; a planarization device; an electroplating system; a test device for optical, electrical, and thermal properties, which can included lifespan (through operation cycling) measurements; a metrology tool, a wafer cleaning machine, and the like.
In the process conducted by tool system 310, sensors and probes comprising sensor component 325 can collect data (e.g., data assets) on different physical properties (e.g., pressure, temperature, humidity, mass density, deposition rate, layer thickness, surface roughness, crystalline orientation, doping concentration, etc.) as well as mechanical properties (valve aperture or valve angle, shutter on/off operation, gas flux, substrate angular velocity, substrate orientation, and the like) through various transducers and techniques with varying degrees of complexity depending on the intended use of the gathered data. Such techniques can include, but are not limiting to including, X-ray diffraction, transmission electron microscopy (TEM), scanning electron microscopy (SEM), mass spectrometry, light-exposure assessment, magnetoelectric transport measurements, optical properties measurements, and so on. Additional data assets that are relevant to a product (e.g., a semiconductor substrate) are development inspection (DI) critical dimension (CD), and final inspection (FI) CI. It should be appreciated that probes can be external to tool system 310 and can be accessed through an interface component (not shown). For instance, such external probes can provide DI CI and FI CI. It should be appreciated that such data assets 328 effectively characterize output assets, or physical products manufactured or fabricated by tool system 310.
In an aspect, data sources in sensor component 325 can be coupled to adaptor component 335, which can be configured to gather data assets 328 in analog or digital form. Adaptor component 335 can facilitate data 368 collected in a process run to be composed or decomposed according to the intended utilization of the data in autonomous learning system 310 before the data is deposited into memory platform 365. Adaptors in adaptor component 335 can be associated with one or more sensors in sensor component 325 and can read the one or more sensors at specific frequencies, or in other specific conditions. An external data source adapter may have the ability to pull data as well as pass through data that is pushed from outside the tool. For example, an MES/historical database adaptor knows how to consult an MES database to extract information for various autobots and package/deposit the data into working memory for one or more components of the autonomous system. As an example, adaptor component 335 can gather wafer-level run data one wafer at a time as the tool processes the wafer. Then, adaptor component 335 can consolidate individual runs in a batch to form “lot-level-data,” “maintenance-interval-data”, etc. Alternatively, if tool system 310 outputs a single file (or computer product asset) for lot-level data, adaptor component 335 can extract wafer-level data, step-level data, and the like. Furthermore, decomposed data elements can relate to one or more components of tool system 300; e.g., variables and times at which a pressure controller in sensor component 325 is operating. Subsequent to processing, or packaging, received data 328 as described above, adaptor component 335 can store processed data in database(s) 355.
Database(s) 355 can include data originated in (i) tool system 370, through measurements performed by sensors in sensor component 325, (ii) a manufacturing execution system (MES) database or a historical database, or (iii) data generated in a computer simulation of tool system 310, e.g., a simulation of semiconductor wafer manufacturing performed by actor 390. In an aspect, an MES is a system that can measure and control a manufacturing process, can track equipment availability and status, can control inventory, and can monitor for alarms.
It is to be appreciated that products, or product assets, fabricated by tool system 310 can be conveyed to actor 390 through interaction component 330. It should be appreciated that product assets can be analyzed by actor 390 and the resulting information, or data assets, conveyed to autonomous learning system 360. In another aspect, interaction component 330 can perform analysis of a product asset 328 via adaptor component 335.
In addition it is to be noted that in embodiment 300 the interaction component 340 and autonomous learning system 360 are externally deployed with respect to tool system 310. Alternative deployment configurations of autonomous biologically based learning tool 300 can be realized, such as embedded deployment wherein interaction component 340 and autonomous biologically based learning system 310 can reside within tool system 370, in a single specific tool component; e.g., single embedded mode, or in a cluster of tool components; e.g., multiple embedded mode. Such deployment alternatives can be realized in a hierarchical manner, wherein an autonomous learning system supports a set of autonomous learning tools that form a group tool, or a tool conglomerate. Such complex configurations are discussed in detail below.
Next, an illustrative tool system 310 is discussed in connection with
In an aspect, process unit 410 comprises a first process unit group 430 which possesses a cooling unit (COL) 435, an alignment unit (ALIM) 440, an adhesion unit (AD) 445, an extension unit (EXT) 450, two prebaking units (PREBAKE) 455, and two postbaking units (POBAKE) 460, which are stacked sequentially from the bottom. Additionally, a second process unit group 465 includes a cooling unit (COL) 435, an extension-cooling unit (EXTCOL) 470, an extension unit (EXT) 475, a second cooling unit (COL) 435, two prebaking units (PREBAKE) 455 and two postbaking units (POBAKE) 460. Cooling unit (COL) 435 and the extension cooling unit (EXTCOL) 470 may be operated at low processing temperatures and arranged at lower stages, and the prebaking unit (PREBAKE) 455, the postbaking unit (POBAKE) 460 and the adhesion unit (AD) 445 are operated at high temperatures and arranged at the upper stages. With this arrangement, thermal interference between units can be reduced. Alternatively, these units can have alternative or additional arrangements. The prebaking unit (PREBAKE) 455, the postbaking unit (POBAKE) 460, and the adhesion unit (AD) 445 each comprise a heat treatment apparatus in which substrates are heated to temperatures above room temperature. In an aspect, temperature and pressure data can be supplied to the autonomous biologically based learning system 360 through interface component 340, from prebaking unit 455, postbaking unit 460, and adhesion unit 445. Rotational speed and positional data for a substrate can be conveyed from alignment unit 440.
Furthermore, autonomous learning system 360 can comprise one or more primary functional units which include a self-awareness component 550, a self-conceptualization component 560, and a self-optimizing component 570. A first feed forward (FF) loop 552 can act as a forward link and can communicate data among self-awareness component 550 and self-conceptualization 560. In addition, a first feed back (FB) loop 558 can act as a reverse link and can communicate data among self-conceptualization component 570 and self-awareness component 550. Similarly, forward link and reverse link data communication among self-conceptualization component 560 and self-optimization component 570 can be accomplished, respectively, through a second FF loop 562 and a second FB loop 568. It should be appreciated that in a FF link, data can be transformed prior to communication to the component that receives the data to further process it, whereas in a FB link a next data element can be transformed by the component that receives the data prior to process it. For example, data transferred through FF link 552 can be transformed by self awareness component 550 prior to communication of the data to self-conceptualizing component 560. It should further be appreciated that FF links 552 and 562 can facilitate indirect communication of data among components 550 and component 570, whereas FB links 568 and 558 can facilitate an indirect communication of data among components 570 and 550. Additionally, data can be conveyed directly among components 550, 360, and 370 through knowledge network 375.
Long term memory 510 can store knowledge supplied through interaction component 330 during initialization or configuration of a tool system (e.g., a priori knowledge) to train the autonomous learning tool system 300 after initialization/configuration. In addition, knowledge generated by autonomous learning system 360 can be stored in long term memory 510. It should be appreciated that LTM 510 can be a part of a memory platform 365 and thus can display substantially the same characteristics thereof. Long term memory 510 can generally comprise a knowledge base that contains information about tool system components (e.g., manufacturing components, probe components, and so on), relationships, and procedures. At least a portion of the knowledge base can be a semantic network that describes or classifies data types (for example as a sequence, an average, or a standard deviation), relationships among the data types, and procedures to transform a first set of data types into a second set of data types.
A knowledge base may contain knowledge elements, or concepts. In an aspect, each knowledge element can be associated with two numeric attributes: a suitability (ξ) and an inertia (ι) of a knowledge element, or concept; collectively such attributes determine a priority of a concept. A well-defined function, e.g., a weighted sum, a geometric average, of these two numeric attributes can be a concept's situation score (σ). For example, σ=ξ+ι. The suitability of a knowledge element can be defined as a relevance of the knowledge element (e.g., concept) to a tool system or a goal component situation at a specific time. In an aspect, a first element, or concept, with a higher suitability score than a second element can be more relevant to a current state of the autonomous learning system 360 and a current state of a tool system 310 than the second element with a lower suitability score. The inertia of a knowledge element, or concept, can be defined as the difficulty associated with utilization of the knowledge element. For example, a low first value of inertia can be conferred to a number element, a list of numbers can be attributed a second inertia value higher than the first value, a sequence of numbers can have a third value of inertia that is higher than the second value, and a matrix of numbers can have a fourth value of inertia which can be higher than the third value. It is noted that inertia can be applied to other knowledge or information structures like graphs, tables in a database, audio files, video frames, code snippets, code scripts, and so forth; the latter items can substantially all be a portion of input 130. The subject innovation provides for a well defined function of the suitability and the inertia that can influence the likelihood that a knowledge element is retrieved and applied. Concepts that have the highest situational score are the most likely concepts to be rendered to short term memory 520 for processing by processing units.
Short term memory 520 is a temporary storage that can be utilized as a working memory (e.g., a workspace or cache) or as a location where cooperating/competing operations, or autobots, associated with specific algorithms or procedures, can operate on data types. Data contained in STM 520 can possess one or more data structures. Such data structures in STM 520 can change as a result of data transformations effected by autobots and planner überbots (e.g., autobots dedicated to planning). The short term memory 305 can comprise data, learning instructions provided by the interaction manager 345, knowledge from the long term memory 310, data provided and/or generated by one or more autobots or überbots, and/or initialization/configuration commands provided by an actor 390. Short term memory 520 can track a state of one or more autobots and/or überbots used to transform data stored therein.
Episodic memory 530 stores episodes which can include an actor-identified set of parameters and concepts which can be associated with a process. In an aspect, an episode can comprise extrinsic data or input 130, and it can provide with a specific context to autonomous learning system 100. It is noted that an episode can generally be associated with a particular scenario identified or generated (e.g., by tool system 110, a goal component 120, or an autonomous learning system 160) while pursuing a goal. An actor that identifies an episode can be a human agent, like a process engineer, a tool engineer, a field support engineer, and so on, or it can be a machine. It should be appreciated that episodic memory 530 resembles a human episodic memory, wherein knowledge associated with particular scenario(s)—e.g., an episode—can be present and accessible without a recollection of the learning process that resulted in the episode. Introduction, or definition, of an episode typically is a part of a training cycle or substantially any extrinsic provision of input, and it can lead to an attempt by the autonomous biologically based learning system 360 to learn to characterize data patterns, or input patterns, that can be present in data associated with the episode. A characterized pattern of data associated with an episode can be stored in episodic memory 530 in conjunction with the episode and an episode's name. The addition of an episode to episodic memory 530 can result in a creation of an episode-specific autobot that can become active when a set of parameters in a process conducted by a tool system 310, or a generally a goal component 120, enter an operating range as defined in the episode; the episode-specific autobot receives sufficient activation energy when the first feature associated with a pursued goal or process is recognized. If the parameters meet the criteria established through a received episode, the episode-specific autobot compares the pattern of data in the episode with the current data available. If the current situation (as defined by the recognized pattern of data) of the tool system 310, or a goal component, matches the stored episode, an alarm is generated to ensure the tool maintenance engineers can become aware of the situation and can take preventive action(s) to mitigate additional damage to functional component 315 or sensor component 325 or material utilized in a tool process.
Autobot component 540 comprises a library of autobots that perform a specific operation on an input data type (e.g., a matrix, a vector, a sequence, and so on). In an aspect, autobots exist in an autobot semantic net, wherein each autobot can have an associated priority; a priority of an autobot is a function of its activation energy (EA) and its inhibition energy (E1). Autobot component 540 is an organized repository of autobots that can include autobots for the self-awareness component 550, self-conceptualization component 560, self-optimization component 570, and additional autobots that can participate in transforming and passing data among components and among the various memory units. Specific operations that can be performed by an autobot can include a sequence average; a sequence ordering; a scalar product among a first and a second vector; a multiplication of a first matrix and a second matrix; a time sequence derivative with respect to time; a sequence autocorrelation computation; a crosscorrelation operation between a first and a second sequence; a decomposition of a function in a complete set of basis functions; a wavelet decomposition of a time sequence numeric data stream, or a Fourier decomposition of a time sequence. It should be appreciated that additional operations can be performed depending on input data; namely, feature extraction in an image, sound record, or biometric indicator, video frame compression, digitization of environment sounds or voice commands, and so on. Each of the operations performed by an autobot can be a named function that transforms one or more input data types to produce one or more output data types. Each function for which there exists an autobot in autobot component 540 can possess an element in LTM, so that itherbots can make autobot activation/inhibition energy decisions based on the total “attention span” and needs of the autonomous learning system 360. Analogously to the autonomous learning system 360, an autobot in autobot component 540 can improve its performance over time. Improvements in an autobot can include better quality of produced results (e.g., outputs), better execution performance (e.g., shorter runtime, capability to perform larger computations, and the like), or enhanced scope of input domain for a particular autobot (e.g., inclusion of additional data types that the autobot can operate on).
Knowledge—concepts and data—stored in LTM 510, STM 520 and EM 530 can be employed by primary functional units, which confer autonomous biologically based learning system 360 a portion of its functionality.
Self-awareness component 550 can determine a level of tool system degradation between a first acceptable operating state of the tool system 310 and a subsequent state, at a later time, in which tool system has degraded. In an aspect, autonomous learning system 360 can receive data that characterizes an acceptable operating state, and data associated with a product asset fabricated in such acceptable state; such data assets can be identified as canonical data. Autonomous biologically based learning system 360 can process the canonical data and the associated results (e.g., statistics about important parameters, observed drift in one or more parameters, predictive functions relating tool parameters, and so on) can be stored by self-awareness component 550 and employed for comparison to data supplied as information input 358; e.g., production process data or test run data. If a difference between generated, learnt results of the canonical data and the device process run-data is small, then the manufacturing system degradation can be considered to be low. Alternatively, if the difference between stored learnt results of the canonical data and the sample process data is large, then there can be a significant level of tool system (e.g., semiconductor manufacturing system) degradation. A significant level of degradation can lead to a process, or goal, contextual adjustment. Degradation as described herein can be computed from a degradation vector (Q1, Q2, . . . , QU) where each component Qλ (λ=1, 2, . . . , U) of the degradation vector is a different perspective of an available data set—e.g., Q1 may be a multivariate mean, Q2 the associated multivariate deviation, Q3 a set of wavelet coefficients for a particular variable in a process step, Q4 may be the mean difference between a predicted pressure and measured pressure, etc. Normal training runs produce a specific set of values (e.g., a training data asset) for each component, which can be compared with component Q1-QU generated with run data (e.g., a run data asset) from each component. To assess degradation, a suitable distance metric can be to employed to compare the (e.g., Euclidean) distance of a run degradation vector from its “normal position” in {Q} space; the large such Euclidean distance, the more a tool system is said to be degraded. In addition, a second metric can be to compute a cosine similarity metric among the two vectors.
Self-conceptualization component 560 can be configured to build an understanding of important tool system 310 relationships (e.g., one or more tool behavior functions) and descriptions (e.g., statistics regarding requested and measured parameters, influence of parameters on degradation, etc.). It is to be appreciated that relationships and descriptions are also data, or soft, assets. The understanding is established autonomously (e.g., by inference and contextual goal adaptation originated from input data; inference can be accomplished, for example, via multivariate regression or evolutionary programming, such as genetic algorithms) by autonomous learning system 360, or through an actor 390 (e.g., a human agent) supplied guidance. Self-conceptualization component 560 can construct a functional description of a behavior of a single parameter of a tool system 310, or generally a goal component like component 120, such as pressure in a deposition chamber in a semiconductor manufacturing system as a function of time during a specific deposition step. In addition, self-conceptualization component 560 can learn a behavior associated with a tool system, like a functional relationship of a dependent variable on a specific set of input information 358. In an aspect, self-conceptualization component 560 can learn the behavior of pressure in a deposition chamber of a given volume, in the presence of a specific gas flow, a temperature, exhaust valve angle, time, and the like. Moreover, self-conceptualization component 560 can generate system relationships and properties that may be used for prediction purposes. Among learnt behaviors, self-conceptualization component can learn relationships and descriptions that characterize a normal state. Such normal state typically is employed by autonomous learning system 360 as a reference state with respect to which variation in observer tool behavior is compared.
Self-optimization component 570 can analyze a current health or performance of an autonomous biologically based learning system 300 based on the level of a tool system 310 deviation between predicted values (e.g., predictions based on functional dependence or relationships learnt by self-conceptualization component 560 and measured values) in order to identify (a) a potential cause of failure of tool system 360, or (b) one or more sources of root cause of the tool system degradation based on information gathered by autonomous learning system 360. Self-optimizing component 570 can learn over time whether autonomous learning system 360 initially incorrectly identifies an erroneous root cause for a failure, the learning system 300 allows for input of maintenance logs or user guidance to correctly identify an actual root cause. In an aspect, the autonomous learning system 360 updates a basis for its diagnosis utilizing Bayesian inference with learning to improve future diagnosis accuracy. Alternatively, optimization plans can be adapted, and such adapted plans can be stored in an optimization case history for subsequent retrieval, adoption, and execution. Moreover, a set of adaptations to a process conducted by tool system 310, or generally a goal pursued by a goal component 120, can be attained through the optimization plans. Self-optimization component 570 can exploit data feedback (e.g., loop effected through links 565, 555, and 515) in order to develop an adaptation plan that can promote process or goal optimization.
In embodiment 500, autonomous biologically based learning system 360 can further comprise a planner component 580 and a system context component 590. The hierarchy of functional memory components 510, 520, and 530, and the primary functional units 550, 560, and 570 can communicate with planner component 580 and the system context component 590 through knowledge network 375.
Planner component 580 can exploit, and comprise, higher level autobots in autobot component 540. Such autobots can be identified as planner überbots, and can implement adjustments to various numeric attributes like a suitability, an importance, an activation/inhibition energy, and a communication priority. Planner component 580 can implement a rigid, direct global strategy; for instance, by creating a set of planner überbots that can force specific data types, or data structures, to be manipulated in short term memory 520 through specific knowledge available in short term memory 505 and specific autobots. In an aspect, autobots created by planner component 580 can be deposited in autobot component 540, and be utilized over the knowledge network 375. Alternatively, or in addition, planner component 580 can implement an indirect global strategy as a function of a current context of an autonomous learning system 360, a current condition of a tool system 310, a content of short term memory 520 (which can include associated autobots that can operate in the content), and a utilization cost/benefit analysis of various autobots. It should be appreciated that the subject autonomous biologically based learning tool 300 can afford dynamic extension of planner components.
Planner component 580 can act as a regulatory component that can ensure process, or goal, adaptation in an autonomous biologically based tool 300 does not result in degradation thereof. In an aspect, regulatory features can be implemented through a direct global strategy via creation of regulatory überbots that infer operational conditions based on planned process, or goal, adaptation. Such an inference can be effected through a semantic network of data types on which the regulatory überbots act, and the inference can be supported or complemented by cost/benefit analysis. It should be appreciated that planner component 580 can preserve goals drifting within a specific region of a space of goals that can mitigate specific damages to a goal component, e.g., a tool system 310.
System context component 590 can capture the current competency of an autonomous biologically based learning tool 300 that exploits autonomous learning system 360. System context component 590 can include a state identifier that comprises (i) a value associated with an internal degree of competency (e.g., a degree of effectiveness of a tool system 310 in conducting a process (or pursuing a goal), a set of resources employed while conducting the process, a quality assessment of a final product or service (or an outcome of a pursued goal), a time-to-delivery of devices, and so on), and (ii) a label, or identifier, to indicate the state of the autonomous learning tool 300. For instance, the label can indicate states such as “initial state,” “training state,” “monitoring state,” “learning state,” or “applying knowledge.” The degree of competency can be characterized by a numerical value, or metric, in a determined range. Further, the system context component 590 can include a summary of learning performed by the autonomous learning system 360 over a specific time interval, as well as a summary of possible process or goal adaptations that can be implemented in view of the performed learning.
An autobot (e.g., autobot 660) can also be self-describing in that the autobot can specify (a) one or more types of input data the autobot can manipulate or require, (b) a type of data the autobot can generate, and (c) one or more constraints on input and output information. In an aspect, interface 672 can facilitate autobot 660 to self-describe and thus express the autobot's availability and capability to überbots, in order for the überbots to supply activation/inhibition energy to the autobots according to a specific tool scenario.
Awareness working memory (AWM) 710 is a S™ that can include a special region of memory identified as awareness sensory memory (ASM) 720 that can be utilized to store data, e.g., information input 358, that can originate in a sensor in sensor component 325 or in actor 390, can be packaged by one or more adaptors in adaptor component 335, and can be received by knowledge network 375. Self-awareness component 550 can also comprise multiple special functionality autobots, which can reside in autobot component 540 and include awareness planner überbots (APs).
In addition, self-awareness component 550 can comprise an awareness knowledge memory (AKM) 730 which is a part of a L™ and can include multiple concepts—e.g., an attribute; an entity such as a class or a causal graph; a relationship, or a procedure−relevant to the operation of self-awareness component 550. In an aspect, a self-awareness component 550 for a semiconductor manufacturing tool can include domain specific concepts like a step, a run, a batch, a maintenance-interval, a wet-clean-cycle, etc., as well as general purpose concepts like a number, a list, a sequence, a set, a matrix, a link, and so on. Such concepts can enter a higher level of abstraction; for instance, a wafer run can defined as an ordered sequence of steps where a step has both recipe parameter settings (e.g., desired values), and one or more step measurements. Furthermore, AKM 730 can include functional relationships that can link two or more concepts like an average, a standard deviation, a range, a correlation, a principal component analysis (PCA), a multi-scale principal component analysis (MSPCA), a wavelet or substantially any basis function, etc. It should be noted that multiple functional relationships can be applicable, and hence related, to a same concept; for example, a list of numbers is mapped to a real number instance by the average, which is a (functional) relation and a standard-deviation relation, as well as a maximum relation, and so forth). When a relationship from one or more entities to another entity is a function or a functional (e.g., a function of a function), there can be an associated procedure that can executed by an überbot in order to effect the function. A precise definition of a concept can be expressed in a suitable data schema definition language, such as UML, OMGL, etc. It should be further noticed that a content of AKM 730 can be augmented dynamically at (tool system) runtime without shutting the system down.
Each concept in AKM 730, as any concept in a knowledge base as described herein, can be associated with a suitability attribute and an inertia attribute, leading to the concept's specific situation score. Initially, before the autonomous system is provided with data, the suitability value for all elements in AKM 730 is zero, but the inertia for all concepts can be tool dependent and can be assigned by an actor, or based on historical data (e.g., data in database(s) 355). In an aspect, inertia of a procedure that produces an average from a set of numbers can be substantially low (e.g., t=1) because computation of an average can be regarded as a significantly simple operation that can be applicable to substantially all situations involved collected data sets, or results from computer simulations. Similarly, maximize and minimize procedures, which transform a set of numbers, can be conferred a substantially low inertia value. Alternatively, compute a range and compute a standard deviation can be afforded higher inertia values (e.g., t=2) because such knowledge elements are more difficult to apply, whereas calculate a PCA can display a higher level of inertia and calculate a MSPCA can have a yet higher value of inertia.
A situation score can be employed to determine which concept(s) to communicate among from AKM 730 and AWM 710 (see below). Knowledge elements, or concepts, that exceed a situation score threshold are eligible to be conveyed to AWM 710. Such concepts can be conveyed when there is sufficient available storage in AWM 710 to retain the concept and there are no disparate concepts with a higher situation score that have not been conveyed to AWM 710. A concept's suitability, and thus a concept's situation score, in AWM 710 can decay as time progresses, which can allow new concepts with a higher suitability to enter awareness working memory 710 when one or more concepts already in memory are no longer needed or are no longer applicable. It is noted that the larger the concept's inertia the longer it takes the concept to both be conveyed to and be removed from AWM 710.
When a tool system state changes, e.g., a sputter target is replaced, an electron beam gun is added, a deposition process is finished, an in situ probe is initiated, an annealing stage is completed, and so on, awareness planner 550 überbots can document which concepts (e.g., knowledge elements) can be applied in the new state, and can increase a suitability value, and thus a situation score, of each such a concept in AKM 730. Similarly, the activation energy of autobots 6151-615N can be adjusted by uberbots in order to reduce the activation energy of specific autobots, and to increase EA for autobots that are appropriate to a new situation. The increment in suitability (and situation score) can be spread by planner überbots to those concepts' first neighbors and then to second neighbors, and so forth. It should be appreciated that a neighbor of a first concept in AKM 730 can be a second concept that resides, in a topological sense, within a specific distance from the first concept according to a selected measure, e.g. number of hops, Euclidean distance, etc.) It is noted that the more distant a second concept is from a first concept that received an original increment in suitability, the smaller the second concept's increment in suitability. Thus, suitability (and situation score) increments present a dampened spread as a function of “conceptual distance.”
In architecture 500, self-awareness component 550 comprises an awareness schedule adapter (ASA) 760 which can be an extension of awareness planner component 750 and can request and effect changes in collection extrinsic data or intrinsic data (e.g., via sensor component 325 through interaction component 330, via input 130, or via (feedback) link 155). In an aspect, awareness schedule adapter 760 can introduce data sampling frequency adjustments—e.g., it can regulate a rate at which different adaptors in adaptor component 335 can convey data to knowledge network 375 (e.g., information input 358) intended for ASM 720. Moreover, awareness schedule adapter 760 can sample at low frequency, or substantially eliminate, collection of data associated with process variables that are not involved in the description of normal patterns of data, or variables that fail to advance the accomplishment of a goal as inferred from data received in an adaptive inference engine. Conversely, ASA 760 can sample at higher frequency a set of variables extensively used in a normal pattern of data, or that can actively advance a goal. Furthermore, when the autonomous learning system 360 acknowledges a change of state tool system 310 (or a change in a situation associated with a specific goal) wherein data indicate that product quality or process reliability are gradually deviating from normal data patterns (or a goal drift is resulting in significant departure from an initial goal in the space of goals), the autonomous learning system can request, via ASA 760, a more rapid sampling of data to collect a larger volume of actionable information (e.g., input 130) that can effectively validate the degradation and trigger an appropriate alarm accordingly. In an aspect, a goal component can display a goal drift summary to an actor that entered an initial goal; e.g., a customer in an electronics store that has substantially departed from an initial expenditure goal when procuring a home entertainment system can be displayed a log with changes in a projected expense after budget adaptation; or a database architect can be shown costs associated with memory space and associated infrastructure upon adaptation of a goal to optimize a data warehouse.
An actor 390 (e.g., a human agent) can train self-awareness component 550 in multiple manners, which can include a definition of one or more episodes (including, for instance, illustrations of successfully adapted goals). A training of the autonomous learning system 360, through self-awareness component 550, for an episode can occur as follows. The actor 390 creates an episode and provides the episode with a unique name. Data for the newly created episode can then be given to autonomous learning system 360. The data can be data for a specific sensor during a single specific operation step of a tool system, a set of parameters during a single specific step, a single parameter average for a run, etc.
Alternatively, or additionally, more elementary guidance can be provided by actor 390. For example, a field support engineer can perform preventive tool maintenance (PM) on tool system 310. PM can be planned and take place periodically, or it can be unplanned, or asynchronous. It should be appreciated that preventive tool maintenance can be performed on the manufacturing system in response to a request by the autonomous learning system 360, in response to routine preventive maintenance, or in response to unscheduled maintenance. A time interval elapses between consecutive PMs, during such a time interval one or more processes (e.g., wafers/lots manufacturing) can take place in the tool system. Through data and product assets and associated information, such as effected planner and unplanned maintanence, autonomous learning system can infer a “failure cycle.” Thus, the autonomous learning system can exploit asset(s) 328 to infer a mean time between failures (MTBF). Such inference is supported through a model of time-to-failure as a function of critical data and product assets. Furthermore, autonomous learning system 360 can develop models, through relationships among disparate assets received as information I/O 358 or through historic data resulting from supervised training sessions delivered by an expert actor. It should be appreciate that an expert actor can be a disparate actor that interacts with a trained disparate autonomous learning system.
Actor 390 can guide the autonomous system by informing the system that it can average wafer level run data and assess a drift in critical parameters across PM intervals. A more challenging exercise can also be performed by the autonomous system, wherein the actor 390 indicates through a learning instruction to autonomous learning system 360 to learn to characterize a pattern of data at the wafer average level before each unplanned PM. Such an instruction can promote the autonomous learning system 360 to learn a pattern of data prior to an unplanned PM, and if a pattern of data can be identified by an awareness autobot, the self-awareness component 550 can learn such a pattern as time evolves. During learning a pattern, awareness component 550 can request assistance (or services) from self-conceptualization component 560 or awareness autobots that reside in autobot component 540. When a pattern for the tool system is learned with a high degree of confidence (e.g. measured by a degree of reproducibility of the pattern as reflected in coefficients of a PCA decomposition, a size of a dominant cluster in a K-cluster algorithm, or a prediction of the magnitude of a first parameter as a function of a set of disparate parameters and time, and so forth), autonomous biologically based learning system 360 can create a reference episode associated with the malfunction that can lead to the need of tool maintenance so that an alarm can be triggered prior to occurrence of the reference episode. It is noted that awareness autobots, which can reside in autobot component 540, can fail to characterize completely a data pattern for the malfunction reference episode, or substantially any specific situation that can require unplanned maintenance, before it is necessary. It should be appreciated nonetheless that such a preventive health management of a tool system 310, which can include a deep behavioral and predictive functional analysis, can be performed by autobots in self-conceptualization component 560.
Embodiment 900 illustrates a conceptualization knowledge memory (CKM) 910 that includes concepts (e.g., attributes, entities, relationships, and procedures) necessary for operation of self-conceptualization component 570. Concepts in CKM 910 include (i) domain specific concepts such as a step, a run, a lot, a maintenance-interval, a wet-clean-cycle, a step-measurements, a wafer-measurements, a lot-measurements, a location-on-wafer, a wafer-region, a wafer-center, a wafer-edge, a first-wafer, a last-wafer, etc.; and (ii) general purpose, domain independent concepts like a number, a constant (e.g., e, π), a variable, a sequence, a time-sequence, a matrix, a time-matrix, a fine-grained-behavior, a coarse-grained-behavior, etc. Self-conceptualization component also includes a vast array of general purpose functional relations such as add, subtract, multiply, divide, square, cube, power, exponential, log, sine, cosine, tangent, of and so forth, as well as other domain specific functional relations that can present various levels of detail and reside in adaptive conceptualization template memory (ACTM) 920.
ACTM 920 is an extension of CKM 910 that can hold functional relationships that are either completely or partially known to an actor (e.g., an end user) that interacts with a tool system 310 (e.g., a semiconductor manufacturing tool). It should be noted that while ACTM is a logical extension of CKM, autobots, planners, and other functional components are not affected by such separation, as the actual memory storage can appear a single storage unit within self-conceptualization component 560. Self-conceptualization component 560 can also include a conceptualization goal memory (CGM) 930 which is an extension of a conceptualization working memory (CWM) 940. CGM 930 can facilitate autobots of a current goal, e.g., to learn (f, pressure, time, step); for a particular process step, learn a function f of pressure wherein the function depends on time. It should be noted that learning function f represents a sub-goal that can facilitate accomplishing the goal of manufacturing a semiconductor device utilizing tool system 310.
Concepts in ACTM 920 also have a suitability numeric attribute and an inertia numeric attribute, which can lead to a situation score. A value of inertia can indicate a likelihood of a concept to be learnt. For example, a higher inertia value for a matrix concept and a lower inertia for a time-sequence concept can lead to a situation where self-conceptualization component 560 can learn a functional behavior of time-sequences rather than a functional behavior of data in a matrix. Similarly to self-awareness component 550, concepts with lower inertia are more likely to be conveyed from CKM 910 to CWM 940.
Conceptual planners (CPs) provide activation energy to the various autobots and provide situation energy to various concepts in CKM 910 and ACTM 920, as a function of a current context, a current state of tool system 310 (or generally a goal component 120), a content of CWM 940, or current autobot(s) active in CWM 940. It should be appreciated that activation energy and situation energy alterations can lead to goal adaptation based on the knowledge generated (e.g., based on learning) as a result of the altered semantic network for concepts in CWM 940 or CKM 910—as inference by an adaptive inference engine can be based on propagation aspects of concepts.
Contents of CTM 920 are concepts which can describe the knowledge discussed above, and thus those concepts can have suitability and inertia numeric attributes. The contents of CTM 920 can be used by autobots to learn the functional behavior of the tool system 310 (subject to the constraint that concepts with lower inertia are more likely to be activated over concepts with higher inertia.). It is not necessary for all guidance to have the same inertia; for instance, a first complete function can be provided a lower inertia than a second complete function even though both concepts represent complete functions.
When partial knowledge like a partially-defined equation is uploaded in CWM 940, it can be completed, e.g., with existing knowledge—CPs coordinate autobots to employ available data to first identify values for unknown coefficients. A set of ad hoc coefficients can thus complete the partially-defined equation concept into a complete function concept. The complete equation concept can then be utilized in a pre-built functional-relation concept such as add, multiply, etc. Basic knowledge with output (e.g., relationship(output(κE),T)) can facilitate autobots in CWM 940 to construct and evaluate various functional descriptions that involve data for κE and T in order to identify the best function that can describe a relationship among κE and T. Alternatively, basic knowledge without output can facilitate autobots, with assistance of CPs, to specify a variable as an output, or independent, variable and attempt to express it as a function of the remaining variables. When a good functional description is not found, an alternative variable can be specified as an independent variable the process is iterated until it converges to an adequate functional relationship or autonomous learning system 360 indicates, for example to actor 390, that an adequate functional relationship is not found. An identified good functional relationship can be submitted to CKM 910 to be utilized by autobots in autonomous learning system 360 with a level of inertia that is assigned by the CPs. For instance, the assigned inertia can be a function of the mathematical complexity of the identified relationship—a linear relationship among two variables can be assigned an inertia value that is lower than the assigned inertia to a non-linear relationship that involve multiple variables, parameters, and operators (e.g., a gradient, a Laplacian, a partial derivative, and so on).
Conceptualization engine 945 can be a “virtual component” that can present coordinated activities of awareness autobots and conceptualization autobots. In an aspect, self-awareness component 550 can feed forward (through FF loop 552) a group of variables (e.g., variables in the group can be those that display good pairwise correlation properties) to self-conceptualization component 560. Forwarded information can facilitate self-conceptualization component 560 to check CKM 910 and ACTM 920 for function relation templates. The availability of a template can allow an autobot of a conceptualization learner (CL), which can reside in the conceptualization engine 945, to more quickly learn a functional behavior among variables in a forwarded group. It should be appreciated that learning such a functional behavior can be a sub-goal of a primary goal. A CL autobot with the assistance of a CP autobot can also use autobots of a conceptualization validator (CV). CV autobots can evaluate a quality of proposed functional relationships (e.g., average error between a predicted value and a measurement is within instrument resolution). A CL autobot can independently learn a functional relationship either autonomously or through actor-supplied guidance; such actor supplied guidance can be regarded as extrinsic data. Functions learned by a CL can be fed back (e.g., via FB link 558) to self-awareness component 550 as a group of variables of interest. For example, after learning the function κE=κ0 exp(−U/T), wherein κ0 (e.g., an asymptotic etch rate) and U (e.g., an activation barrier) possess specific values known to the CL, self-conceptualization component 560 can feed back the guidance group (output(κE,T) to self-awareness component 550. Such feed back communication can afford self-awareness component 550 to learn patterns about such group of variables so that degradation with respect to the group of variables can be quickly recognized and, if necessary, an alarm generated (e.g., an alarm summary, an alarm recipient list verified) and triggered. Memory 960 is a conceptualization episodic memory.
The following two aspects related to CL and CV should be noted. First, CL can include autobots that can simplify equations (e.g., through symbolic manipulation), which can facilitate to store a functional relationships as a succinct mathematical expression. As an example, the relationship P=((2+3)Φ)((1+0)÷θ) is simplified to P=3Φ÷θ, where P, Φ and θ indicate, respectively, a pressure, a flow and an exhaust valve angle. Second, CV can factor in the complexity of the structure of an equation when it determines a quality of the functional relationship—e.g., for parameters with substantially the same characteristics, like average error of predicted values versus measurements, a simpler equation can be preferred instead of a more complicated equation (e.g., simpler equation can have lower concept inertia).
Additionally, important FF 552 communication of information from self-awareness component 550 to self-conceptualization component 560, and FB 558 communication from self-conceptualization component 560 to self-awareness component 550, can involve cooperation of awareness autobots and conceptualization autobots to characterize a pattern of data for an episode. As discussed above in connection with
Similarly, a prediction of an unscheduled PM can rely on knowledge of temporal fluctuations of critical measurements of tool system data and the availability of a set of predictive functions conveyed by self-conceptualization component 570. The predictive functions can assist a self-awareness component (e.g., component 550) to predict an emerging situation of an unplanned PM in cases where the prediction depends on projected values of a set of variables as a function of time.
Optimization knowledge memory (OKM) 1010 contains concepts (e.g., knowledge) related to diagnosis and optimization of the behavior of tool system 310. It should be appreciated that a behavior can include a goal or a sub-goal. Accordingly, OKM 1010 contains domain, or goal, specific concepts such as step, step-data, run, run-data, lot, lot-data, PM-time-interval, wet-clean-cycle, process-recipe, sensor, controller, etc. The latter concepts are associated with a tool system 310 that manufactures semiconductor devices. In addition, OKM 1010 comprises domain independent concepts, which can include a reading (e.g., readings from a pressure sensor in sensor component 325), a sequence, a comparator, a case, a case-index, a case-parameter, a cause, an influence, a causal-dependency, an evidence, a causal-graph, etc. Furthermore, OKM 1010 can comprise a set of functional relations like compare, propagate, rank, solve, etc. Such functional relations can be exploited by autobots, which can reside in autobot component 540 and can confer OKM 1010 at least a portion of its functionality through execution of procedures. Concepts stored in OKM 1010 possess a suitability numeric attribute and an inertia numeric attribute, and a situation score attribute derived there from. The semantics of suitability, inertia and situation score is substantially the same as that for self-awareness component 550 and self-conceptualization component 560. Therefore, if a run-data is provided with a lower inertia than step-data, self-optimization component 570 planners (e.g., überbots) are more likely to communicate the concept of run-data from OMK 1010 to optimizing working memory (OWM) 1020. In turn, such inertia relationship between run-data and step-data can increase the activation rate of optimization autobots that work with run related concepts.
It should be noted that through FF links 552 and 562, self-awareness component 550 and self-conceptualization component 560 can influence the situation score of concepts stored on OKM 1010, and the activation energy of optimization autobots through optimization planners (OPs), which can reside in optimization planner component 1050. It should be appreciated that concepts which are stored in OKM 1010, and are influenced through self-awareness component 550 and self-conceptualization component 560, can determine aspects of a specific goal to be optimized as a function of a specific context. As an illustration, if self-awareness component 550 recognizes that a pattern of data for a process step has degraded significantly, the situation score of the associated step concept can be increased. Accordingly, OPs can then supply additional activation energy to optimizing autobots related to the step concept in order to modify a set of steps executed during a process (e.g., while pursuing a goal). Similarly, if self-conceptualization component 560 identifies a new functional relationship among tool measurements for a product lot, FF information received from self-conceptualization component 560 (via FF 562, for example) self-optimization component 570 can increase (1) a situation score of a lot concept and (2) an activation energy of an optimization autobot with a functionality that relies on a lot concept; therefore, modifying aspects of the lot concept (e.g., number or type of wafers in a lot, cost of a lot, resources utilized in a lot, and so on).
Health assessment of a tool system 310 can be performed through diagnosing engine 825 as discussed next. It should be noted that a health assessment can be a sub-goal of a manufacturing process. Diagnosing engine 825 autonomously creates a dependency graph and allows actor 390 to augment the dependency graph. (Such a dependency graph can be regarded as extrinsic data or as intrinsic data.) The causal graph can be conveyed incrementally, according to the dynamics of the process conducted by the tool system 310, and a diagnosis plan that can be devised by the actor 390. For example, a causal graph can show that a “pressure” malfunction is caused by one of four causes: a deposition chamber has a leak, gas flow into the chamber is faulty, exhaust valve angle (which controls the magnitude of gas flow) is faulty, or a pressure sensor is in error. Components of tool system 310 have a priori probabilities of failure (e.g., a chamber leak can occur with probability 0.01, a gas flow can be faulty with probability 0.005, and so on). In addition, actor 390, or self-conceptualization component 560, can define a conditional dependency for pressure malfunction which can be expressed as a conditional probability; e.g., probability of pressure being at fault given that the chamber has a leak can be p(P|leak). Generally, conditional probabilities causally relating sources of tool failure can be provided by actor 390. It should be noted that autonomous learning system 360 assumes that probability assignments defined by actor 390 can be approximate estimates, which in many cases can be significantly different from a physical probability (e.g., actual probability supported by observations). Examples of causal graphs are presented and discussed next in connection with
Self-optimization component 570 can also comprise a prognostic component 1060 which can generate a set of prognostics regarding performance of tool system 360 through information I/O 358 associated with the tool 360. Such information can comprise quality of materials employed by functional component, physical properties of product assets 328 produced by tool system 360, such as index of refraction, optical absorption coefficient, or magnetotransport properties in cases product assets 328 are doped with carriers, etc. Multiple techniques can be utilized by prognostics component 1060. The techniques comprise first characterization techniques substantially the same as those techniques that can be employed by self-awareness component when processing information 358; namely, such as (i) frequency analysis utilizing Fourier transforms, Gabor transforms, wavelet decomposition, non-linear filtering based statistical techniques, spectral correlations; (ii) temporal analysis utilizing time dependent spectral properties (which can be measured by sensor component 325), non-linear signal processing techniques such as Poincaré maps and Lyapunov spectrum techniques; (iii) real- or signal-space vector amplitude and angular fluctuation analysis; (iv) anomaly prediction techniques and so forth. Information, or data assets generated through analysis (i), (ii), (iii) or (iv) can be supplemented with predictive techniques such as neural-network inference, fuzzy logic, Bayes network propagation, evolutionary algorithms, like genetic algorithm, data fusion techniques, and so on. The combination of analytic and predictive techniques can be exploited to facilitate optimization of tool system 310 via identification of ailing trends in specific assets, or properties, as probed by sensor component 325, as well as information available in OKM 101, with suitable corrective measures generated by optimization planner component 1050, and optimization autobots that can reside in component 540.
To generate the dependency graph 1100 self-conceptualization component 530 can proceed in two steps. (i) Comparator 1120 is introduced as a root node that receives as input a single learnt function 1110. A failure in comparator 1120 implies a failure in a tool (e.g., tool system 310) that employs a biologically based autonomous learning system. A comparator failure can be a Boolean value (e.g., “PASS/FAIL” 1130) result which can be based on comparing a measured value of the pressure with a predicted value generated through learnt function 1110. Self-conceptualization component 530 flags a failure in comparator 1120 when the average difference between predicted pressure values and collected pressure data (e.g., as reported by a pressure sensor residing in sensor component 378) fails to remain within user-specified bounds—e.g., average difference is to remain within 5% of predicted pressure. A failure of comparator 1120 is made dependent on the output of the predictive function 1110. Thus a comparator failure depends on (is influenced by) the failure of the pressure reading (PR 1140); which can fail because a pressure sensor (PS 1143) has failed or a physical pressure (e.g., the physical quantity PP 1146) has failed. Physical pressure PP 1146 can fail because a pressure mechanism (PM 1149) can fail. Thus the system autonomously creates the dependencies between PR 1140 and {PS 1143, PP 1146} and between PP 1140 and {PM 1149}.
(ii) Dependent variables in learnt function 1110 are employed to complete the dependency graph as follows. Physical mechanism PM 1149 can fail when a gas-flow reading (ΦR 1150) fails or a valve-angle reading (θR 1160) fails—dependent variables in learnt function 1110. Thus, self-conceptualization component 530 creates dependencies between PM 1149 and {θR 1150, ΦR 1160}. Substantially the same processing, or reasoning, for a failure in a reading can be employed by self-conceptualization component 530 to create dependencies between ΦR 1150 and {ΦS 1153, ΦP 1156} and between θR 1160 and {θS 1163, θP 1166}. Self-conceptualization component 530 then can add the dependency between ΦP 1156 and { M 1159} and between θP and {θM}. It is to be noted that the relationship between the physical quantity (e.g., PP 1146, ΦP 1156, θP 1166) and the associated mechanism (e.g., PM 1149, ΦM 1159, and θM 1169) is redundant and presented to enhance clarity—mechanism nodes (e.g., nodes 1149, 1159, and 1169) can be removed, and their children made the children of the associated physical magnitude nodes (e.g., nodes 1146, 1156, and 1169).
In a dependency graph such as dependency graph 900, leaf-level nodes are physical points of failure; e.g., nodes 1140, 1143, 1146, and 1149; nodes 1140, 1153, 1156, and 1159; and 1160, 1163, 1166, and 1169. In an aspect, an actor (e.g., actor 390, which can be a user) can supply a biologically autonomous learning system with a priori probabilities for all physical points of failure. Such a priori probabilities can be obtained from manufacturing specifications for the component, field data, MTBF data, etc., or can be generated by simulation of the performance of parts present in a manufacturing tool and involved in a relevant manufacturing processing. The actor can also supply conditional probabilities based on prior experience, judgment, field data, and possible failure modes (e.g., the presence of a first failure can eliminate the possibility of a second failure, or the first failure can increase the probability of occurrence of the second failure, etc.). Upon receiving a priori and conditional probabilities, for example via an interaction component, such as component 340, the autonomous system can use Bayesian network propagation with learning to update the probabilities based on actual failure data submitted to the autonomous system. Thus, in case the initial probabilities provided by the actor are erroneous, the autonomous system adjusts the probabilities as field data contradicts or supports a failure outcome; namely, a PASS or FAIL result of a comparator.
It should be noted that an actor (e.g., actor 390, which can be a user) can add dependencies to an autonomously generated dependency graph (e.g., dependency graph 900) rooted at mechanism failures. Such an addition can be effected, for instance, through interaction manager 355. In an aspect, as an illustration, dependency graph 1100 is augmented with two nodes labeled PLEAK 1170 and PALT 1173 that result in a dependency of PM 1149 on {ΦR 1150, θR 1160, PLEAK 1170, and PALT 1173}. It is to be appreciated that dependency graph 1100 can be augmented with a deeper graph as well. Addition of node PLEAK 1170 informs the autonomous system, through self-conceptualization component 530, that besides a failure of a gas flow reading or a valve angle reading, the pressure mechanism can also fail should a leak be present in the tool. Node PALT 1173 is complementary to node 1170 in that it represents the likelihood that mechanisms alternative to a leak results in system failure. Upon addition of a node, or a deeper graph, the actor is to assign a priori probabilities for the node and associated conditional probabilities describing the dependencies.
It should be appreciated that learnt functions can be more complex than the function P=F(Φ,θ) discussed above, and can include substantially more independent variables; however, causal graphs can be prepared in substantially the same manner.
In order to identify a root cause, e.g., the physical point of failure with the highest probability of failure, a biologically based autonomous learning system can utilize a failure of one or more predictor or recipe comparators to rank all physical points of failure present in a dependency graph. In an aspect, for a complete dependency graph with one or more comparators, the biologically based autonomous learning system can use Bayesian inference to propagate the probabilities given the failure signature of the comparators. Thus the system can compute the probability of failure for a particular PASS/FAIL outcome (e.g., outcome 1198A for comparator A 1195A or outcome 1198B for comparator B 1195B) for each comparator. As an example, suppose that predictor comparator 1120 and recipe comparator A 1195A fail whereas comparator B 1195E passes. The autonomous system can compute the failure probability for each physical point of failure given the comparator failures. (For example what is the probability of the pressure sensor failure given that comparator 1195A and comparator A 1195A fail whereas comparator B 1195E passes). Each point of failure is then ordered from most likely to fail (highest computed probability), or the most likely root cause, to least likely to fail (lowest computed probability). Identification of a root cause, which can be deemed as actionable intelligence (e.g., output 140), can be conveyed to an actor via an interaction manager for further process; e.g., order a new part, request a maintenance service (an actor communicates with or resides in the tool's manufacturer location), download a software update, schedule a new training session, and the like.
In an aspect, autonomous learning system 360 can learn (through learning mechanisms described above in connection with system 300) expected values for the critical output parameters during normal (e.g., non-faulty) group tool 1200 operation. In an aspect, when measured output 1265 deviates from an expected output, autonomous learning system 360 can identify a performance metric of group 1200 performance as degraded. It should be appreciated that the latter assessment can proceed in substantially the same manner as described in connection with single autonomous tool system 300; namely, through a self-awareness component in autonomous learning system 360. It is to be noted that even though autonomous group tool 1200 can present a degraded performance, a subset of autonomous tool system 12201-1220K can provide output that is not degraded and meet individual expectation values for a predetermined metric.
In addition, similarly to the scenario of a single tool system (e.g., tool system 310), autonomous learning system 360 can construct a predictive model for a critical output parameter as a function of individual tool related output parameters. It should be appreciated that such output parameters can be collected through asset 328 input/output. It is to be noted that in group tool 1200, measurements of tool output (e.g., 12601-1260K) can be available to autonomous biologically based learning system 360 via sensor components residing in each of tool systems 12201-1220K, which can be accessed through deployed knowledge network extant in each autonomous learning system (e.g., 360, or 1250).
Furthermore, the autonomous system 360 can also construct a predictive model of group time-to-failure as a function of assets 328 of group 1200; e.g., group input data, group outputs, group recipes, or group maintenance activities. In an aspect, to determine a group time-to-failure, autonomous learning system 360 can gather failure data, including time between detected (e.g., through a set of sensor components) failures, associated assets 12501-1250K, outputs 12601-1260K, and maintenance activities for substantially all operation tools in the set of tools 12201-1220K. (It should be appreciated that as a consequence of prior failure assessments, specific tools (e.g., tool system 212201 and tool system K 1220K) in the set of tools (e.g., tools 12201-1220K) in group 1200 can be out of operation.) Collected data can be autonomously analyzed (e.g., through a processing component 385 in autonomous learning system 360) to learn a predictive function for time-to-failure as a function of the group assets (e.g., inputs, recipes, . . . ), outputs, and maintenance activities. It should be appreciated that the group time-to-failure model constructed from the collected data can readily display substantially dominant factors that impact performance of group tool 1200.
In an aspect, time-to-failure models constructed for individual components of tool systems (e.g., 12201-1220K) in group tool 1200 can be employed by actor 390 (e.g., a group level controller) to optimize part inventory and optimize maintenance scheduling. It should be appreciated that such optimization can be conducted, at least in part, by autonomous system 360. For example, the autonomous system accesses the MES (or ERP) system to identify the number of available parts. When a set of parts that provide functionality to tool systems 12201-1220K (e.g., parts in one or more of components within a functional component like a component 315 in system 310), and can be expected to be necessary (e.g., for replacement) within a specific time period Δτ, exceeds an available supply in stock, additional parts can be ordered. Alternatively, or in addition, when parts are available, an expected schedule of necessary parts can be analyzed to determine an optimal, or adequate, time to place a new order.
It should be appreciated that maintenance schedules can be reassessed and optimized during a necessary, previously scheduled, maintenance activity, in order to exploit an opportunity available to autonomous system 360 to analyze parts and identify parts that can fail in a substantially short period of time. It should further be appreciated that a group or individual time-to-failure schedule can be complemented, autonomously in an aspect, with additional information such as cost of parts, time to replace parts, and so forth, to determine whether replacement of a part during a current maintenance cycle is beneficial with respect to the replacement of the part in a forthcoming scheduled maintenance cycle. It is noted that autonomous system 360 can also take as input various costs associated with the operation of group tool 1200 in order to compute a cost per output product (e.g., a wafer, a car, a computer, etc.) for the group, and a total cost to produce a specific order during operation of the group tool 1200. After building a model of cost as a function of individual tool assets 12501-1250K (e.g., recipes), outputs 12601-1260K, and maintenance activities, autonomous system 360 can rank individual tool systems 12201-1220K in increasing order of operation cost. A combined cost data asset can be utilized construct a predictive model of cost versus assets, outputs, and maintenance activities associated with the individual tool systems—for example, such an assessment can identify operational assets and variables that affect substantially an operation or maintenance cost for the group tool. In an aspect, autonomous system 360 can utilize available historic data assets to redesign a production line, or equipment configuration in a floor plant, in order to minimize costs. In addition, during such an optimization process, autonomous system 360 can rely on shutdown of various tool systems in order to exploit alternative patterns of operation. Furthermore, autonomous system 360 can utilize cost-benefit analysis to determine a set of trade-off scenarios in which production of specific output proceeds without output for specific, highly costly tool systems.
Tools system 12201-1220K can be substantially the same, or can be disparate (e.g., tool systems 12201-12203 are steppers, tool 1220j is a stepper, and 1220K-4-1220K are turbomolecular vacuum pumps). Typically, a central difference amongst homogeneous (e.g., tool systems are alike) and heterogeneous (e.g., tools are disparate) can lie in that input and output measurements (e.g., measurement assets) are distinct. For example, a critical output of interest for tool group 1200 can be D1 CD uniformity, but a coating system that is part of the group tool 1200 can fail to provide such output measurements. Accordingly, autonomous system 360 can construct a model for expressing a tool group's outputs as a function of individual tool (e.g., 12201-1220K) outputs. Thus, when a group performance appears degraded, individual performances associated with individual tools can be analyzed to isolate a tool that has the largest weight in causing the performance degradation.
Conglomerate system 1310 can be autonomously supported by an autonomous learning system comprising an interaction component 340, an actor 390, and an autonomous learning system 360. In an aspect, autonomous support can be directed toward improving an overall fabrication effectiveness (OFE) metric of output assets (e.g., output 1365 or 1265). In addition, each of the autonomous tool conglomerates 13201-1320Q can be in turn autonomously supported by an interaction component 1330, and an autonomous learning system 1340. Interface component 1330 facilitates interaction between autonomous learning system 1340 and actors 3901-390Q. Functionality of each of such components is substantially the same as the functionality of respective component described above in connection with system 360 and system 1200. Information 13481 (I=1, 2, . . . , Q) communicated among interaction component 1330 and autonomous system 1340 is associated with the respective autonomous tool conglomerate I 13201. Similarly, assets 13501 conveyed to and received from an autonomous tool conglomerate I 13201 are specific thereof.
To address performance in an autonomous tool conglomerate 13101-1310Q, the multi-step characteristics of a fabrication process can be incorporated through a performance tag that identifies products utilizing a composite conglomerate index Ca, wherein the index a indicates a specific tool group within conglomerate C (e.g., autonomous conglomerate 1320Q), and a run index (R); thus, a product quality, or performance metric associated with a specific product is identified via a label (Ca;R), which can be termed “group-layer output.” Such label facilitates identifying each autonomous operation group as an individual component Ca. Therefore, autonomous system 360 can map quality and performance metrics as a function of fabrication conglomerate (e.g., autonomous tool conglomerate 13102) and as a function of tool group within each fabrication conglomerate. The latter facilitates root-cause analysis of poor performance or quality, by first identifying a conglomerate (e.g., a fabrication facility) and subsequently performing the analysis for the tool associated with the assessed degradation. It should be appreciated that index Ca to account for the fact that output assets generated in an autonomous system comprised of multiple conglomerate tools can be transported from a first conglomerate (N) to a second conglomerate (N′). Thus, the composite symbol for tracking performance associated with a transfer of assets (e.g., as a part of a multi-step fabrication process) can read Cα;N→N′.
Performance of an autonomous tool conglomerate can be performed as a function of product yield. Such yield is utilized to rank disparate conglomerates. In an aspect, autonomous learning system 360 can develop a model for yield based at least in part on output assets from each autonomous tool, or autonomous group tool. For example, for tools, or group of tools, employed in semiconductor manufacturing, yield can be expressed as a function of a wafer thickness, a device uniformity, an impurity (e.g., extrinsic and intrinsic dopant concentration) concentration, a DI CD, an FI CD, and so on. Moreover, other yield metrics can be utilized to determine a model for a yield, specially in an autonomous learning systems comprising tool conglomerates systems (e.g., 13201-1320Q) wherein output assets can be transported among conglomerates: an overall equipment efficiency (OEE), a cycle time efficiency, an on-time-delivery rate, a capacity utilization rate, a rework rate, a mechanical line yield, a probe yield and final test yield, an asset production volume, a startup or ramp-up performance rate, etc. It is to be noted that an autonomous system that supports operation of a set of autonomous tool conglomerates can autonomously identify relationships amongst yield metrics in order to redesign processes or communicate with actors 3901-390Q with respect to adjustments in connection to said yield metrics.
The yield function mentioned supra can be analyzed through a combination of static and dynamic analysis (e.g., simulation) to rank group layer outputs according to degree of influence, or weight, in leading to a specific yield. It is to be noted that ranking tools, group of tools, or conglomerates, at a group-layer-output level based at least in part on influence in affecting asset output, or yield, can afford a group or conglomerate autonomous learning system 360 to autonomously identify, through autonomous systems associated with each of the tools in a group or group in a conglomerate, whether a specific tool can be isolated as a dominant tool in yield deterioration. When such a tool is located, the group or conglomerate level autonomous system 360 can issue an alarm to a maintenance department with information regarding ranking the faults that can be candidates for performance degradation.
In addition, yield for the lowest ranking autonomous tool conglomerate can be employed to identify the group layer outputs of the tool group that is dominant in its impact on yield. The time-to-failure for such tool-group can be compared with substantially the same tool groups in disparate autonomous conglomerates in order to identify cause(s) of poor performance. Furthermore, an autonomous tool conglomerate system ranks tools within a specific tool group in disparate tool conglomerates. It is to be noted that an autonomous learning system that supports and analyzes a group of autonomous tool conglomerates (e.g., 13201-1320Q) can rank each of the conglomerates according to inferred time-to-failure for each conglomerate. Since time-to-failure can change over operational time intervals in view of, e.g., input/output asset (e.g., asset 358) load, a database with time-to-failure projection can be updated at specified periods of time (e.g., weekly, monthly, quarterly, or yearly).
Further yet, when an individual tool that is primarily responsible for a group tool's poor performance (e.g., the tool ranks the lowest in performance within a group tool, such as a tool that most frequently fails to output assets with specified target properties of quality like uniform doping concentration or uniform surface reflection coefficient) is identified, an autonomous system associated with the lowest performing tool, or with the conglomerate system that includes such poor performing tool, can analyze the tool's outputs to identify those outputs that most significantly affect the output of the lowest performing group. For example, a tool in a tool group or conglomerate that outputs assets with low uniformity as illustrates above, can lead to a substantial percentage (e.g., 60%) of tool groups uniformity variation (for example, variation in uniformity change of surface reflectivity of an optical display due to uniformity issues on surface reflectivity of coatings on otherwise high-quality displays). To that end, in an aspect, for each output in the group the tool autonomous system constructs a function that expresses tool output as a function of tool assets (e.g., inputs, recipes, and process parameters, tool operator or actor, and so on). This model is then analyzed to identify the dominant factors in poor performance. It is to be noted that an autonomous system can identify best performing tools in a group tool and analyze causes that result in the tool having the best performance; e.g., the vacuum level of the tool during operation is consistently lower than vacuum level of disparate tools in the group tool, or during epitaxial deposition a wafer in the best performing tool rotates at a lower speed than in disparate tool carrying out a deposition, thus the tool consistently achieves greater device quality. Such factors in highest ranking and lowest ranking tools can be compared with same parameters in other tools in conglomerate system. In case the comparison indicates that the factors identified as the root causes of highest and lowest ranking performance appear to be substantially the same throughout the tool conglomerate system, then a new model can be developed and alternative root causes can be identified. Such iterative, autonomous processes of model development and validation can continue until root causes are identified and best practices are emulated (e.g., a coating recipe utilized in tool conglomerate 1320p is adopted in substantially all tool conglomerates in view that it increases output asset performance by a specific, desirable margin) and root causes for low performance are mitigated (e.g., abandoning a specific brand of paint whose viscosity at the operating temperature of a painting tunnel results in non-uniform coloration of painted products). Ranking of tools, group of tools, or conglomerate of tools is autonomous and proceeds in substantially the same manner as in a single autonomous tool system (e.g., system 360). Autonomous systems that support operation of a conglomerate of autonomous tools considers such autonomous conglomerates as a single component regardless of the complexity of its internal structure, which can be accessed and managed through an autonomous system associated with the conglomerate.
In backward chaining, action flow (e.g., process flow 1530) which leads to an output typically counters a probe flow (e.g., assessment flow 1540) which typically assesses the action flow. Thus, assessment generally takes place in a top-bottom manner, in which assessment is conducted on a high-level stage of a specific action, e.g., a finalized asset output 1520, and proceeds to lower-level stages in a quest to focus the assessment on a specific stage prior to completion of a specific action. As applied by autonomous system 1504, output asset 1520 is received via process station N 1510N. The autonomous system 1504 can evaluate, as illustrated by 1546, a set of performance metrics {P(C)N-1→N} leading to a specific degradation vector (not shown), based at least in part on a expected performance, for substantially all operational components (e.g., tool, group or conglomerate tool) in the process station 1510N. Additionally, it should be appreciated that in process 1530, output assets (e.g., assets 1515) can be transported across disparate geographical areas, therefore the degradation vector assessed by autonomous system 1504 can comprise metrics associated with the in-transit portion of the process that leads to a partially finished asset 1515. For example, when process 1530 regards accelerometers for vehicular airbag deployment, mechanical pieces in a transported accelerometer can be damaged as a consequence of utilizing an alternative route for transporting the accelerometers rather than employing a route disclosed in the process 1530. When result(s) 1549 of such an assessment indicate that N-station output 1520 is faulty, autonomous system 1504 isolates a faulty tool, or group of tools, associated with process station N, and generates a report (e.g., assessment report 1550, repair(s) report 1560, or maintenance schedule 1570). The generated report(s) can contain information to be utilized by one or more actors (e.g., actors 3901-390Q). In addition, reports can be stored to create a legacy of solutions (or “fixes”) for specific issues with performance, especially issues that appear infrequently so that an actor's intervention can be preferred with respect to an autonomously developed solution which typically can benefit from extensively available data. Moreover, availability of reports can facilitate failure simulations or forensic analysis of a failure episode, which can reduce manufacturing costs in at least two levels: (a) costly, infrequently failing equipment can be predicted to fail under rare conditions, which can be simulated by autonomous system 360, arising from operation of equipment by an actor with a background non-commensurate with the complexity of the equipment, (b) optimization of parts inventory through prediction of various failure scenarios based at least in part on historical data stored in assessment reports 1550 and repair reports 1560.
In case results 1549 of process station N 1510N yield no faulty tool, or group of tools, assessment is performed on a lower-level process station N−1 1510N-1 that generates a partially processed output asset 1515, and is a part in the process cycle 1530 to generate output 1520. Through analysis of a set of disparate performance metrics {P(C)N-2→N-1}, a degree of degradation can be extracted and associated tool, or group of tools (e.g., conglomerate C) can be located. In instances that no faulty conglomerate of autonomous tools, or group of autonomous tools, or individual autonomous tool, autonomous system 1504 continues the backward, top-bottom assessment flow 1540 with the object to locate sources of poor performance in final output 1520.
Additionally, in an aspect, management component 1635 can access (i) an asset store 1683, which typically contains assets scheduled to be distributed or assets that have been distributed; (ii) a partner store 1686 comprising commercial partners associated in the distribution or completion of specific assets; (iii) a customer store 1689 which can contain current, past, or prospective customers to which the selected asset has been, or can be distributed; (iv) a policy store that can determine aspects associated to the distribution of assets, such as licensing, customer support and relationships, procedures for asset packaging, scheduling procedures, enforcement of intellectual property rights, and so on. It should be appreciated that information contained in policy store can change dynamically based at least in part on knowledge, e.g., information asset, learned or generated by autonomous biologically based learning system.
Once an asset has been packaged, which can include adding to a package a monitoring device like RFID tags, either active or passive, or bar codes (e.g., two-dimensional codes, Aztec codes, etc.), and it has been scheduled for distribution, a record of distribution can be stored, or if the asset is a data asset then a copy of the asset can be stored. Then, the asset can be delivered to a disparate autonomous tool conglomerate P 1320P.
Conglomerate 1730 can receive the data asset and, as a non-limiting example, initiates a deposition process to fabricate a solid-state device according to the received asset. To that end, conglomerate 1730 can partner with conglomerate 1740 and both can be regarded as fabrication facilities that are part of a two-conglomerate autonomous conglomerate tool 1310. Such conglomerates can produce multiple devices according to the received specification asset, once a device is fabricated it can be tested, and assigned a quality and performance metric, such metrics can lead to backward chaining to located “poor performers” among the autonomous tools that enter conglomerates 1730 and 1740. Through determination of multiple metrics, it is possible to autonomously adjust the operation of conglomerates 1720 and 1730 to optimize production of the device, or output asset. It is noted that link 1724 indicates an internal link, wherein conglomerates 1730 and 1740 are part of a same fabrication plant; thus the asset can be transported in substantially different conditions than when utilizing link 1724 which provides a vehicular transportation route. Link 1744 can be employed to ship devices for commercial packaging in a disparate geographic location (such transportation can be motivated by advantageous packaging costs, skillful labor, corporate tax incentives, and so on). It should be appreciated that an autonomous learning system at conglomerate 1740 can optimize the shipping times (via a scheduler, for example) and routes (e.g., link 1744) in order to ensure timely and cost effective delivery. At conglomerate 1750 assets are packed and remotely tested, via a wireless link, in conglomerate 1760. In an aspect, the volume of devices tested and the lots from which devices are tested can be determined by an autonomous system in conglomerate 1760. Once packed devices have been approved for commercialization, the assets are shipped through road link 1744 ato conglomerate 1740, and subsequently shipped via road link 1770 to a disparate class of conglomerate 1775. Such conglomerate can be a partner vendor, and conglomerate 1775 can be storage wharehouse, which can be considered a tool group conglomerate. Such conglomerate is linked, internally, to conglomerate 1765 which can be a showroom for the received assets.
In view of the example systems presented and described above, a methodology that may be implemented in accordance with the disclosed subject matter, will be better appreciated with reference to the flowchart of
In a further aspect, the received data can be associated with data types or with procedural, or functional, units. A data type is a high level abstraction of actual data; for instance, in an annealing state in the tool system a temperature can be controlled at a programmed level during the span of the annealing cycle, the time sequence of temperature values measured by a temperature sensor in the tool system can be associated a sequence data type. Functional units can correspond to libraries of received instructions, or processing code patches that manipulate data necessary for the operation of the tool or for analyzing data generated by the tool. Functional units can be abstracted into concepts related to the specific functionality of the unit; for example, a multiplication code snippet can be abstracted into a multiply concept. Such concepts can be overloaded, in that a single concept can be made dependent on a plurality of data types, such as multiply(sequence), multiply(matrix), or multiply(constant, matrix). Moreover, concepts associated with functional units can inherit other concepts associated with functional units, like derivative(scalar_product(vector, vector)) which can illustrate a concept that represents a derivative of a scalar product of two vectors with respect to an independent variable. It should be appreciated that functional concepts are in direct analogy with classes, which are in themselves concepts. Furthermore, data types can be associated a priority and according to the priority can be deposited in a semantic network. Similarly, functional concepts (or autobots), can also be associated with a priority, and deposited in a disparate semantic network. Concept priorities are dynamic, and can facilitate concept activation in the semantic networks.
At act 1830 knowledge is generated from the received data, which can be represented in semantic networks, as discussed above. Generation of knowledge can be accomplished by propagating activation in the semantic networks. Such propagation can be determined by a situation score assigned to a concept in addition to a score combination. In an aspect, score combination can be a weighted addition of two scores, or an average of two or more scores. It should be appreciated that a rule for score combination can be modified as necessary, depending on tool system conditions or information input received from an external actor. It should be appreciated that a priority can decay as time progresses to allow concepts that are seldom activated to became obsolete, allowing new concepts to become more relevant.
The generated knowledge can be complete information; for instance, a steady-state pressure in a deposition step is a precise, well-defined mathematical function (e.g., a single-valued function with all parameters that enter the function deterministically assessed, rather than being stochastic or unknown) of two independent variables like steady-state flow and steady state exhaust valve angle. Alternatively, the generated knowledge can represent a partial understanding; for example, an etch rate can be possess a known functional dependence on temperature (e.g., an exponential dependence), yet the specific relationship—e.g., precise values of parameters that determine the functional dependence—between etch rate and temperature is unknown.
At act 1840 the generated knowledge is stored for subsequent utilization of for autonomous generation of further knowledge. In an aspect, knowledge can be stored in a hierarchy of memories. A hierarchy can be determined on the persistence of knowledge in the memory and the readability of knowledge for creation of additional knowledge. In an aspect, a third tier in the hierarchy can be an episodic memory (e.g., episodic memory 530), wherein received data impressions and knowledge can be collected. In such a memory tier manipulation of concepts is not significant, the memory acting instead as a reservoir of available information received from a tool system or an external actor. In an aspect, such a memory can be identified as a metadatabase, in which multiple data types and procedural concepts can be stored. In a second tier, knowledge can be stored in a short term memory wherein concepts can be significantly manipulated and spread activation in semantic networks can take place. In such a memory tier, functional units or procedural concepts operate on received data, and concepts to generate new knowledge, or learning. A first tier memory can be a long term memory (e.g., LTM 510) in which knowledge is maintained for active utilization, with significant new knowledge stored in this memory tier. In addition, knowledge in a long term memory can be utilized by functional units in short term memory.
At act 1850 the generated or stored knowledge is utilized. Knowledge can be employed to (i) determine a level of degradation of a goal component (e.g., tool system 310) by identifying differences between stored knowledge and newly received data (see self-awareness component 550), wherein the received data can be extrinsic (e.g., input 130) or intrinsic (e.g., a portion of output 140); (ii) characterize either extrinsic or intrinsic data or both, for example by identifying data patterns or by discovering relationships among variables (such as in a self-conceptualization component 560), wherein the variables can be utilized to accomplish the established goal; or (iii) generate an analysis of the performance of the tool system that generates the data (e.g., self-optimization component 570), providing indications of root cause for predicted failures or existing failures as well as necessary repairs or triggering alarms for implementing preventive maintenance before degradation of the tool system causes tool failure. It is to be noted that utilization of the stored and generated knowledge is affected by the received data—extrinsic or intrinsic—and the ensuing generated knowledge.
Act 1860 is a validation act in which the degree of accomplishment of a goal can be inspected in view of generated knowledge. In case the established goal is accomplished, example method 1800 can end. Alternatively, if the established goal has not been accomplished, the established goal can be reviewed at act 1870. In the latter, flow of method 1800 can lead to establishing a new goal in case a current goal is to be revised or adapted; for instance, goal adaptation can be based on generated knowledge. In case no revision of a current goal is to be pursued, flow of method 1800 is returned to generate knowledge, which can be utilized to continue pursuing the currently established goal.
At act 2130 the processed asset is distributed. Distribution typically depends on the asset features and characteristics, as well as on the destination of the asset. For example, assets can be distributed within a factory plant, in order to complete asset production like in an assembly line wherein an unfinished vehicle (e.g., an asset) can be transported through different stages of assembly. Similarly, in the food industry, a frozen meal (e.g., asset) is distributed throughout a food preparation plant. Alternatively, or in addition, depending on industry, an unfinished asset can be distributed to overseas to be finished in order to benefit from cost-effective production markets.
At act 2140, an distributed asset is monitored in order to ensure, for example, the asset distribution adheres to applicable distribution regulation, or to ensure adequate inventory replenishment by having access to distribution status of the asset. In addition, monitoring distribution of the asset can mitigate losses and damages, as well as can facilitate interaction with commercial partners and customers.
Various aspects or features described herein may be implemented as a method, apparatus, or article of manufacture using standard programming and/or engineering techniques. The term “article of manufacture” as used herein is intended to encompass a computer program accessible from any computer-readable device, carrier, or media. For example, computer readable media can include but are not limited to magnetic storage devices (e.g., hard disk, floppy disk, magnetic strips . . . ), optical disks [e.g., compact disk (CD), digital versatile disk (DVD) . . . ], smart cards, and flash memory devices (e.g., card, stick, key drive . . . ).
What has been described above includes examples of the claimed subject matter. It is, of course, not possible to describe every conceivable combination of components or methodologies for purposes of describing the claimed subject matter, but one of ordinary skill in the art may recognize that many further combinations and permutations of the claimed subject matter are possible. Accordingly, the claimed subject matter is intended to embrace all such alterations, modifications and variations that fall within the spirit and scope of the appended claims. Furthermore, to the extent that the term “includes” is used in either the detailed description or the claims, such term is intended to be inclusive in a manner similar to the term “comprising” as “comprising” is interpreted when employed as a transitional word in a claim.
The subject application is related to co-pending, and commonly assigned, U.S. patent application Ser. No. ______, entitled “AUTONOMOUS ADAPTIVE SEMICONDUCTOR MANUFACTURING,” filed on Mar. 8, 2008. The entirety of this application is incorporated herein by reference.