Patent Application
20210166121
This patent application is based on and claims priority to Japanese Patent Application No. 2019-217440 filed on Nov. 29, 2019, the entire contents of which are incorporated herein by reference.
The present disclosure relates to a predicting device, a predicting method, and a predicting computer program product.
Conventionally, in the field of various manufacturing processes, by managing the number of objects processed or a cumulative amount of treatment time, estimation is performed with respect to various items, such as a state in a manufacturing apparatus. Based on a result of the estimation, prediction of replacement time of each part, prediction of timing of maintenance of the manufacturing apparatus, and the like are performed.
Meanwhile, during the manufacturing process, various data are measured along with processing of the objects, and a set of the measured data (a set of multiple types of time series data; hereinafter referred to as a “time series data set”) includes data necessary for the estimation regarding the items to be estimated.
The present disclosure provides a predicting device, a predicting method, and a predicting program utilizing time series data sets measured during processing of an object in a manufacturing process.
A predicting device according to one aspect of the present disclosure includes a processor, and a non-transitory computer readable medium that has stored therein a computer program that, when executed by the processor, configures the processor to acquire one or more time series data sets measured along with processing of an object at a predetermined unit of process in a manufacturing process performed by a manufacturing device, and to acquire device state information acquired when the object is processed; and apply the one or more time series data sets in a neural network to develop a trained model. The neural network includes a plurality of network sections each configured to process the acquired time series data sets and the device state information, and a concatenation section configured to combine output data output from each of the plurality of network sections as a result of processing the acquired time series data sets, and to output, as a combined result, a result of combining the output data output from each of the plurality of network sections. The computer program further configures the processor to compare the combined result with a quality indicator to train the trained model such that the combined result output from the concatenation section progressively approaches the quality indicator.
Hereinafter, embodiments will be described with reference to the drawings. For substantially the same components in the present specification and drawings, overlapping descriptions are omitted by giving the same reference numerals.
First, the overall configuration of a manufacturing process (a semiconductor manufacturing process in the present embodiment) and a system including a predicting device will be described.
In the semiconductor manufacturing process, an object (e.g., wafer before processing 110) is processed at a predetermined unit of process 120 to produce a result (e.g., wafer after processing 130). The unit of process 120 described herein is a specialized term related to a particular semiconductor manufacturing process performing in a processing chamber, and details will be described below. Also, a wafer before processing 110 refers to a wafer (substrate) before being processed at the chamber(s) that perform unit of process 120, and wafer after processing 130 refers to a wafer (substrate) after being processed in the chamber (s) that that perform the unit of process 120.
The time series data acquiring devices 140_1 to 140_n each acquire time series data measured along with processing of the wafer before processing 110 at the unit of process 120. The time series data acquiring devices 140_1 to 140_n each measure different properties. It should be noted that the number of measurement items that each of the time series data acquiring devices 140_1 to 140_n measures may be one, or more than one. The time series data measured in accordance with the processing of the wafer before processing 110 includes not only time series data measured during the processing of the wafer before processing 110 but also time series data measured during preprocessing or post-processing of the wafer before processing 110. These processes may include preprocessing and post-processing performed without a wafer (substrate).
The time series data sets acquired by the time series data acquiring devices 140_1 to 140_n are stored in a training data storage unit 163 (a non-transitory memory device) in the predicting device 160, as training data (input data in the training data).
When a wafer before processing 110 is processed at the unit of process 120, device state information is acquired, and the device state information is stored, as training data (input data), in the training data storage unit 163 of the predicting device 160, in association with the time series data sets. Examples of the device state information include:
accumulated data, such as
information indicating deterioration of various parts (e.g., F/R, C/R, cell, electrode, and the like) of the semiconductor manufacturing device;
information indicating deterioration of members (e.g. inner walls) in a processing space (e.g., chamber) of the semiconductor manufacturing device; and
information such as thickness of deposits that have formed on the parts in the semiconductor manufacturing device.
The device state information is managed for each item individually, and the device state information is reset when parts are replaced or when cleaning is performed.
When a wafer before processing 110 is processed at the unit of process 120, a quality indicator is acquired and stored in the training data storage unit 163 of the predicting device 160 as the training data (correct answer data, or ground truth data) in association with the time series data set. The quality indicator is information representing a result (quality) of the semiconductor manufacturing process, and may be any value that reflects a result or a state of the processed object (wafer) or a result or a state of the processing space, such as an etch rate, CD, film thickness, film quality, or number of particles. The quality indicator may be a value measured directly, or may be a value obtained indirectly (i.e., estimated value).
A predicting program (code that is executed on a processor to implement the algorithms discussed herein) is installed in the predicting device 160. By executing the predicting program, the predicting device 160 functions as a training unit 161 and an inference unit 162.
The training unit 161 performs machine learning using the training data (time series data sets acquired by the time series data acquiring devices 140_1 to 140_n, and the device state information and the quality indicator associated with the time series data sets) to develop a trained model.
Specifically, the training unit 161 processes the time series data sets and the device state information (input data) using multiple network sections, and performs machine learning with respect to the multiple network sections such that a result of combining output data output from the multiple network sections approaches the quality indicator (correct answer data).
The inference unit 162 inputs device state information and time series data sets acquired by the time series data acquiring devices 140_1 to 140_n along with processing of a new object (wafer before processing) at the unit of process 120, to the multiple network sections to which machine learning has been applied. Accordingly, the inference unit 162 infers the quality indicator based on the device state information and the time series data sets acquired along with the processing of the new wafer before processing.
In the inference unit 162, the time series data sets are input repeatedly while changing a value of the device state information, to infer the quality indicator for each of the values of the device state information. The inference unit 162 specifies a value of the device state information when the quality indicator reaches a predetermined threshold. Thus, according to the inference unit 162, it is possible to accurately predict replacement time of parts in the semiconductor manufacturing device, maintenance timing of the semiconductor manufacturing device, and the like. Once trained by the training unit 161, the interference unit embodies a learned model that is able to accurately replacement time for parts, maintenance timing, and/or process adjustments based on age and/or use of equipment. Thus, the trained model can be used to control/adjust semiconductor manufacturing equipment to and the process steps used to make the produced object. While the term “unit” is used herein for devices such as the training unit and the inference unit, it should be understood that the term “circuitry” may be used as well (e.g., “training circuitry” or “inference circuitry”). This is because the circuit device(s) that execute the operations implemented as software code and/or logic operations are configured by the software code and/or logic operations to execute the algorithms described herein.
As described above, the predicting device 160 according to the present embodiment estimates the quality indicator based on the time series data sets acquired along with processing of an object, and predicts replacement time of each part or maintenance timing of the semiconductor manufacturing device based on the estimated quality indicator. This improves the accuracy of the prediction as compared to a case in which replacement time of each part or maintenance timing of the semiconductor manufacturing device is predicted based on only the number of objects processed or cumulative values of processing time and the like.
In addition, the predicting device 160 according to the present embodiment processes time series data sets acquired along with processing of an object, by using multiple network sections. Accordingly, it is possible to analyze time series data sets at a predetermined unit of process in a multifaceted manner, and it is possible to realize a higher inference accuracy as compared to a case, for example, in which time series data sets are processed using a single network section.
Next, the predetermined unit of process 120 in the semiconductor manufacturing process will be described.
Time series data sets measured in accordance with processing of the wafer before processing 110 in the unit of process 120 of
a time series data set output in accordance with a wafer process performed in the chamber A (first processing space),
a time series data set output in accordance with a wafer process performed in the chamber B (second processing space), and
a time series data set output in accordance with a wafer process performed in the chamber C 10. (third processing space).
Meanwhile,
Further, in reference to
A diagram (a) of
In the unit of process 120 of the time-diagram (a) in
The time-diagram (a) in
In contrast, a diagram (b) of
Further, in the unit of process 120 of the time-diagram (b) in
Next, the hardware configuration of the predicting device 160 will be described.
The predicting device 160 further includes an auxiliary storage device 405, a display device 406, an operating device 407, an interface (I/F) device 408, and a drive device 409. Each hardware element in the predicting device 160 is connected to each other via a bus 410.
The CPU 401 is an arithmetic operation processing device that executes various programs (e.g., predicting program) installed in the auxiliary storage device 405.
The ROM 402 is a non-volatile memory that functions as a main memory unit. The ROM 402 stores programs and data required for the CPU 401 executing the various programs installed in the auxiliary storage device 405. Specifically, the ROM 402 stores a boot program such as BIOS (Basic Input/Output System) or EFI (Extensible Firmware Interface).
The RAM 403 is a volatile memory, such as a DRAM (Dynamic Random Access Memory) or an SRAM (Static Random Access Memory), and functions as a main memory unit. The RAM 403 provides a work area on which the various programs installed in the auxiliary storage device 405 are loaded when the various programs are executed by the CPU 401.
The GPU 404 is an arithmetic operation processing device for image processing. When the CPU 401 executes the predicting program, the GPU 404 performs high-speed calculation of various image data (i.e., the time series data sets in the present embodiment) by using parallel processing. The GPU 404 includes an internal memory (GPU memory) to temporarily retain information needed to perform parallel processing of the various image data.
The auxiliary storage device 405 stores the various programs (computer executable code) and various data used when the various programs are executed by the CPU 401. For example, the training data storage unit 163 is implemented by the auxiliary storage device 405.
The display device 406 displays an internal state of the predicting device 160. The operating device 407 is an input device used by an administrator of the predicting device 160 when the administrator inputs various instructions to the predicting device 160. The I/F device 408 is a connecting device for connecting and communicating with a network (not illustrated).
The drive device 409 is a device into which a recording medium 420 is loaded. Examples of the recording medium 420 include a medium for optically, electrically, or magnetically recording information, such as a CD-ROM, a flexible disk, and a magneto-optical disk. In addition, examples of the recording medium 420 may include a semiconductor memory or the like that electrically records information, such as a ROM, and a flash memory.
The various programs installed in the auxiliary storage device 405 are installed when, for example, a recording medium 420 distributed is loaded into the drive device 409 and the various programs recorded in the recording medium 420 are read out by the drive device 409. Alternatively, the various programs installed in the auxiliary storage device 405 may be installed by being downloaded via a network (not illustrated).
Next, training data that is read out from the training data storage unit 163 when the training unit 161 performs machine learning will be described.
The “APPARATUS” field stores an identifier indicating a semiconductor manufacturing device (e.g., semiconductor manufacturing device 200) whose quality index is monitored. The “RECIPE TYPE” field stores an identifier (e.g., process recipe I) indicating a process recipe, which is performed when a corresponding time series data set is measured, among process recipes performed in the corresponding semiconductor manufacturing device (e.g., EqA).
The “TIME SERIES DATA SET” field stores time series data sets measured by the time series data acquiring devices 140_1 to 140_n when processing according to the process recipe indicated by the “RECIPE TYPE” is performed in the semiconductor manufacturing device indicated by the “APPARATUS”.
The “DEVICE STATE INFORMATION” field stores device state information that is acquired just after the corresponding time series data sets (for example, time series data set 1) are measured by the time series data acquiring devices 140_1 to 140_n.
The “QUALITY INDICATOR” field stores a quality indicator acquired just after the corresponding time series data sets (for example, time series data set 1) are measured by the time series data acquiring device 140_1 to 140_n.
Next, specific examples of the time series data sets measured by the time series data acquiring devices 140_1 to 140_n will be described.
In contrast,
Specifically, the time series data acquiring devices 140_1 to 140_n may acquire time series data measured during preprocessing, as the time series data set 1. The time series data acquiring devices 140_1 to 140_n may acquire time series data measured during wafer processing, as the time series data set 2. Further, the time series data acquiring devices 140_1 to 140_n may acquire time series data measured during post-processing, as the time series data set 3.
Alternatively, the time series data acquiring devices 140_1 to 140_n may acquire time series data measured during processing in accordance with the process recipe I, as the time series data set 1. The time series data acquiring devices 140_1 to 140_n may acquire time series data measured during processing in accordance with the process recipe II, as the time series data set 2. Further, the time series data acquiring devices 140_1 to 140_n may acquire time series data measured during processing in accordance with the process recipe III, as the time series data set 3.
Next, the functional configuration of the training unit 161 will be described.
The branch section 710 is an example of an acquisition unit, and reads out time series data sets and device state information associated with the time series data sets from the training data storage unit 163.
The branch section 710 controls input to the network sections of the first network section 720_1 to the M-th network section 720_M, so that the time series data sets and the device state information are processed by the network sections of the first network section 720_1 to the M-th network section 720_M.
The first to M-th network sections (720_1 to 720_M) are configured based on a convolutional neural network (CNN), which include multiple layers.
Specifically, the first network section 720_1 has a first layer 720_11, a second layer 720_12, . . . , and an N-th layer 720_1N. Similarly, the second network section 720_2 has a first layer 720_21, a second layer 720_22, . . . , and an N-th layer 720_2N. Other network sections are also configured similarly. For example, the M-th network section 720_M has a first layer 720_M1, a second layer 720_M2, . . . , and an N-th layer 720_MN.
Each of the first to N-th layers (720_11 to 720_1N) in the first network section 720_1 performs various types of processing such as normalization processing, convolution processing, activation processing, and pooling processing. Similar types of processing are performed at each of the layers in the second to M-th network sections (720_2 to 720_M).
The concatenation section 730 combines each output data output from the N-th layers (720_1N to 720_MN) of the first to M-th network sections (720_1 to 720_M), and outputs a combined result to the comparing section 740. Similar to the network sections (720_1 to 720_M), the concatenation section 730 may be configured to be trained by machine learning. The concatenation section 730 may be implemented as a convolutional neural network or other type of neural network.
The comparing section 740 compares the combined result output from the concatenation section 730, with the quality indicator (correct answer data) read out from the training data storage unit 163, to calculate error. The training unit 161 performs machine learning with respect to the first to M-th network sections (720_1 to 720_M) and the concatenation section 730 by error backpropagation, such that error calculated by the comparing section 740 satisfies the predetermined condition.
By performing the machine learning, model parameters of each of the first to M-th network sections 720_1 to 720_M and the model parameters of the concatenation section 730 are optimized to predict device state information for adjustment of processes used in the manufacture of a processed substrate.
Next, details of the processing performed in each part (in particular, the branch section) of the training unit 161 will be described with reference to specific examples.
(1) Details of Processing (1) Performed in the Branch Section
First, the processing of the branch section 710 will be described in detail.
The branch section 710 also generates time series data set 2 (second time series data set) by processing the time series data sets measured by the time series data acquiring devices 140_1 to 140_n in accordance with a second criterion, and inputs the time series data set 2 into the second network section 720_2.
The branch section 710 inputs the device state information to one of the first layer 720_11 to the N-th layer 720_1N in the first network section 720_1. Within the layer to which the device state information is entered by the branch section 710, the device state information is combined with a signal to which the convolution processing is applied. It is more preferable that the device state information is input to a layer that is positioned closer to the branch section 710 among the layers (720_11 to 720_1N) in the first network section 720_1, and that is combined, in the layer, with the signal to which the convolution processing is applied.
The branch section 710 inputs the device state information to one of the first layer 720_21 to the N-th layer 720_2N in the second network section 720_2. Within the layer to which the device state information is entered by the branch section 710, the device state information is combined with a signal to which the convolution processing is applied. It is more preferable that the device state information is input to a layer that is positioned closer to the branch section 710 among the layers (720_21 to 720_2N) in the second network section 720_2, and that is combined, in the layer, with the signal to which the convolution processing is applied.
As described above, because the training unit 161 is configured such that multiple sets of data (e.g., time series data set 1 and time series data set 2 in the above-described example) are generated by processing the time series data sets in accordance with each of the different criteria (e.g., first criterion and second criterion) and that each of the multiple sets of data is processed in a different network section, and because machine learning is performed on the above-described configuration, time series data sets at the unit of process 120 can be analyzed in a multifaceted manner. As a result, a model (inference unit 162) that realizes a high inference accuracy can be generated as compared to a case in which time series data sets are processed using a single network section.
The example of
(2) Details of Processing (2) Performed in the Branch Section
Next, another processing performed in the branch section 710 will be described in detail.
The branch section 710 inputs the device state information to one of the first layer 720_31 to the N-th layer 720_3N of the third network section 720_3. In the layer to which the device state information is entered by the branch section 710, the device state information is combined with a signal to which the convolution processing is applied. It is more preferable that the device state information is input to a layer that is positioned closer to the branch section 710 among the layers (720_31 to 720_3N) in the third network section 720_3, and that is combined, in the layer, with the signal to which the convolution processing is applied.
The branch section 710 inputs the device state information to one of the first layer 720_41 to the N-th layer 720_4N in the fourth network section 720_4. In the layer to which the device state information is entered by the branch section 710, the device state information is combined with a signal to which the convolution processing is applied. It is more preferable that the device state information is input to a layer that is positioned closer to the branch section 710 among the layers (720_41 to 720_4N) in the fourth network section 720_4, and that is combined, in the layer, with the signal to which the convolution processing is applied.
As described above, because the training unit 161 is configured to classify the time series data sets into multiple sets of data (e.g., time series data set 1 and time series data set 2 in the above-described example) in accordance with data type, and to process each of the multiple sets of data in a different network section, and because machine learning is performed on the above-described configuration, the unit of process 120 can be analyzed in a multifaceted manner. As a result, it is possible to generate a model (inference unit 162) that achieves a high inference accuracy, as compared to a case in which machine learning is performed by inputting time series data sets into a single network section.
In the example of
(3) Details of Processing (3) Performed in the Branch Section
Next, yet another processing performed in the branch section 710 will be described in detail. FIG. is a third diagram illustrating a specific example of the processing performed in the branch section 710. In the case of
The example of
Among these, the normalizing unit 1101 applies a first normalization process to the time series data sets that are input from the branch section 710, to generate the normalized time series data set 1 (first time series data set). The normalized time series data set 1 is combined with the device state information input by the branch section 710, and is input to the convolving unit 1102. The first normalization process and a process of combining the normalized time series data set 1 with the device state information, performed by the normalizing unit 1101, may be performed in another layer in the fifth network section 720_5 other than the first layer 720_51, but more preferably, may be performed in a layer that is positioned closer to the branch section 710 among the layers (720_51 to 720_5N) in the fifth network section 720_5.
In addition, the example of
Among these, the normalizing unit 1111 applies a second normalization process to the time series data sets that are input from the branch section 710, to generate the normalized time series data set 2 (second time series data set). The normalized time series data set 2 is combined with the device state information input by the branch section 710 and is input to the convolving unit 1112. The second normalization process and a process of combining the normalized time series data set 2 with the device state information, performed by the normalizing unit 1111, may be performed in another layer in the sixth network section 720_6 other than the first layer 720_61, but more preferably, may be performed in a layer that is positioned closer to the branch section 710 among the layers (720_61 to 720_6N) in the sixth network section 720_6.
As described above, because the training unit 161 is configured to process time series data sets using multiple network sections each including a normalizing unit that performs normalization using a different method from other normalizing units, and because machine learning is performed on the above-described configuration, the unit of process 120 can be analyzed in a multifaceted manner. As a result, a model (inference unit 162) that achieves a high inference accuracy can be generated, as compared to a case in which a single type of normalization is applied to the time series data sets using a single network section. Moreover, the model developed in the training unit 161 may be employed in the inference unit 162 to identify processes that will likely result in predicted conditions that may adversely affect a quality of a manufactured semiconductor component. By the predicting the condition with the trained model, the trained model may be used to control of semiconductor manufacturing equipment to trigger supervised or automated maintenance operations on a process chamber; adjustment of at least one of a RF power system (e.g., adjustment of RF power levels and/or RF waveform) for generating plasma or a gas input (or process gas composition) and/or gas exhaust operation, supervised or automated calibration operations (e.g., gas flow and/or RF waveforms for generating plasma, supervised or automated adjustment of gas flow levels, supervised or automated replacement of components such as electrostatic chuck, which may become wasted over time, and the like
(4) Details of Processing (4) Performed in the Branch Section
Next, still another processing performed in the branch section 710 will be described in detail.
The branch section 710 inputs the time series data set 2 (second time series data set) measured along with the processing of the wafer in the chamber B to the eighth network section 720_8, among the time series data sets measured by the time series data acquiring devices 140_1 to 140_n.
The branch section 710 inputs the device state information acquired when the wafer is processed in the chamber A to one of the first layer 720_71 to the N-th layer 720_7N in the seventh network section 720_7. In the layer to which the device state information is entered by the branch section 710, the device state information is combined with a signal to which the convolution processing is applied. It is more preferable that the device state information is input to a layer that is positioned closer to the branch section 710 among the layers (720_71 to 720_7N) in the seventh network section 720_7, and that is combined, in the layer, with the signal to which the convolution processing is applied.
The branch section 710 inputs the device state information acquired when the wafer is processed in the chamber B to one of the first layer 720_81 to the N-th layer 720_8N in the eighth network section 720_8. In the layer to which the device state information is entered by the branch section 710, the device state information is combined with a signal to which the convolution processing is applied. It is more preferable that the device state information is input to a layer that is positioned closer to the branch section 710 among the layers (720_81 to 720_8N) in the eighth network section 720_8, and that is combined, in the layer, with the signal to which the convolution processing is applied.
As described above, because the training unit 161 is configured to process different time series data sets, each being measured along with processing in a different chamber (first processing space and second processing space), by using respective network sections, because machine learning is performed on the above-described configuration, the unit of process 120 can be analyzed in a multifaceted manner. As a result, a model (inference unit 162) that achieves a high inference accuracy can be generated, as compared to a case in which each of the time series data sets is configured to be processed using a single network section.
Next, the functional configuration of the inference unit 162 will be described.
The branch section 1310 acquires the time series data sets newly measured by the time series data acquiring devices 140_1 to 140_N after the time series data sets, which were used by the training unit 161 for machine learning, were measured, and acquires the device state information. The branch section 1310 is also configured to cause the first to M-th network sections (1320_1 to 1320_M) to process the time series data sets and the device state information. Note that the device state information can be varied (i.e., the device state information is treated as a configurable parameter in the inference unit 162), and the branch section 1310 repeatedly inputs the same time series data sets to the first to M-th network sections (1320_1 to 1320_M) while changing a value of the device state information.
The first to M-th network sections (1320_1 to 1320_M) are implemented, by performing machine learning in the training unit 161 to optimize model parameters of each of the layers in the first to M-th network sections (720_1 to 720_M).
The concatenation section 1330 is implemented by the concatenation section 730 whose model parameters have been optimized by performing machine learning in the training unit 161. The concatenation section 1330 combines output data output from an N-th layer 1320_1N of the first network section 1320_1 to an N-th layer 1320_1N of the M-th network section 1320_M, to output a result of inference (quality indicator) for each value of the device state information.
The monitoring section 1340 acquires the quality indicators output from the concatenation section 1330 and the corresponding values of the device state information. The monitoring section 1340 generates a graph having the device state information as the horizontal axis and the quality indicator as the vertical axis, by plotting sets of the acquired quality indicators and the corresponding values of the device state information. The graph 1341 illustrated in
The predicting section 1350 specifies the value of the device state information (point 1351 in the example of
It should be noted that the predetermined threshold 1352 may be determined with respect to a quality indicator related to necessity of maintenance of the semiconductor manufacturing device. Alternatively, the predetermined threshold 1352 may be determined with respect to a quality indicator related to necessity of replacement of parts within the semiconductor manufacturing device.
As described above, the inference unit 162 is generated by machine learning being performed in the training unit 161, which analyzes the time series data sets with respect to the predetermined unit of process 120 in a multifaceted manner. Thus, the inference unit 162 can also be applied to different process recipes, different chambers, and different devices. Alternatively, the inference unit 162 can be applied to a chamber before maintenance and to the same chamber after its maintenance. That is, the inference unit 162 according to the present embodiment eliminates the need, for example, to maintain or retrain a model after maintenance of a chamber is performed, which is required in conventional systems.
Next, an overall flow of the predicting process performed by the predicting device 160 will be described.
In step S1401, the training unit 161 acquires time series data sets, device state information, and a quality indicator, as training data.
In step S1402, the training unit 161 performs machine learning by using the acquired training data. Of the acquired training data, the time series data sets and the device state information are used as input data, and the quality indicator is used as correct answer data.
In step S1403, the training unit 161 determines whether to continue the machine learning. If machine learning is continued by acquiring further training data (in a case of YES in step S1403), the process returns to step S1401. Meanwhile, if the machine learning is terminated (in a case of NO in step S1403), the process proceeds to step S1404.
In step S1404, the inference unit 162 generates the first to M-th network sections 1320_1 to 1320_M by reflecting model parameters optimized by the machine learning.
In step S1405, the inference unit 162 initialize the device state information. As the initial value of the device state information, for example, the inference unit 162 may acquire a value of the device state information that has been measured along with processing of a new wafer before processing.
In step S1406, the inference unit 162 infers the quality indicator, by inputting time series data sets measured along with the processing of a new wafer before processing and by inputting the value of the device state information.
In step S1407, the inference unit 162 determines whether or not the inferred quality indicator exceeds a predetermined threshold. If it is determined in step S1407 that the inferred quality indicator does not exceed the predetermined threshold (in the case of NO in step S1407), the process proceeds to step S1408.
In step S1408, the inference unit 162 increments the value of the device state information by a predetermined increment, and the process returns to step S1406. The inference unit 162 continues to increment the value of the device state information until it is determined that the inferred quality indicator exceeds the predetermined threshold.
Meanwhile, if it is determined in step S1407 that the inferred quality indicator exceeds the predetermined threshold (in the case of YES in step S1407), the process proceeds to step S1409.
In step S1409, the inference unit 162 specifies the value of the device state information when the inferred quality indicator exceeds the predetermined threshold. Based on the specified value of the device state information, the inference unit 162 predicts (i.e., estimates) and outputs replacement time of parts of the semiconductor manufacturing device or maintenance timing of the semiconductor manufacturing device.
As is apparent from the above description, the predicting device according to the first embodiment performs the following steps:
a) time series data sets and device state information measured along with processing of an object at a predetermined unit of process in the manufacturing process are acquired;
b) with respect to the acquired time series data sets, the following one of b-1), b-2), and b-3) is performed;
c) machine learning is performed with respect to the multiple network sections, such that a result of the combining of the output data output from each of the multiple network sections approaches the quality indicator obtained when processing the object at the predetermined unit of process in the manufacturing process;
d) while changing a value of the device state information, newly obtained time series data sets, which are measured by time series data acquiring devices along with processing of a new object, are processed by using the multiple network sections to which a result of the machine learning is applied, to infer the quality indicator for each value of the device state information, by outputting, for each of the values of the device state information, a result of combining output data output from each of the multiple network sections to which machine learning has been applied; and
e) whether the quality indicator, inferred for each of the values of the device state information, satisfies the predetermined conditions, is determined, and replacement time of parts of the semiconductor manufacturing device or maintenance timing of the semiconductor manufacturing device is predicted by using a value of the device state information when the quality indicator satisfies the predetermined conditions.
Thus, according to the first embodiment, it is possible to provide a predicting device that utilizes time series data sets measured along with processing of an object in a semiconductor manufacturing process and device state information acquired during the processing of the object.
In the predicting device 160 according to the first example embodiment, with respect to the configuration in which acquired time series data sets and device state information are processed using multiple network sections, four types of configurations are illustrated. The second embodiment further describes, among these four configurations, a configuration in which time series data sets and device state information are processed using multiple network sections each including a normalizing unit that performs normalization using a different method from other normalizing units. In the following description, a case in which
a time series data acquiring device is an optical emission spectrometer, and
time series data sets are optical emission spectroscopy data (hereinafter referred to as “OES data”), which are data sets including the number, corresponding to the number of types of wavelengths, of sets of time series data of emission intensity will be described.
Hereinafter, the second embodiment will be described focusing on the differences from the above-described first embodiment.
First, the overall configuration of a system including a device performing a semiconductor manufacturing process and a predicting device will be described, in which the time series data acquiring device in the system is an optical emission spectrometer.
In the system 1500 illustrated in
Next, the training data, which is read out from the training data storage unit 163 when the training unit 161 performs machine learning, will be described.
Next, a specific example of the OES data measured in the optical emission spectrometer 1501 will be described.
In
As illustrated in the graph 1710, the OES data measured in the optical emission spectrometer 1501 differs in length of time in each wafer to be processed.
In the example of
Meanwhile, the lateral size (width) of the OES data 1720 depends on the length of time measured by the optical emission spectrometer 1501. In the example of
Thus, the OES data 1720 can be said to be a set of time series data that groups together a predetermined number of wavelengths, where there is one-dimensional time series data of a predetermined length of time for each of the wavelengths.
When the OES data 1720 is input to the fifth network section 720_5 and the sixth network section 720_6, the branch section 710 resizes the data on a per minibatch basis, such that the data size is the same as that of the OES data of other wafer identification numbers.
Next, a specific example of processing performed by the normalizing units in the fifth network section 720_5 and the sixth network section 720_6, into each of which the OES data 1720 is input from the branch section 710, will be described.
As illustrated in
Thus, even though the same OES data 1720 is used, information that will be found out from the same OES data 1720 differs depending on what is used as a reference (i.e., depending on analysis methods). The predicting device 160 according to the second embodiment causes different network sections, each of which is configured to perform a different normalization, to process the same OES data 1720. Thus, by combining multiple normalization processes, it is possible to analyze the OES data 1720 in the unit of process 120 in a multifaceted manner. As a result, a model (inference unit 162) that realizes the high inference accuracy can be generated, as compared to a case in which a single type of normalization process is applied to the OES data 1720 using a single network section.
The above-described example describes a case in which normalization is performed using an average value of emission intensity and a standard deviation of emission intensity. However, a statistical value used for normalization is not limited thereto. For example, the maximum value and a standard deviation of emission intensity may be used for normalization, or other statistics may be used. In addition, the predicting device 160 may be configured such that a user can select types of a statistical value to be used for normalization.
Next, a specific example of the processing performed by the pooling units included in the final layer of the fifth network section 720_5 and in the final layer of the sixth network section 720_6 will be described.
Because data size differs between minibatches, the pooling units 1104 and 1114 included in the respective final layers of the fifth network section 720_5 and the sixth network section 720_6 perform pooling processes such that fixed-length data is output between minibatches (i.e., size of output data according to each minibatch becomes the same).
In
Feature data 2012_1 to 2012_m represent feature data generated based on the OES data belonging to the minibatch 2, and are input to the pooling unit 1104 of the N-th layer 720_5N of the fifth network section 720_5. Each of the feature data 2012_1 to 2012_m represents feature data corresponding to one channel.
Also, feature data 2031_1 to 2031_m and feature data 2032_1 to 2032_m are similar to the feature data 2011_1 to 2011_m or the feature data 2012_1 to 2012_m. However, each of the feature data 2031_1 to 2031_m and 2032_1 to 2032_m is feature data corresponding to Nλ channels.
Here, the pooling units 1104 and 1114 calculate an average value of feature values included in the input feature data on a per channel basis, to output the fixed-length output data. Thus, the data output from the pooling units 1104 and 1114 can have the same data size between minibatches.
Next, the functional configuration of the inference unit 162 will be described.
The branch section 1310 acquires OES data newly measured by the optical emission spectrometer 1501 after the OES data used by the training unit 161 for machine learning was measured, and acquires device state information. The branch section 1310 is also configured to cause both the fifth network section 1320_5 and the sixth network section 1320_6 to process the OES data and the device state information. The device state information can be varied, and the branch section 1310 repeatedly inputs the same time series data sets while changing a value of the device state information.
The fifth network section 1320_5 and the sixth network section 1320_6 are implemented, by performing machine learning in the training unit 161 to optimize model parameters of each of the layers in the fifth network section 720_5 and the sixth network section 720_6.
The concatenation section 1330 is implemented by the concatenation section 730 whose model parameters have been optimized by performing machine learning in the training unit 161. The concatenation section 1330 combines output data that is output from an N-th layer 1320_5N of the fifth network section 1320_5 and from an N-th layer 1320_6N of the sixth network section 1320_6, to output an inference result (quality indicator) for each value of the device state information.
As the monitoring section 1340 and the predicting section 1350 are the same as the monitoring section 1340 and the predicting section 1350 illustrated in
As described above, the inference unit 162 is generated by machine learning being performed in the training unit 161, which analyzes the OES data with respect to the predetermined unit of process 120 in a multifaceted manner. Thus, the inference unit 162 can also be applied to different process recipes, different chambers, and different devices. Alternatively, the inference unit 162 can be applied to a chamber before maintenance and to the same chamber after its maintenance. That is, the inference unit 162 according to the present embodiment eliminates the need, for example, to maintain or retrain a model after maintenance of the chamber is performed, which was required in conventional systems.
Next, an overall flow of the predicting process performed by the predicting device 160 will be described.
In step S2201, the training unit 161 acquires OES data, device state information, and a quality indicator, as training data.
In step S2202, the training unit 161 performs machine learning by using the acquired training data. Specifically, the OES data and the device state information in the acquired training data are used as input data, and the quality indicator in the acquired training data is used as correct answer data.
In step S2203, the inference unit 162 infers the quality indicator, by inputting OES data sets measured along with processing of a new wafer before processing, and by inputting the value of the device state information.
As is apparent from the above description, the predicting device according to the second embodiment performs the following steps:
acquiring, at a predetermined unit of process in a manufacturing process, OES data measured by an optical emission spectrometer along with processing of an object and device state information acquired during the processing of the object;
inputting the acquired OES data and device state information to two network sections each of which performs normalization using a different method from each other;
combining output data output from each of the two network sections;
performing machine learning with respect to the two network sections such that a result of the combining of the output data output from each of the two network sections approaches a quality indicator obtained during the processing of the object at the predetermined unit of process in the manufacturing process;
while changing a value of the device state information, processing OES data measured along with processing of a new object by the optical emission spectrometer, by using the two network sections to which machine learning has been applied;
inferring the quality indicator for each value of the device state information, by outputting a result of combining output data output from each of the two network sections to which machine learning has been applied;
determining whether the quality indicator, inferred for each of the values of the device state information, satisfies the predetermined conditions; and
predicting (estimating) replacement time of parts of the semiconductor manufacturing device or maintenance timing of the semiconductor manufacturing device, by using a value of the device state information when the quality indicator satisfies the predetermined conditions.
Thus, according to the second embodiment, it is possible to provide a predicting device that utilizes OES data, which is time series data sets measured along with processing of an object in a semiconductor manufacturing process, and the device state information acquired during the processing of the object.
In the second embodiment, as an example of a time series data acquiring device, an optical emission spectrometer is described. However, types of the time series data acquiring device applicable to the first embodiment are not limited to the optical emission spectrometer.
For example, examples of the time series data acquiring device described in the first embodiment may include a process data acquiring device that acquires various process data, such as temperature data, pressure data, or gas flow rate data, as one-dimensional time series data. Alternatively, the time series data acquiring device described in the first embodiment may include a radio-frequency (RF) power supply device for plasma configured to acquire various RF data, such as voltage data of the RF power supply, as one-dimensional time series data.
The above-described first and second embodiments are described such that a machine learning algorithm for each of the network sections in the training unit 161 is configured based on a convolutional neural network. However, the machine learning algorithm for each of the network sections in the training unit 161 is not limited to the convolutional neural network, and may be based on other machine learning algorithms.
The first and second embodiments described above have been described such that the predicting device 160 functions as the training unit 161 and the inference unit 162. However, an apparatus serving as the training unit 161 needs not be integrated with an apparatus serving as the inference unit 162, and an apparatus serving as the training unit 161 and an apparatus serving as the inference unit 162 may be separate apparatuses. That is, the predicting device 160 may function as the training unit 161 not including the inference unit 162, or the predicting device 160 may function as the inference unit 162 not including the training unit 161.
The above-described functions of the predicting device 160, such as functions of the training unit 161 and the inference unit 162, may be implemented in a controller of the semiconductor manufacturing device 200, and the controller (inference unit 162) of the semiconductor manufacturing device 200 may predict replacement time of each part in the semiconductor manufacturing device 200. Based on the predicted replacement time, the controller (inference unit 162) of the semiconductor manufacturing device 200 may display a warning message on a display device of the controller, or may operate the semiconductor manufacturing device 200. For example, if the current time reaches the predicted replacement time of a part of the semiconductor manufacturing device 200, the controller (inference unit 162) may stop operations of the semiconductor manufacturing device in order to replace the part.
It should be noted that the present invention is not limited to the above-described configurations, such as configurations described in the embodiments described above, or configurations combined with other elements. Configurations may be changed to an extent not departing from the spirit of the invention, and can be appropriately determined in accordance with their application forms.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2019-217440 | Nov 2019 | JP | national |
