Patent Application
20230004803
The present invention generally relates to material science, particularly to reconfigurable memtransistors for continuous learning in spiking neural networks, fabricating methods, and applications of the same.
The background description provided herein is to present the context of the invention generally. The subject matter discussed in the background of the invention section should not be assumed to be prior art merely due to its mention in the background of the invention section. Similarly, a problem mentioned in the background of the invention section or associated with the subject matter of the background of the invention section should not be assumed to have been previously recognized in the prior art. The subject matter in the background of the invention section merely represents different approaches, which in and of themselves may also be inventions. Work of the presently named inventors, to the extent it is described in the background of the invention section, as well as aspects of the description that may not otherwise qualify as prior art at the time of filing, are neither expressly nor impliedly admitted as prior art against the invention.
Exponential improvement in solid-state digital electronics over the past several decades has led to an array of modern ubiquitous technologies such as the Internet of Things, edge computing, artificial intelligence (AI), and machine learning (ML) that are impacting nearly all aspects of society. Recent progress in AI/ML has been primarily driven by software improvements exemplified by the DeepMind AlphaGo program that defeated a world champion in the game of Go. However, running AI/ML algorithms on conventional von Neumann hardware platforms results in substantial energy consumption, which is orders of magnitude higher than that of the human brain. AI/ML hardware accelerators based on neuromorphic architectures are being actively pursued using memristors, phase change memory, and synaptic transistors to improve energy efficiency. These devices imitate specific biological responses, such as synaptic plasticity, where the conductance state is modified by a temporal relation between pre-synaptic and post-synaptic neuron spikes. However, synaptic plasticity in biology is more complex than current neuromorphic demonstrations and involves more than two neurons to regulate the strength of synaptic connections. Therefore, to better mimic complex biological synapses, three-terminal neuromorphic devices have emerged to improve energy efficiency, linearity, and reconfigurability.
In parallel with advances in neuromorphic device concepts, two-dimensional (2D) materials have attracted significant attention as a platform for next-generation electronics. The atomic-level thickness of 2D materials imparts weak screening that allows strong electrostatic tunability and reconfigurability of device responses. For example, monolayer polycrystalline MoS2 memtransistors have achieved gate-tunable memristive switching. The memtransistor is a promising building block for next-generation bio-realistic neuromorphic systems by co-locating memory and transistor functionality. Dual-gated MoS2 memtransistors also minimize crosstalk and sneak currents in scalable crossbar architectures, thus simplifying integration challenges that have hindered memristive architectures based on bulk materials. Despite the unique attributes of memtransistors, their implementation in neuromorphic architectures has been limited to standard artificial neural networks, suggesting that their full potential for AI/ML has not yet been realized.
Therefore, a heretofore unaddressed need exists in the art to address the aforementioned deficiencies and inadequacies.
In one aspect, this invention relates to a memtransistor. The memtransistor comprises a polycrystalline monolayer film of an atomically thin material, wherein the polycrystalline monolayer film is grown directly on a first substrate and transferred onto a second substrate; and a gate electrode defined on the second substrate; and source and drain electrodes spatially-apart formed on the polycrystalline monolayer film to define a channel region in the polycrystalline monolayer film therebetween, wherein the gate electrode is capacitively coupled with the channel region.
In one embodiment, the atomically thin material comprises two-dimensional (2D) semiconductor material.
In one embodiment, the 2D semiconductor material comprises MoS2, MoSe2, WS2, WSe2, InSe, GaTe, black phosphorus (BP), or related two-dimensional materials.
In one embodiment, the polycrystalline monolayer film of MoS2 has well-defined grain boundaries, sub-stoichiometric S:Mo ratio, and predominantly monolayer coverage.
In one embodiment, the first substrate is formed of sapphire, quartz, graphene, or hexagonal boron nitride.
In one embodiment, the second substrate is an SiO2/Si substrate, or an substrate of a high-k dielectric layer including Al2O3 or HfO2.
In one embodiment, the SiO2/Si substrate comprises a silicon substrate with a silicon dioxide overlayer.
In one embodiment, the gate, source, and drain electrodes comprise a same conductive material or different conductive materials.
In one embodiment, the polycrystalline monolayer film of MoS2 has well-defined grain boundaries, sub-stoichiometric S:Mo ratio, and predominantly monolayer coverage.
In one embodiment, the memtransistor is reconfigurable with gate tunability that enables continuous learning that allows selective forgetting of inessential tasks, thus freeing up neural resources to learn new tasks.
In one embodiment, by growing the polycrystalline monolayer film grown directly on the sapphire, quartz, graphene, or hexagonal boron nitride substrate, lattice defects in the polycrystalline monolayer film are reduced and the crystallographic registry is improved, thereby enabling accentuation of the vertical field effect from the gate compared to drain bias induced resistive switching, and heightening reconfigurability of a synaptic learning behavior from long-term potentiation (LTP) to long-term depression (LTD).
In one embodiment, the LTP and the LTD are controlled by the gate bias polarity and not the drain pulse polarity, which parallels biological systems' synaptic weight update and neuroplasticity.
In one embodiment, by mimicking the biological systems, LTP/LTD tuning is achieved by biasing the gate without changing the polarity of drain pulses.
In one embodiment, additional learning behaviors are achieved by varying the temporal evolution of gate bias pulses.
In one embodiment, the gate pulses are used to modulate potentiation and depression, resulting in diverse learning curves and simplified spike-timing-dependent plasticity that facilitate unsupervised learning in a simulated spiking neural network (SNN).
In one embodiment, a library of learning curves obtained from temporal evolution of the pulsing amplitude is used to perform unsupervised image recognition in the SNN with functions of continuous learning.
In one embodiment, the unsupervised learning in the SNN is performed using an experimental memtransistor learning behavior modeled in a simplified spike-timing-dependent plasticity (STDP) scheme.
In another aspect, the invention relates to a circuit comprising one or more memtransistors as disclosed above.
In yet another aspect, the invention relates to an electronic device comprising one or more memtransistors as disclosed above.
In a further aspect, the invention relates a system for continuous learning in a spiking neural network, comprising one or more synaptic units, wherein each synaptic unit comprises one or more memtransistors as disclosed above.
In one embodiment, each synaptic unit has learning and/or unlearning behaviors, with the gate-tunable characteristics of the memtransistors.
In one embodiment, switching LTP-LTD learning behavior is achieved by only reversing the polarity of the gate pulses, while further adjustments in the gate amplitude produced diverse learning curves and thus learning behaviors.
In one aspect, the invention relates a method for fabricating a memtransistor, comprising growing a polycrystalline monolayer film of an atomically thin material on a first substrate; transferring the polycrystalline monolayer film to a second substrate; and forming a gate electrode on the second substrate and source and drain electrodes on the grown polycrystalline monolayer film, wherein the source and drain electrodes define a channel region in the polycrystalline monolayer film therebetween, and wherein the gate electrode is capacitively coupled with the channel region.
In one embodiment, the first substrate is formed of sapphire, quartz, graphene, or hexagonal boron nitride.
In one embodiment, the second substrate is an SiO2/Si substrate, or an substrate of a high-k dielectric layer including Al2O3 or HfO2.
In one embodiment, the polycrystalline monolayer film is grown by chemical vapor deposition (CVD) on the substrates.
In one embodiment, said transferring comprises coating a polymer film on the polycrystalline monolayer film grown on the first substrate; separating the polymer film with the polycrystalline monolayer film from the first substrate; adhering the separated polymer film with the polycrystalline monolayer film to the second substrate; and removing the polymer film.
In one embodiment, the polymer film is formed of polycarbonate (PC).
In one embodiment, said forming is performed by photolithography.
In one embodiment, the atomically thin material comprises 2D semiconductor material.
In one embodiment, the 2D semiconductor material comprises MoS2, MoSe2, WS2, WSe2, InSe, GaTe, BP, or related two-dimensional materials.
These and other aspects of the present invention will become apparent from the following description of the preferred embodiment taken in conjunction with the following drawings, although variations and modifications therein may be affected without departing from the spirit and scope of the novel concepts of the invention.
The accompanying drawings illustrate one or more embodiments of the invention and together with the written description, serve to explain the principles of the invention. Wherever possible, the same reference numbers are used throughout the drawings to refer to the same or like elements of an embodiment.
The invention will now be described more fully hereinafter with reference to the accompanying drawings, in which exemplary embodiments of the invention are shown. However, this invention may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this specification will be thorough and complete and fully convey the invention's scope to those skilled in the art. Like reference numerals refer to like elements throughout.
The terms used in this specification generally have their ordinary meanings in the art, within the context of the invention, and in the specific context where each term is used. Certain terms used to describe the invention are discussed below, or elsewhere in the specification, to provide additional guidance to the practitioner regarding the description. For convenience, certain terms may be highlighted, for example using italics and/or quotation marks. The use of highlighting has no influence on the scope and meaning of a term; the scope and meaning of a term are the same, in the same context, whether or not it is highlighted. It will be appreciated that same thing can be said in more than one way. Consequently, alternative language and synonyms may be used for any one or more of the terms discussed herein, nor is any special significance to be placed upon whether or not a term is elaborated or discussed herein. Synonyms for certain terms are provided. A recital of one or more synonyms does not exclude the use of other synonyms. The use of examples anywhere in this specification including examples of any terms discussed herein is illustrative only, and in no way limits the scope and meaning of the invention or of any exemplified term. Likewise, the invention is not limited to various embodiments given in this specification.
It will be understood that, as used in the description herein and throughout the claims that follow, the meaning of “a”, “an”, and “the” includes plural reference unless the context clearly dictates otherwise. Also, it will be understood that when an element is referred to as being “on” another element, it can be directly on the other element or intervening elements may be present therebetween. In contrast, when an element is referred to as being “directly on” another element, there are no intervening elements present. As used herein, the term “and/or” includes any and all combinations of one or more of the associated listed items.
It will be understood that, although the terms first, second, third, etc. may be used herein to describe various elements, components, regions, layers and/or sections, these elements, components, regions, layers and/or sections should not be limited by these terms. These terms are only used to distinguish one element, component, region, layer or section from another element, component, region, layer or section. Thus, a first element, component, region, layer or section discussed below could be termed a second element, component, region, or section without departing from the invention's teachings.
Furthermore, relative terms, such as “lower” or “bottom” and “upper” or “top,” may be used herein to describe one element's relationship to another element as illustrated in the figures. It will be understood that relative terms are intended to encompass different orientations of the device in addition to the orientation depicted in the figures. For example, if the device in one of the figures. is turned over, elements described as being on the “lower” side of other elements would then be oriented on “upper” sides of the other elements. The exemplary term “lower”, can, therefore, encompasses both an orientation of “lower” and “upper,” depending on the particular orientation of the figure. Similarly, if the device in one of the figures is turned over, elements described as “below” or “beneath” other elements would then be oriented “above” the other elements. Therefore, the exemplary terms “below” or “beneath” can encompass both an orientation of above and below.
It will be further understood that the terms “comprises” and/or “comprising,” or “includes” and/or “including” or “has” and/or “having”, or “carry” and/or “carrying,” or “contain” and/or “containing,” or “involve” and/or “involving, and the like are to be open-ended, i.e., to mean including but not limited to. When used in this specification, they specify the presence of stated features, regions, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, regions, integers, steps, operations, elements, components, and/or groups thereof.
Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and this specification, and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.
As used in this specification, “around”, “about”, “approximately” or “substantially” shall generally mean within 20 percent, preferably within 10 percent, and more preferably within 5 percent of a given value or range. Numerical quantities given herein are approximate, meaning that the term “around”, “about”, “approximately” or “substantially” can be inferred if not expressly stated.
As used in this specification, the phrase “at least one of A, B, and C” should be construed to mean a logical (A or B or C), using a non-exclusive logical OR. As used herein, the term “and/or” includes any and all combinations of one or more of the associated listed items.
The description below is merely illustrative in nature and is in no way intended to limit the invention, its application, or uses. The broad teachings of the invention can be implemented in a variety of forms. Therefore, while this invention includes particular examples, the true scope of the invention should not be so limited since other modifications will become apparent upon a study of the drawings, the specification, and the following claims. For purposes of clarity, the same reference numbers will be used in the drawings to identify similar elements. It should be understood that one or more steps within a method may be executed in a different order (or concurrently) without altering the principles of the invention.
Artificial intelligence and machine learning are growing computing paradigms, but current algorithms incur undesirable energy costs on conventional silicon-based hardware, motivating the exploration of efficient neuromorphic architectures.
One of the objectives of this invention is to provide a novel device concept in the class of three-terminal memtransistors with gate-tunable dynamic learning behavior. Unprecedented synaptic behavior is achieved by fabricating memtransistors from monolayer MoS2 grown on sapphire by chemical vapor deposition (CVD). Due to reduced lattice defects in CVD MoS2 grown on sapphire, the vertical field effect from the gate is enhanced compared to drain bias induced resistive switching, heightening the reconfigurability of the synaptic learning behavior from long-term potentiation (LTP) to long-term depression (LTD). Mimicking biological systems, LTP/LTD tuning is achieved by biasing the gate terminal without changing the polarity of drain terminal pulses. Furthermore, additional learning behaviors emerge by varying the temporal evolution of gate bias pulses. The resulting spike-timing-dependent plasticity facilitates unsupervised learning in simulated spiking neural networks (SNN). The gate tunability of these reconfigurable MoS2 memtransistors uniquely enables continuous learning, which is an underexplored cognitive concept that allows selective forgetting of inessential tasks, thus freeing up neural resources to learn new tasks. Overall, this invention demonstrates that the reconfigurability of memtransistors provides unique opportunities for energy-efficient artificial intelligence and machine learning.
The previous reports that fabricate memristors, memtransistors, or similar resistive switching devices use polycrystalline monolayer MoS2 film grown directly on SiO2/Si for ease of fabrication, and do not use transferred MoS2 initially grown on sapphire wafers. Growing MoS2 on sapphire imparts crystalline registry to the film and reduced density of lattice defects that are responsible for resistive switching. Thus, in memtransistors fabricated from MoS2 grown on sapphire, the electric field from the gate has a disproportionally large effect compared to the lateral source-drain field, which incurs qualitative changes in the learning behavior from potentiation to depression. The library of learning curves obtained from temporal evolution of the pulsing amplitude can then be used to perform unsupervised image recognition in a simulated spiking neural network where the concept of continuous learning is demonstrated. An electronic device with potential applications in continuously evolving neural networks has not been demonstrated previously and thus the present reconfigurable MoS2 memtransistor solves this problem uniquely.
In addition, current commercial solutions for brain-inspired neuromorphic hardware cannot adapt to dynamically varying application needs. For example, a neuromorphic chip intended for automated digit recognition cannot reconfigure itself on-demand to perform both digit recognition and character recognition, which is in stark contrast to real biological systems. Bridging this critical gap between artificial and natural intelligence, we demonstrate synaptic units that can learn and forget by the first demonstration of continuous learning by a solid-state electronic device, namely reconfigurable memtransistor devices using MoS2 grown on sapphire. Therefore, systems comprised of such synaptic units can assimilate new functionalities by replacing older (unused) functions. Our reconfigurable memtransistors can also dynamically reconfigure themselves to a diverse range of tasks over the device lifetime. This dynamic learning greatly enhances commercial opportunities for artificial intelligence hardware accelerators.
More specifically, the invention relates to memtransistors, fabricating methods, and applications of the same.
In one aspect of the invention, the memtransistor comprises a polycrystalline monolayer film of an atomically thin material, wherein the polycrystalline monolayer film is grown directly on a first substrate and transferred onto a second substrate; and a gate electrode defined on the second substrate; and source and drain electrodes spatially-apart formed on the polycrystalline monolayer film to define a channel region in the polycrystalline monolayer film therebetween, wherein the gate electrode is capacitively coupled with the channel region.
In one embodiment, the atomically thin material comprises two-dimensional (2D) semiconductor material.
In one embodiment, the 2D semiconductor material comprises MoS2, MoSe2, WS2, WSe2, InSe, GaTe, black phosphorus (BP), or related two-dimensional materials.
In one embodiment, the polycrystalline monolayer film of MoS2 has well-defined grain boundaries, sub-stoichiometric S:Mo ratio, and predominantly monolayer coverage.
In one embodiment, the first substrate is formed of sapphire, quartz, graphene, or hexagonal boron nitride.
In one embodiment, the second substrate is an SiO2/Si substrate, or an substrate of a high-k dielectric layer including Al2O3 or HfO2.
In one embodiment, the SiO2/Si substrate comprises a silicon substrate with a silicon dioxide overlayer.
In one embodiment, the gate, source and drain electrodes comprise a same conductive material or different conductive materials.
In one embodiment, the memtransistor is reconfigurable with gate tunability that enables continuous learning that allows selective forgetting of inessential tasks, thus freeing up neural resources to learn new tasks.
In one embodiment, by growing the polycrystalline monolayer film grown directly on the sapphire, quartz, graphene, or hexagonal boron nitride substrate, lattice defects in the polycrystalline monolayer film are reduced and the crystallographic registry is improved, thereby enabling accentuation of the vertical field effect from the gate compared to drain bias induced resistive switching, and heightening reconfigurability of a synaptic learning behavior from long-term potentiation (LTP) to long-term depression (LTD). It should be noted that other substrates can also be utilized to practice the invention if a similar reduction in defect density can be achieved.
In one embodiment, the LTP and the LTD are controlled by the gate bias polarity and not the drain pulse polarity, which parallels the synaptic weight update and neuroplasticity in biological systems.
In one embodiment, by mimicking the biological systems, LTP/LTD tuning is achieved by biasing the gate without changing the polarity of drain pulses.
In one embodiment, additional learning behaviors are achieved by varying the temporal evolution of gate bias pulses.
In one embodiment, the gate pulses are used to modulate potentiation and depression, resulting in diverse learning curves and simplified spike-timing-dependent plasticity that facilitate unsupervised learning in a simulated spiking neural network (SNN).
In one embodiment, a library of learning curves obtained from temporal evolution of the pulsing amplitude is used to perform unsupervised image recognition in the SNN with functions of continuous learning.
In one embodiment, the unsupervised learning in the SNN is performed using an experimental memtransistor learning behavior modeled in a simplified spike-timing-dependent plasticity (STDP) scheme.
In another aspect, the invention relates a circuitry, comprising one or more memtransistors as disclosed above.
In yet another aspect, the invention relates an electronic device, comprising one or more memtransistors as disclosed above.
In a further aspect, the invention relates a system for continuous learning in a spiking neural network, comprising one or more synaptic units, wherein each synaptic unit comprises one or more memtransistors as disclosed above.
In one embodiment, each synaptic unit has learning and/or unlearning behaviors, with the gate-tunable characteristics of the memtransistors.
In one embodiment, switching LTP-LTD learning behavior is achieved by only reversing the polarity of the gate pulses, while further adjustments in the gate amplitude produced diverse learning curves and thus learning behaviors.
In one aspect, the invention relates a method for fabricating a memtransistor, comprising growing a polycrystalline monolayer film of an atomically thin material on a first substrate; transferring the polycrystalline monolayer film to a second substrate; and forming a gate electrode on the second substrate and source and drain electrodes on the grown polycrystalline monolayer film, wherein the source and drain electrodes define a channel region in the polycrystalline monolayer film therebetween, and wherein the gate electrode is capacitively coupled with the channel region.
In one embodiment, the first substrate is formed of sapphire, quartz, graphene, or hexagonal boron nitride.
In one embodiment, the second substrate is an SiO2/Si substrate, or an substrate of a high-k dielectric layer including Al2O3 or HfO2.
In one embodiment, the polycrystalline monolayer film is grown by chemical vapor deposition (CVD) on the substrates.
In one embodiment, said transferring comprises coating a polymer film on the polycrystalline monolayer film grown on the first substrate; separating the polymer film with the polycrystalline monolayer film from the first substrate; adhering the separated polymer film with the polycrystalline monolayer film to the second substrate; and removing the polymer film.
In one embodiment, the polymer film is formed of polycarbonate (PC).
In one embodiment, said forming is performed by photolithography.
In one embodiment, the atomically thin material comprises two-dimensional (2D) semiconductor material.
In one embodiment, the 2D semiconductor material comprises MoS2, MoSe2, WS2, WSe2, InSe, GaTe, black phosphorus (BP), or related two-dimensional materials.
Among other things, the invention has at least the following advantages.
The invention may have widespread applications in neuromorphic computing, edge computing, artificial intelligence, machine learning, artificial neural networks, non-volatile memory, sensors, hardware accelerators, and the likes.
These and other aspects of the invention are further described below. Without intent to limit the scope of the invention, exemplary instruments, apparatus, methods, and their related results according to the embodiments of the invention are given below. Note that titles or subtitles may be used in the examples for convenience of a reader, which in no way should limit the scope of the invention. Moreover, certain theories are proposed and disclosed herein; however, in no way they, whether they are right or wrong, should limit the scope of the invention so long as the invention is practiced according to the invention without regard for any particular theory or scheme of action.
Artificial intelligence (AI) and machine learning (ML) are growing computing paradigms, but current algorithms incur undesirable energy costs on conventional hardware platforms, thus motivating the exploration of more efficient neuromorphic architectures.
This example discloses a memtransistor with gate-tunable dynamic learning behavior. By fabricating memtransistors from monolayer MoS2 grown on sapphire, the relative importance of the vertical field effect from the gate is enhanced, thereby heightening reconfigurability of the device response. Inspired by biological systems, gate pulses are used to modulate potentiation and depression, resulting in diverse learning curves and simplified spike-timing-dependent plasticity that facilitate unsupervised learning in simulated spiking neural networks. This capability also enables continuous learning, which is a previously underexplored cognitive concept in neuromorphic computing. Overall, this work demonstrates that the reconfigurability of memtransistors provides unique hardware accelerator opportunities for energy efficient artificial intelligence and machine learning.
Specifically, the range of learning behaviors of memtransistors is expanded through a combination of enhanced electrostatic control and tailored gate bias pulsing profiles. Utilizing monolayer MoS2 grown on sapphire, memtransistors are efficiently modulated by the gate electrode. Long-term potentiation (LTP) and long-term depression (LTD) are controlled by the gate bias polarity and not the drain pulse polarity, which parallels the synaptic weight update and neuroplasticity in biological systems. This unique capability imparted by 2D materials is leveraged to perform unsupervised learning in a simulated spiking neural network (SNN) using the experimental memtransistor learning behavior modelled in a simplified spike-timing-dependent plasticity (STDP) scheme. The experimental learning curves further enable undemonstrated unsupervised continuous learning in simulated SNNs, which circumvents traditional tradeoffs between image recognition accuracy and resource allocation. This proof-of-concept demonstration is crucial to developing lifelong learning capabilities in artificial intelligence (AI) and machine learning (ML) algorithms in addition to addressing catastrophic forgetting, which is a persistent challenge in neuromorphic computing that requires continuous, energy intensive task updates.
Material growth: Continuous MoS2 films were synthesized by chemical vapor deposition (CVD) using molybdenum trioxide (Millipore-Sigma, 99.97% trace metals basis) and sulfur powder (Millipore-Sigma, 99.98% trace metals basis). Sapphire (<0001>, MTI Corporation) was used as the substrate for CVD growth. Prior to growth, the substrates were bath-sonicated for 10 min in acetone and 10 min in isopropyl alcohol, followed by deionized water rinsing and nitrogen drying. An oxygen plasma step was applied for 3 min at about 200 mTorr to further clean the substrates. Substrates were then placed in the middle of a 1-inch tube furnace (Lindberg/Blue), with 12 mg molybdenum trioxide and 150 mg sulfur positioned approximately 2 cm and 32 cm away from the substrates upstream. The tube furnace was purged with ultra-high purity Ar (99.99%) at 200 sccm for 10 min and flushed twice (increased pressure to 400 Torr and then pumped down to about 78 mTorr) to create an inert environment. The pressure was kept at 150 Torr at 25 sccm Ar for the remainder of the procedure. To begin the growth, the furnace was first heated to 150° C. in 5 min and held at this temperature for 20 min, then ramped to 800° C. at a rate of 12° C./min and maintained for 20 min, followed by natural cooling. Meanwhile, sulfur power was heated up to 50° C. for 5 min by a heating tape wrapped around the quartz tube and maintained at that temperature for 49 min, then further heating to 150° C. at a rate of 4.5° C./min and held for 23 min, and finally natural cooling to room temperature.
Device fabrication: Synthesized MoS2 films on sapphire were transferred to silicon substrates with a 285 nm thick silicon dioxide overlayer. Polycarbonate (PC) solution was spin-coated on MoS2 on sapphire at 1600 rpm for 1 min. After heating the film at 100° C. for 1 min, the sapphire substrate was submerged into a water bath. The PC film with MoS2 was then separated from sapphire due to the hydrophobic nature of PC film and floated on top of the water bath. A clean oxidized silicon substrate was used to scoop up the floated film. Heating the silicon substrate gradually from 80° C. to 180° C. in 30 minutes allowed the MoS2 film to adhere fully to the silicon substrate. Finally, the PC film was removed by a chloroform bath for 4 h, followed by isopropyl alcohol rinsing and nitrogen drying preceding device fabrication.
MoS2 memtransistor devices were fabricated by standard photolithography. In the first step, to pattern metal electrodes, negative photoresist Futurrex NR-9 1000PY was spin-coated at 3,000 rpm for 40 s. The photoresist was baked at 150° C. for 60 s and then exposed for 30 s under a 365 nm wavelength UV light for a dose of 390 mJ/cm2. The post-exposure bake was performed at 100° C. for 60 s. Then, the patterns were developed by immersing the substrate in Futurrex resist developer RD 6 for 10 s. Subsequently, Ti/Au (5 nm/50 nm) was evaporated by thermal evaporation and the photoresist was lifted off by MicroChem resist Remover PG. Next, another step of photolithography was performed to define the channel region. Positive resist Microposit S1813 was spin-coated on the substrate at 4,000 rpm for 60 s and then baked at 100° C. for 60 s. After exposing for 15 sat a dose of 35 mJ/cm2, the resist was developed in MF-319 for 60 s. Reactive ion etching using an Ar plasma was used to remove the exposed MoS2 film outside the channel region. Finally, the photoresist was lifted off by Remover PG.
Spiking neuron network simulation: A simulated spiking neural network (SNN) was developed with Python 3.7 and the BRIAN 2.2 simulator package. A simplified spike-timing-dependent plasticity (STDP) learning model was used following previous reports. The MNIST (Modified National Institute of Standards and Technology) dataset of handwritten digits was used to train and test the two-layer neural network, which included one input layer of 784 input neurons and one output layer of M output neurons (M=10, 20, 50, 100, 200). The output neurons were modeled after the leaky-integrate-and-fire model in Equation (1), where t is time, τ=1 (leak time constant), g=γ=1 (multiplicative factor for leaky integrate and fire output neurons), and Iinput is the current resulting from resistance modulation of memtransistors connected to the output neurons (summation of the product of weights and internal state variable X for input neurons). The weights of the 784×M synaptic connections were randomly initialized from a Gaussian distribution between 0 and 1, which represented the minimum and maximum normalized conductance.
Linearity and symmetry of the device long-term potentiation and depression curves influenced the SNN weight update rule. This was achieved by fitting the learning curves against the STDP exponential models, as illustrated in Equations (2)-(3) where α and β designate learning rate and linearity fitting parameters, δG normalized conductance change, and subscripts p and m potentiation and depression. Notably, the equations do not specify an explicit time dependence, which may seem contrary to the fundamental basis of STDP. However, the simplified framework adopted here implies the time dependence through the exponential function, since repeated voltage pulses (i.e., large time window between input and output neuron spike events) are likely to evoke smaller changes in conductance than singular voltage pulses (i.e., smaller time window between input and output neuron spike events). The weight update scheme in the simulation also assumed that 784×M synaptic weights were stored on the fabricated devices. Likewise, output neuron dynamics aided in network stability. Lateral inhibition resets the internal state variable (for each output neuron) to zero after an output neuron spiking event to prevent simultaneous spiking among neighboring output neurons. Furthermore, homeostasis corrected each input neuron firing threshold every 200 training digits to aid in network stability. In the corresponding Equation (4), Xth is the threshold X internal state variable, γ=5 (multiplicative factor to increase threshold modifications for homeostasis), A is the average spiking activity, and T=1/M (target activity to achieve equal firing rates in homeostasis). The simulation used 60,000 training digits and 10,000 testing MNIST digits to measure the recognition rate of the network and was run ten times for reproducibility.
Alternate architectures like multilayer perceptron (MLP) networks were previously implemented to demonstrate the efficacy of memristive hardware and have reported recognition accuracies greater than 90%. While the recognition rates reported by the SNN were lower than those previous MLP reports, we emphasize that recognition rate is not the only figure of merit for neural networks. In this work, we highlight the unique capability of SNNs to conduct unsupervised continuous learning, a significant step in developing lifelong learning capabilities without forgoing energy considerations for neuromorphic computing. Here, key compositional differences between the two networks was briefly distinguished to clarify why SNN was used in this context. MLPs usually have fewer output neurons and layers than SNNs and use backpropagation and continuous activation functions (relying on multiply-and-accumulate summation functions) to propagate strictly spatial information. These methods are relatively straightforward to implement, but are also sensitive to noise and typically more power intensive. Due to these factors, MLPs are more suitable for supervised learning. SNNs in contrast are more structurally complex (with additional layers and input neurons, and thus synaptic connections), and communicate spatiotemporal information via spiking patterns of the input and output neurons, where past spiking history and timing affect such patterns. The bio-realistic STDP algorithm, which informs the frequency and timing of the spiking trains, consequently helps promote Hebbian learning in the connecting synapses. In the SNN, a homeostasis mechanism was also integrated into the structure where the threshold values of the internal states of the output neurons dynamically adjusted based on the activity of the neuron (i.e., the greater post-synaptic neuronal activity would lead to a gradual increase in the threshold X value). This added feature makes the network extremely robust against device-to-device variation and other non-idealities and noise, and thus more desirable for unsupervised learning.
Furthermore, the concept presented in this work of the learning-unlearning capability in memtransistor devices for application to continuous learning can be generalized to other datasets, such as the recently devised dynamic analog to MNIST known as Neuromorphic-MNIST (NMNIST). The use of other datasets can often achieve improved recognition rates. However, we emphasize that recognition rate is not the chief figure of merit that distinguishes STDP-SNN from other neural networks, but rather the unique capability of unsupervised continuous learning that emerges from the integration of 2D memtransistors with SNNs. Additionally, more recent datasets (e.g., NMNIST) have not yet reached a standardized protocol in the literature, which complicates performance benchmarking. Consequently, we limited ourselves to the most established MNIST database in this work.
Continuous learning simulation: Similar to the setup above, the simulation framework for SNN-based continuous learning was developed using Python, BRIAN 2.0, and PyTorch packages. A three-layer network including an input-layer with N=784 neurons, a hidden-layer with H=200 neurons, and an output layer with M=10 neurons was used to classify MNIST handwritten digits in a continuous learning setup. While training the network, synaptic weights between the input neurons and hidden neurons were updated using unsupervised STDP learning following Equations (2)-(3). In contrast, the weights between hidden and output neurons were updated using supervised learning to automate the learning-digit association process. Meanwhile, the hidden layer was divided into group A and group B with H/2 neurons each. Such segregation of neurons in groups allowed dynamic programming of neurons in each group for either learning or unlearning. Selective learning/unlearning for neurons in group A and group B was achieved by using corresponding α and β from the learning curves in panels a-b of
Experimental setup: All electrical measurements were carried out in a vacuum probe station (Lakeshore CRX 4K) at a base pressure of 5×10−5 Torr. The DC voltage sweep, pulse potentiation/depression, and retention measurements were conducted using source meters (Keithley, 2400) and home-built LabVIEW programs.
As shown in panel a of
Program and read pulses were applied to the drain electrode with the source grounded. The resulting memristive response showed a strong gate dependence, where LTP and LTD were achieved by reversing the gate bias's polarity without changing the source-drain pulsing's polarity, as shown in panel e of
Sulfur vacancies create defect states that can be filled/vacated with a change in the gate bias, as shown in
A representative pulsing scheme to achieve gate-tunable LTP and LTD is shown in panel d of
To explore the MoS2 memtransistor device response in neural networks, unsupervised image recognition learning tasks were performed by simulating a two-layer SNN operating on a simplified STDP algorithm. Previous multilayer perceptron demonstrations using backpropagation reported higher recognition accuracies using fewer neurons in supervised learning. However, as accuracy is not a critical figure of merit here, SNN is more appropriate for continuous and unsupervised learning by exploiting the gate-tunable characteristics of memtransistors to achieve bio-realistic STDP (un)learning functions. Panel a of
Panel a of
After the network trained for 60,000 training MNIST digits, the output neurons were classified and tested against another set of 10,000 digits, resulting in the calculated recognition rates, as shown in panel d of
Since LTP and LTD responses can be tuned by the magnitude of VG, qualitatively diverse learning curves can also be realized by varying the VG profile during constant VD pulsing (presented as square waves in panel d of
Dynamic modulation of synaptic weights opens opportunities for SNNs such as continuous learning, an emerging AI/ML framework that enables lifelong adaptation of learning systems in response to dynamic real-world conditions. Conventional AI/ML models learn and exceed human-level performance in certain tasks, although their inherent rigidity can lead to “catastrophic forgetting” of learned information while processing incoming information. Conversely, continuous learning models learn new tasks without forgetting older high-priority tasks, enabling flexibility to perform diverse AI/ML tasks on the same processing resource. Continuous learning models neuro-cognitive mechanisms in the human brain that are responsible for continually acquiring new knowledge by selectively unlearning and overwriting unused, insignificant knowledge. Therefore, under limited implementation resources, continuous learning models are compelling for AI/ML to unlearn lower priority knowledge to accommodate new task learning, which can be realized using tunable LTP and LTD memtransistor learning curves in a modified SNN.
Panel a of
Two recognition tasks were considered: Task-1 was to learn MNIST digits 0 and 1, and Task-2 was to learn MNIST digits 3 and 4. The synaptic platform was initially trained to perform Task-1 such that all H neurons were trained with training images of 0s and 1s using memtransistor learning curve 1, as shown in panel a of
Conductance maps in panel c of
Panel f of
In conclusion, the atomically thin and gate-tunable nature of 2D materials were leveraged to fabricate MoS2 memtransistors with gate-selective control of individual synapses for dynamic reconfigurability of synapses for unsupervised continuous learning. Monolayer MoS2 on sapphire was critical in enhancing the effect of the gate voltage in the memtransistor device response. Switching LTP-LTD learning behavior was achieved by only reversing the polarity of the gate potential, while further adjustments in the gate amplitude produced diverse learning curves and thus learning behaviors. The resulting learning and unlearning behaviors in a simulated STDP-SNN setup permitted dynamic reallocation of resources for different on-field tasks, thus demonstrating hardware implementation of continuous learning. Further efforts in 2D materials defect engineering, device fabrication optimization, and improved simulation methods are likely to further enhance neuromorphic performance and help realize the full potential of memtransistors as a reconfigurable platform for advanced neuromorphic functionality.
The foregoing description of the exemplary embodiments of the invention has been presented only for the purposes of illustration and description and is not intended to be exhaustive or to limit the invention to the precise forms disclosed. Many modifications and variations are possible in light of the above teaching.
The embodiments were chosen and described to explain the principles of the invention and their practical application to enable others skilled in the art to utilize the invention and various embodiments and with various modifications as are suited to the particular use contemplated. Alternative embodiments will become apparent to those skilled in the art to which the invention pertains without departing from its spirit and scope. Accordingly, the scope of the invention is defined by the appended claims rather than the foregoing description and the exemplary embodiments described therein.
Some references, which may include patents, patent applications, and various publications, are cited and discussed in the description of this invention. The citation and/or discussion of such references is provided merely to clarify the description of the invention and is not an admission that any such reference is “prior art” to the invention described herein. All references cited and discussed in this specification are incorporated herein by reference in their entireties and to the same extent as if each reference was individually incorporated by reference.
This application claims priority to and the benefit of U.S. Provisional Application No. 63/245,997, filed Sep. 20, 2021, which is incorporated herein in its entirety by reference. This application is also a continuation-in-part application of U.S. application Ser. No. 17/036,428, filed Sep. 29, 2020, which itself claims priority to and the benefit of U.S. Provisional Application Ser. No. 62/908,841, filed Oct. 1, 2019, which are incorporated herein in its entireties by reference. This application is also a continuation-in-part application of U.S. application Ser. No. 16/770,662, filed Jun. 8, 2020, which is a U.S. national stage application of PCT Application No. PCT/US2018/065929, filed Dec. 17, 2018, which itself claims priority to and the benefit of U.S. Provisional Application No. 62/599,946, filed Dec. 18, 2017, which are incorporated herein in their entireties by reference.
This invention was made with government support under 1720139 and 1542205 awarded by the National Science Foundation, and DE-NA0003525 awarded by the Department of Energy. The government has certain rights in the invention.
| Number | Date | Country | |
|---|---|---|---|
| 63245997 | Sep 2021 | US | |
| 62908841 | Oct 2019 | US | |
| 62599946 | Dec 2017 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 17036428 | Sep 2020 | US |
| Child | 17939057 | US | |
| Parent | 16770662 | Jun 2020 | US |
| Child | 17036428 | US |
