Memristive device with dropout function, memristive device array and manufacturing method

Abstract

The present invention relates to the field of semiconductor technology, and particularly relates to a memristive device with a dropout function, a memristive device array, and a manufacturing method. The memristive device with a dropout function of the present invention includes a vertical heterostructure formed by successively arranging a first electrode, a ferroelectric material layer, a two-dimensional graphene material layer, a volatile device dielectric layer, and a silver electrode layer by layer. Among them, the two-dimensional graphene material layer serves as a barrier layer for silver ions. With the help of the two-dimensional graphene material layer, the memristive device of the embodiment of the present invention realizes the combination of volatile and non-volatile dynamic behaviors. The random dynamic growth process of volatile silver conductive filaments in the volatile device dielectric layer can suppress sneak current while realizing the dropout function, and the non-volatile ferroelectric material layer can achieve high-performance synaptic plasticity behavior through ferroelectric polarization reversal.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor technology, and particularly to a memristive device with a dropout function, a memristive device array, and a manufacturing method thereof. Background Art

[0002] As the chip manufacturing process approaches the physical limit of semiconductor materials, Moore's Law is gradually failing. At the same time, with the advent of the big data and artificial intelligence era, it brings insurmountable "power consumption wall" and "memory wall" problems to the von Neumann architecture system based on traditional complementary metal oxide semiconductor transistors (CMOS). A memristor is a new type of two-terminal electronic device whose resistance value is related to the flowing charge. With characteristics such as high density, low latency, low power consumption, continuously adjustable resistance state, and the ability to perfectly simulate neuron synapses, and its physical mechanism of intrinsic computing, it brings new hope for solving the von Neumann bottleneck problem. As Figure 1 shown, a memristor array based on a crossbar switch structure can map a numerical matrix to the analog conductance values of each node in the memristor crossbar array, and thus perform efficient multiply-accumulate (MAC) operations in a large-scale parallel manner based on Ohm's Law and Kirchhoff's Law. The MAC operation is one of the main operations in neural networks. Due to high data density, it consumes a large amount of computing time and power consumption. However, using a memristor array can complete the MAC operation in one time cycle, avoiding the time and energy overhead caused by repeated data transfer between the storage and computing units, and fundamentally solving the "memory wall" and "power consumption wall" problems caused by the traditional von Neumann architecture.

[0003] However, the sneak current problem of the memristor array limits the large-scale integration of the memristor array. At the same time, the memristive neural network will have an overfitting problem during small-sample or large-scale network training, which limits the recognition accuracy of the memristive neural network. Summary of the Invention

[0004] In order to solve the overfitting problem that occurs in the memristive neural network during small-sample or large-scale network training, the present invention provides a memristive device with a dropout function, including:

[0005] A vertical heterostructure formed by sequentially arranging a first electrode, a ferroelectric material layer, a two-dimensional graphene barrier layer, a volatile device dielectric layer, and a silver electrode layer by layer. Among them, the stacked structure of the first electrode, the ferroelectric material layer, and the two-dimensional graphene material layer serves as the synaptic plasticity functional structure layer of the memristive device, and the two-dimensional graphene material layer, the volatile device dielectric layer, and the silver electrode serve as the random switching functional structure layer of the memristive device.

[0006] Optionally, the first electrode is a gold electrode or an inert metal electrode, the ferroelectric material layer is one of two-dimensional α-In2Se3, two-dimensional transition metal dichalcogenides, two-dimensional group-V sulfides, two-dimensional layered perovskites, two-dimensional indium selenide, two-dimensional copper indium phosphorus sulfide, barium titanate, lead zirconate titanate, HZO, and the volatile device dielectric layer is one of two-dimensional hexagonal boron nitride, two-dimensional transition metal dichalcogenides, and oxide dielectric layers.

[0007] Optionally, the two-dimensional transition metal dichalcogenide is MX2, where M is W / Mo and X is S / Se / Te.

[0008] Optionally, when the two-dimensional ferroelectric material layer is two-dimensional α-In2Se3, the two-dimensional α-In2Se3 has out-of-plane ferroelectric polarization reversal.

[0009] Optionally, the first electrode is located on the surface of the substrate or the silver electrode is located on the surface of the substrate.

[0010] The present invention also provides a manufacturing method of a memristive device with a dropout function, including:

[0011] Providing a substrate, forming a first electrode on the surface of the substrate; sequentially forming a ferroelectric material layer, a two-dimensional graphene material layer, and a volatile device dielectric layer on the surface of the first electrode; selectively etching the ferroelectric material layer, the two-dimensional graphene material layer, and the volatile device dielectric layer; forming a silver electrode on the surface of the volatile device dielectric layer;

[0012] Or providing a substrate, forming a silver electrode on the surface of the substrate; sequentially forming a volatile device dielectric layer, a two-dimensional graphene material layer, and a ferroelectric material layer on the surface of the silver electrode; selectively etching the ferroelectric material layer, the two-dimensional graphene material layer, and the volatile device dielectric layer; forming a first electrode on the surface of the ferroelectric material layer.

[0013] Optionally, the specific process of forming the first electrode includes: forming a photoresist first electrode pattern on the surface of the substrate or the ferroelectric material layer by electron beam lithography, then preparing and forming a first electrode metal layer by electron beam evaporation process, and stripping the photoresist first electrode pattern to form the first electrode.

[0014] Optionally, the specific process of forming the silver electrode includes: forming a photoresist first electrode pattern on the surface of the volatile device dielectric layer or the substrate by electron beam lithography, then preparing and forming a silver metal layer by electron beam evaporation process, and stripping the photoresist first electrode pattern to form the silver electrode.

[0015] Optionally, the volatile device dielectric layer is one of two-dimensional hexagonal boron nitride, two-dimensional transition metal dichalcogenide, and oxide dielectric layer. When the volatile device dielectric layer is two-dimensional hexagonal boron nitride or two-dimensional transition metal dichalcogenide, a dry transfer process is used to form the volatile device dielectric layer on the surface of the two-dimensional graphene material layer. When the volatile device dielectric layer is one of the oxide dielectric layers, a deposition process is used to form the volatile device dielectric layer on the surface of the two-dimensional graphene material layer.

[0016] The present invention also provides a memristive device array with a dropout function. The memristive device array has a crossbar switch array structure that suppresses sneak current, and a memristive device with a dropout function as described above is connected at the cross position.

[0017] In summary, the advantages and beneficial effects of the present invention are as follows:

[0018] The memristive device with a dropout function of the present invention includes a vertical heterostructure formed by sequentially arranging a first electrode, a ferroelectric material layer, a two-dimensional graphene material layer, a volatile device dielectric layer, and a silver electrode layer by layer. Among them, the two-dimensional graphene material layer acts as a barrier layer for silver ions. With the help of the two-dimensional graphene material layer, the memristive device of the embodiment of the present invention realizes the combination of volatile and non-volatile dynamic behaviors. The random dynamic growth process of volatile silver conductive filaments in the volatile device dielectric layer can suppress sneak current while realizing the dropout function, and the non-volatile ferroelectric material layer can realize high-performance synaptic plasticity behavior through ferroelectric polarization reversal. Description of the Drawings

[0019] Figure 1 Shown is a schematic structural diagram of a multiply-accumulate operation (MAC) based on a memristive neural network in the prior art.

[0020] Figure 2 Shown is a schematic diagram of the working principle of the Dropout algorithm in the prior art.

[0021] Figure 3 Shown is a schematic structural diagram of a memristive device with a dropout function according to an embodiment of the present invention.

[0022] Figure 4 Shown is a schematic flowchart of a manufacturing method of a memristive device with a dropout function according to an embodiment of the present invention.

[0023] Figure 5 Shown is a schematic structural diagram of a memristive device array with a dropout function according to an embodiment of the present invention.

[0024] Figure 6The figure shows the switching characteristics of the memristive device according to an embodiment of the present invention.

[0025] Figure 7 The figure shows the linear synaptic plasticity characteristics of the memristive device according to an embodiment of the present invention. Detailed implementation manners

[0026] For the convenience of those skilled in the art, the present invention will be further described in detail below in conjunction with specific embodiments.

[0027] During the machine learning process, if the number of training samples is too small or the network scale is too complex, the trained model is likely to exhibit overfitting. This means that the network shows poor generalization ability. Even if the accuracy of the training data is very high, the ability to predict data from the training set is very limited. The Dropout algorithm can effectively solve the occurrence of overfitting. As Figure 2 shown, during the training process, a part of the neuron connections are randomly disconnected with a fixed probability. This averaging process dilutes the noise in the training data, simplifies the complex structure of the network, and enhances the generalization ability of the network in the case of insufficient sample quantity.

[0028] In order to implement the Dropout algorithm on hardware, a memristive device with a random dynamic switching mechanism needs to be designed. Many research reports show that memristive devices based on silver conductive filaments have intrinsic random characteristics, but there is no memristive device that can implement the Dropout function yet.

[0029] Therefore, an embodiment of the present invention provides a memristive device with a Dropout function. Please refer to Figure 3 , including:

[0030] A vertical heterostructure formed by sequentially arranging a first electrode 10, a ferroelectric material layer 20, a two-dimensional graphene material layer 30, a volatile device dielectric layer 40, and a silver electrode 50 layer by layer.

[0031] The vertical heterostructure is a van der Waals heterostructure with self-selection ability. The stacked structure of the first electrode 10, the ferroelectric material layer 20, and the two-dimensional graphene material layer 30 serves as the synaptic plasticity structure of the memristive device, and the two-dimensional graphene material layer 30, the volatile device dielectric layer 40, and the silver electrode 50 serve as the random switching structure of the memristive device.

[0032] In this embodiment, the first electrode is a gold electrode. In other embodiments, the first electrode can also be an inert metal electrode such as TiW, TiN, Ta, TaN, Pt, or W.

[0033] In this embodiment, the ferroelectric material layer is two-dimensional α-In2Se3, and the two-dimensional α-In2Se3 has out-of-plane ferroelectric polarization reversal.

[0034] Since two-dimensional α-In2Se3 is a layered ferroelectric material with stable out-of-plane ferroelectric polarization reversal performance, please refer to Figure 7 , and gradually adjust the ferroelectric polarization reversal of α-In2Se3 through a pulsed electric field to adjust the Schottky barrier height at the interface. Because the Schottky barrier at the interface changes due to ferroelectric reversal, the conductance resistance value can change continuously, resulting in a change in the device conductance. Therefore, it can be used to achieve a series of synaptic plasticity behaviors, providing a basis for realizing non-volatile conductance continuously adjustable synaptic plasticity behaviors.

[0035] In other embodiments, the ferroelectric material layer can also be one of two-dimensional transition metal dichalcogenides, two-dimensional group-V chalcogenides, two-dimensional layered perovskites, two-dimensional indium selenide, two-dimensional copper indium phosphorus sulfide, or one of barium titanate, lead zirconate titanate, HZO, etc. Among them, the two-dimensional transition metal dichalcogenide is MX2, where M is W / Mo and X is S / Se / Te, such as MoS2, MoSe2, WSe2, etc.

[0036] Since the stacked structure of the first electrode 10, the ferroelectric material layer 20, and the two-dimensional graphene material layer 30 serves as the synaptic plasticity structure of the memristive device, high-linearity and high-symmetry weight updates can be achieved by controlling the pulse coding method.

[0037] The two-dimensional graphene material layer 30, the volatile device dielectric layer 40, and the silver electrode 50 serve as the random switching structure of the memristive device.

[0038] In this embodiment, the volatile device dielectric layer is two-dimensional hexagonal boron nitride (hBN). The strong in-plane atomic bonding of the two-dimensional hexagonal boron nitride layer provides a platform for an ultra-low off-state current and volatile silver conductive filaments. Because hBN serves as a volatile dielectric layer and has a very high bandgap, when the silver conductive filaments are not connected, the device is in the off state and the off-state current is extremely low, so it provides a platform for the ultra-low off-state current. Please refer to Figure 5 , when the voltage is less than the threshold voltage (0.9 V), the current of the device is only 10 fA, so that the current of the non-selected device is extremely low, suppressing the sneak current.

[0039] Since silver is an active electrode, under the action of an electric field, silver is oxidized into silver ions and then migrates in the hBN layer. When it encounters electrons, it will be reduced to silver atoms. At this time, silver conductive filaments are formed and gradually grow. Due to the high insulation of hBN and the tunneling current mechanism, the current is extremely low during the above process. Once the silver conductive filaments connect the silver electrode and the two-dimensional graphene material layer, the current will increase sharply, and the corresponding device remains in the on state. When the electric field is removed, due to the surface tension and concentration gradient of the silver conductive filaments, the silver conductive filaments will break spontaneously, thus forming a volatile non-linear I-V characteristic, asFigure 6 as shown

[0040] Meanwhile, the strong planar atomic bonding energy of the two-dimensional graphene material layer in the embodiment of the present invention effectively blocks the diffusion of silver conductive filaments. With the help of the two-dimensional graphene material layer, the memristive device in the embodiment of the present invention realizes the combination of volatile and non-volatile dynamic behaviors. The random dynamic growth process of volatile silver conductive filaments in hBN can suppress the sneak current while realizing the dropout function, and the non-volatile two-dimensional α-In2Se3 can realize high-performance synaptic plasticity behavior through ferroelectric polarization reversal.

[0041] In other embodiments, the volatile device dielectric layer can also be one of two-dimensional transition metal dichalcogenides, SiO2, Al2O3, HfO2, Ta2O5 and other oxide dielectric layers, where the two-dimensional transition metal dichalcogenide is MX2, where M is W / Mo and X is S / Se / Te, such as MoS2, MoSe2, WSe2, etc.

[0042] The embodiment of the present invention also provides a manufacturing method of the above-mentioned memristive device with a dropout function, including:

[0043] Providing a substrate and forming a first electrode on the surface of the substrate;

[0044] Sequentially forming a ferroelectric material layer, a two-dimensional graphene material layer, and a volatile device dielectric layer on the surface of the first electrode;

[0045] Selectively etching the ferroelectric material layer, the two-dimensional graphene material layer, and the volatile device dielectric layer;

[0046] Forming a silver electrode on the surface of the volatile device dielectric layer.

[0047] In this embodiment, the substrate is a silicon substrate or a silicon substrate with a silicon oxide layer. In other embodiments, the substrate can also be a substrate of other materials, or the surface of the substrate has one or more electrode layers, and a memristive device with a dropout function is formed on the surface of the electrode layer.

[0048] The memristive device with a dropout function is sequentially provided with a first electrode, a ferroelectric material layer, a two-dimensional graphene blocking layer, a volatile device dielectric layer, and a silver electrode on the surface of the substrate, or a silver electrode, a volatile device dielectric layer, a two-dimensional graphene blocking layer, a ferroelectric material layer, and a first electrode can also be sequentially provided on the surface of the substrate.

[0049] In this embodiment, since the first electrode is a gold electrode, the specific process for forming the first electrode includes: forming a photoresist first electrode pattern on the surface of the substrate or the ferroelectric material layer by electron beam lithography, then preparing and forming a first electrode metal layer by electron beam evaporation process, and stripping the photoresist first electrode pattern to form the first electrode.

[0050] The specific process for forming the silver electrode includes: forming a photoresist first electrode pattern on the surface of the dielectric layer of the volatile device or the surface of the substrate by electron beam lithography, then preparing and forming a silver metal layer by electron beam evaporation process, and stripping the photoresist first electrode pattern to form the silver electrode.

[0051] In this embodiment, the two-dimensional ferroelectric material layer is two-dimensional α-In2Se3, and the two-dimensional α-In2Se3 has out-of-plane ferroelectric polarization reversal. The two-dimensional α-In2Se3 is formed on the surface of the first electrode by a dry transfer process.

[0052] The dielectric layer of the volatile device is two-dimensional hexagonal boron nitride or two-dimensional transition metal dichalcogenide, and the dielectric layer of the volatile device is formed on the surface of the two-dimensional graphene material layer by a dry transfer process.

[0053] In other embodiments, when the dielectric layer of the volatile device is one of oxide dielectric layers such as SiO2, Al2O3, HfO2, Ta2O5, etc., the dielectric layer of the volatile device is formed on the surface of the two-dimensional graphene material layer by a deposition process.

[0054] The embodiment of the present invention also provides another manufacturing method of the above-mentioned memristive device with a dropout function, including:

[0055] Providing a substrate and forming a silver electrode on the surface of the substrate;

[0056] Sequentially forming a dielectric layer of a volatile device, a two-dimensional graphene material layer, and a ferroelectric material layer on the surface of the silver electrode;

[0057] Selectively etching the ferroelectric material layer, the two-dimensional graphene material layer, and the dielectric layer of the volatile device;

[0058] Forming a first electrode on the surface of the ferroelectric material layer.

[0059] The embodiment of the present invention also provides a memristive device array with a dropout function. Please refer to Figure 4 , the memristor array has a crossbar switch array structure for suppressing sneak current, and a memristive device with the above-mentioned dropout function is connected at the cross position of the upper electrode and the lower electrode.

[0060] Since the Dropout algorithm improves the generalization ability of the network by randomly discarding some synaptic connections with a certain probability, thus avoiding the overfitting problem. During the training phase, a Bernoulli probability function is applied to each synapse. The random switching characteristic of the silver conductive wire can well implement the Bernoulli probability function. Since the distribution during the growth process of the silver conductive wire is random, the resulting periodic threshold turn-on voltage is also random. At a specific read voltage, the device will achieve the on-state or off-state with a certain probability. As Figure 5 and Figure 6 shown, when the threshold turn-on voltage (close to 0.9V) is turned on, the probability of achieving the on-state is close to 60%, similar to the randomly generated 1 or 0 in the Bernoulli function. Therefore, it is expected to implement the dropout function from the device.

[0061] Finally, it is stated that any modification or equivalent replacement of some or all of the technical features carried out by relying on the device structure of the present invention and the technical solutions of the embodiments, and the essence obtained without departing from the corresponding technical solutions of the present invention, all belong to the patent scope of the device structure of the present invention and the embodiments.

Claims

1. A memristive device with dropout function, characterized in that, Including: A vertical heterostructure formed by sequentially setting a first electrode, a ferroelectric material layer, a two-dimensional graphene barrier layer, a volatile device dielectric layer, and a silver electrode layer by layer. Among them, the stacked structure of the first electrode, the ferroelectric material layer, and the two-dimensional graphene material layer serves as the synaptic plasticity functional structure layer of the memristive device, and the two-dimensional graphene material layer, the volatile device dielectric layer, and the silver electrode serve as the random switching functional structure layer of the memristive device.

2. The memristive device with dropout function according to claim 1, wherein, The first electrode is an inert metal electrode, the ferroelectric material layer is one of two-dimensional transition metal dichalcogenides, two-dimensional group-V chalcogenides, two-dimensional layered perovskites, two-dimensional indium selenide, two-dimensional copper indium phosphosulfide, barium titanate, lead zirconate titanate, HZO, and the volatile device dielectric layer is one of two-dimensional hexagonal boron nitride, two-dimensional transition metal dichalcogenides, and oxide dielectric layers.

3. A memristive device with a dropout function according to claim 2, characterized in that, The two-dimensional transition metal dichalcogenide is MX2, where M is W or Mo, and X is S or Se or Te.

4. A memristive device with dropout function according to claim 1, characterized in that, When the two-dimensional ferroelectric material layer is two-dimensional α-In2Se3, the two-dimensional α-In2Se3 has out-of-plane ferroelectric polarization reversal.

5. A memristive device with dropout function according to claim 1, characterized in that, The first electrode is located on the surface of the substrate or the silver electrode is located on the surface of the substrate.

6. A manufacturing method of a memristive device with a dropout function, characterized in that, Including: Providing a substrate and forming a first electrode on the surface of the substrate; Sequentially forming a ferroelectric material layer, a two-dimensional graphene material layer, and a volatile device dielectric layer on the surface of the first electrode; Selectively etching the ferroelectric material layer, the two-dimensional graphene material layer, and the volatile device dielectric layer; forming a silver electrode on the surface of the volatile device dielectric layer; Or providing a substrate and forming a silver electrode on the surface of the substrate; Sequentially forming a volatile device dielectric layer, a two-dimensional graphene material layer, and a ferroelectric material layer on the surface of the silver electrode; selectively etching the ferroelectric material layer, the two-dimensional graphene material layer, and the volatile device dielectric layer; forming a first electrode on the surface of the ferroelectric material layer.

7. The manufacturing method of a memristive device with a dropout function according to claim 6, characterized in that, The specific process for forming the first electrode includes: forming a photoresist first electrode pattern on the surface of the substrate or the ferroelectric material layer by electron beam lithography, then preparing and forming a first electrode metal layer by electron beam evaporation process, and stripping the photoresist first electrode pattern to form the first electrode.

8. The manufacturing method of a memristive device with dropout function according to claim 6, characterized in that, The specific process for forming the silver electrode includes: forming a photoresist first electrode pattern on the surface of the volatile device dielectric layer or the substrate by electron beam lithography, then preparing and forming a silver metal layer by electron beam evaporation process, and stripping the photoresist first electrode pattern to form the silver electrode.

9. The manufacturing method of a memristive device with a dropout function according to claim 6, characterized in that, The volatile device dielectric layer is one of two-dimensional hexagonal boron nitride, two-dimensional transition metal dichalcogenides, and oxide dielectric layers. When the volatile device dielectric layer is two-dimensional hexagonal boron nitride or two-dimensional transition metal dichalcogenides, a dry transfer process is used to form the volatile device dielectric layer on the surface of the two-dimensional graphene material layer. When the volatile device dielectric layer is one of the oxide dielectric layers, a deposition process is used to form the volatile device dielectric layer on the surface of the two-dimensional graphene material layer.

10. A memristive device array with dropout function, characterized in that, The memristor device array is a crossbar switch array structure with suppressed sneak current, and the cross positions are connected with the memristor devices having the dropout function as described in any one of claims 1 to 5.

Images