An adaptive homogeneous reservoir computing system based on oxide memristors

Abstract

The application belongs to the technical field of micro-nano electronics, and discloses an adaptive homogeneous reservoir computing system based on oxide memristor, which comprises a reservoir layer containing a plurality of volatile dynamic memristors with nonlinear short-term decay memory characteristics; a fully connected readout layer containing a plurality of nonvolatile memristors with multi-value long-term memory characteristics; and a feedback loop for adjusting the bias voltage applied during the state decay of the dynamic memristor, wherein the two types of memristors adopt the same laminated structure containing a titanium oxide layer. The application solves the problems of non-uniform material system of the reservoir layer and the readout layer, and single time scale of the reservoir system and inability to adjust, and the method proposed by the application can adaptively adjust the time scale of the reservoir, so that it can better adapt to different types of input signals, thereby improving the reservoir computing performance and completing the dynamic information recognition function.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of micro-nano electronics, and more particularly relates to an adaptive homogeneous reservoir computing system based on oxide memristors. BACKGROUND

[0002] With the high-efficiency parallel computing capability, the hardware artificial neural network greatly improves the computing power of machine learning. Artificial neural networks mainly include two categories: feedforward neural networks (FNNs) and recurrent neural networks (RNNs). Although RNNs can process time-series information, the large number of complex internal loop connections hinders its hardware implementation.

[0003] Reservoir networks, as a derivative of RRNs, can map low-latitude time series nonlinearity into linearly separable reservoir states, providing an optional network framework for processing spatiotemporal information. The main body of the reservoir network consists of two parts: the reservoir layer and the readout layer. The reservoir layer is used to map sequence information to a high-dimensional feature space and is composed of fixed nonlinear nodes. The readout layer is used for simple linear processing and is composed of non-volatile weight units that need to be trained. This network has a very low hardware implementation cost, and since only the weights of the readout layer need to be trained, the training cost is greatly reduced.

[0004] However, due to the different functions of the reservoir layer and the readout layer, most of the current reservoir hardware systems based on new memory use different materials to realize their reservoir layer and readout layer, which limits the improvement of system integration and process compatibility. On the other hand, the time characteristics of most current reservoirs based on microelectronic devices are determined by the manufacturing process. Once the device is prepared, the time characteristics cannot be changed. Therefore, most current reservoir systems can only process sequence information of a specific time scale, which severely limits the scalability of the reservoir system in the time scale. SUMMARY

[0005] In view of the above defects or improvement needs of the prior art, the application provides an adaptive homogeneous reservoir computing system based on oxide memristors, which aims to improve system integration and process compatibility and improve scalability in the time scale.

[0006] To achieve the above purpose, the application provides an adaptive homogeneous reservoir computing system based on oxide memristors, which comprises:

[0007] The reservoir layer comprises a plurality of volatile dynamic memristors with nonlinear short-term decay memory characteristics, which are used to receive time-series signals and extract feature information; the dynamic memristor comprises a top electrode, a bottom electrode, and a first stack located in the middle;

[0008] The full-connection readout layer comprises a plurality of nonvolatile resistive memory devices with multi-value long-term memory characteristics, and is configured to receive feature information output by the reservoir layer and output a recognition result.

[0009] The feedback loop is configured to adjust the bias voltage applied to the dynamic memristor in the reservoir layer during the decay of the dynamic memristor according to the recognition result output by the full-connection readout layer, so that the decay time constant of the dynamic memristor matches the time scale of the input time sequence signal.

[0010] The first stack and the second stack are both formed by stacking a first oxide layer and a second oxide layer, the first oxide layer is a titanium oxide layer, the material of the second oxide layer has a lower electron affinity than titanium oxide and has a nonvolatile resistive switching property, and the oxygen vacancy content of the titanium oxide layer in the first stack is less than a threshold value for generating a conductive filament, and the oxygen vacancy content of the titanium oxide layer in the second stack is greater than or equal to the threshold value for generating a conductive filament.

[0011] Optionally, the full-connection readout layer has a plurality of output terminals, and the recognition result comprises recognition of the content category of the time sequence signal and recognition of the decay speed of the dynamic memristor.

[0012] The feedback loop is configured to adjust the bias voltage applied to the dynamic memristor in the reservoir layer during the decay of the dynamic memristor according to the recognition result output by the full-connection readout layer, so that the decay time constant of the dynamic memristor matches the time scale of the input time sequence signal.

[0013] Optionally, the material of the second oxide layer is aluminum oxide or hafnium oxide.

[0014] Optionally, the dynamic memristor comprises a substrate and a bottom electrode, a first stack and a top electrode stacked on the substrate in order from bottom to top, and the titanium oxide layer in the first stack is located on the side close to the top electrode; the nonvolatile resistive memory device comprises a substrate and a bottom electrode, a second stack and a top electrode stacked on the substrate in order from bottom to top, and the titanium oxide layer in the second stack is located on the side close to the bottom electrode.

[0015] Optionally, the bottom electrode is adhered to the substrate by a bottom electrode adhesive layer, and the material of the bottom electrode adhesive layer is titanium.

[0016] Optionally, the first stack and the second stack are both prepared by an atomic layer deposition process, wherein,

[0017] The titanium oxide layer in the first stack is deposited at a temperature of 110-200°C using tetrakis(dimethylamino) titanium as a titanium source precursor and H2O as an oxygen source.

[0018] The titanium oxide layer in the second stack is deposited at a temperature of 250-350 DEG C using tetrakis(dimethylamino) titanium as a titanium source precursor and H2O as an oxygen source.

[0019] Optionally, the thickness of the same material layer in the first stack and the second stack is the same, wherein the thickness of the titanium oxide layer is 35-45 nm, and the thickness of the second oxide layer is 4-6 nm.

[0020] Optionally, the material of the top electrode and the bottom electrode is selected from Al, Pt, Au, W, Ti or TiN.

[0021] Optionally, the system further comprises an input layer for converting a target to be recognized into a time sequence signal and inputting the time sequence signal into the reservoir layer.

[0022] Optionally, the system further comprises a control module for applying the bias voltage to the dynamic memristor only during a period when no time sequence signal voltage pulse is applied to the dynamic memristor.

[0023] Overall, compared with the prior art, the above technical scheme conceived by the present application mainly has the following beneficial effects:

[0024] 1. The adaptive homogeneous reservoir computing system based on oxide memristors proposed in the present application adopts a stack based on the same material for the reservoir layer and the readout layer, the first oxide layer is a titanium oxide layer, and the second oxide layer is a material with a lower electron affinity than the titanium oxide and a non-volatile resistive switching performance. By controlling the oxygen vacancy content in the titanium oxide layer, the dynamic memristor has a nonlinear short-term decay memory characteristic, and the non-volatile memristor has a long-term memory characteristic. The second oxide layer has a lower electron affinity than the titanium oxide, so that the dynamic memristor has a larger resistance window, and a more abundant reservoir state can be achieved. The present application realizes a homogeneous reservoir computing system, and improves the integration potential and process compatibility of the system. When a time sequence signal is input, a time sequence signal voltage pulse is applied to the dynamic memristor, the conductance of the dynamic memristor changes, the time characteristic is extracted and used for subsequent network reasoning and recognition. When the time sequence signal voltage pulse is removed, the dynamic memristor enters a conductance decay stage. The present application proposes a feedback loop to apply a bias voltage during the conductance state decay of the dynamic memristor. By utilizing the ion drift and diffusion phenomenon in the interface resistive switching layer, a short-term memory characteristic with adjustable decay time constant is realized, and thus the time scale of the reservoir can be modulated. In summary, the present application solves the problems of non-uniform material system of the reservoir layer and the readout layer, and single and unadjustable time scale of the reservoir system. The method proposed in the present application can adaptively adjust the time scale of the reservoir, so that it can better adapt to different types of input signals, thereby improving the reservoir computing performance and completing the dynamic information recognition function.

[0025] 2. In an optional embodiment, the full-connection readout layer can output the identification result of the decay speed of the dynamic memristor, and directly modulate the bias voltage based on the result, so that the modulation can be quickly performed, and the calculation cost is reduced.

[0026] 3. In an optional embodiment, the material of the second oxide layer is aluminum oxide or hafnium oxide, both of which have a lower electron affinity than titanium oxide and have non-volatile resistive switching properties. When the titanium oxide layer is stacked with the titanium oxide layer, a larger interface barrier can be formed, so that the dynamic memristor has a larger resistance window, and a more abundant reservoir state can be achieved. In addition, when the titanium oxide layer is stacked with the titanium oxide layer with a higher oxygen content, a stable conductive filament can be formed under external voltage stimulation, and the non-volatility is achieved.

[0027] 4. In an optional embodiment, the preparation process of the titanium oxide layer is limited to form a stack with different oxygen vacancy concentrations, thereby achieving different resistive switching characteristics. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a structural schematic diagram of an adaptive homogeneous reservoir computing system based on oxide memristors in an embodiment of the present application;

[0029] Figure 2 is a schematic diagram of the conductance change of a dynamic memristor under the joint action of a timing signal and a bias voltage in an embodiment of the present application;

[0030] Figure 3 is a structural schematic diagram of a dynamic memristor in an embodiment of the present application;

[0031] Figure 4 is a schematic diagram of the decay change of a dynamic memristor under different bias voltages in an embodiment of the present application;

[0032] Figure 5 is a structural schematic diagram of a dynamic memristor in an embodiment of the present application;

[0033] Figure 6 is a long-term depression cycle measurement diagram of a non-volatile memristor in an embodiment of the present application;

[0034] Figure 7 is a schematic diagram of the prediction accuracy of a system in an embodiment of the present application;

[0035] Figure 8 is a comparison diagram of the time scale range that can achieve accurate prediction under different bias voltages. DETAILED DESCRIPTION

[0036] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0037] As shown in Figure 1 The structure of the adaptive homogeneous reservoir pool computing system based on oxide memristor in an embodiment of the present application is shown in the schematic diagram, which includes a reservoir layer, a fully connected readout layer and a feedback loop.

[0038] The reservoir layer contains a plurality of volatile dynamic memristors with nonlinear short-term decay memory characteristics, which are used to receive time series signals and extract feature information; the dynamic memristor contains a top electrode, a bottom electrode and a first stack in the middle.

[0039] The fully connected readout layer contains a plurality of non-volatile memristors with multi-value long-term memory characteristics, which are used to receive the feature information output by the reservoir layer and output the recognition result, the non-volatile memristor contains a top electrode, a bottom electrode and a second stack in the middle.

[0040] The first stack and the second stack are both formed by the contact of the first oxide layer and the second oxide layer, the first oxide layer is a titanium oxide layer, the material of the second oxide layer is a material with lower electron affinity than titanium oxide and non-volatile resistive switching performance, and the oxygen vacancy content of the titanium oxide layer in the first stack is less than the content threshold value of the conductive filament, and the oxygen vacancy content of the titanium oxide layer in the second stack is greater than or equal to the content threshold value of the conductive filament.

[0041] The feedback loop is used to adjust the bias voltage applied to the dynamic memristor during the decay of the dynamic memristor in the reservoir layer in the reverse direction according to the recognition result output by the fully connected readout layer, so that the decay time constant of the dynamic memristor matches the time scale of the input time series signal.

[0042] In the above structure, the reservoir layer is composed of volatile dynamic memristors, the conductance of the dynamic memristor will change temporarily after being stimulated by external voltage, and will decay and recover to the original state according to the nonlinear short-term decay memory characteristics, thereby realizing the nonlinear mapping function of the reservoir layer to the time series. The fully connected readout layer is composed of non-volatile memristors, one non-volatile memristor as a non-volatile weight unit of the readout layer, the non-volatile memristor can realize multi-value storage, thereby ensuring the adjustment of the weight represented by it during training, and due to its long-term memory characteristics, i.e. non-volatility, the adjusted weight will not change automatically, ensuring the stability of its weight.

[0043] When the specific structure is designed, the structure of the dynamic memristor and the nonvolatile memristor is basically the same, both of which comprise a stack between the top electrode and the bottom electrode, and the stack is composed of a first oxide layer and a second oxide layer, oxygen vacancies are formed in the oxide layer, and the material of the first oxide layer of the two stacks is the same, which is a titanium oxide layer, and the material of the second oxide layer of the two stacks is the same, which is a material with a lower electron affinity than titanium oxide and having a nonvolatile resistive switching property.

[0044] In the dynamic memristor, the resistive switching mechanism is as follows: when a reverse voltage is applied to one side electrode of the titanium oxide layer, the positively charged oxygen vacancies in the titanium oxide layer move away from the interface, so that the conduction band on the titanium oxide side at the interface is raised, and for the same reason, the oxygen vacancies in the second oxide layer migrate to the interface, and the conduction band on the second oxide layer side at the interface is lowered, finally resulting in a decrease in the interface barrier, and an increase in the effective conductance of the memristor. At this time, since the oxygen vacancy content of the titanium oxide is less than the content threshold for forming a conductive filament, a stable conductive filament will not be formed under external excitation, and therefore, after the external excitation is removed, due to the concentration gradient of oxygen vacancies between the interface and the bulk, the oxygen vacancies spontaneously diffuse, the interface barrier gradually increases, the conductance gradually decreases, and the original state is restored. The above process is the short-term decay memory characteristic. As can be seen, the conductance value of the dynamic memristor is determined by the barrier height of the contact interface of the titanium oxide layer / second oxide layer in the first stack. In the present application, the electron affinity of the material used in the second oxide layer is lower than that of titanium oxide, so the titanium oxide layer / second oxide layer can form a higher interface barrier, thereby enabling the dynamic memristor to have a larger resistance window and to realize a more abundant reservoir state.

[0045] In the nonvolatile memristor, since the second oxide layer is a material with a nonvolatile resistive switching property, the following resistive switching mechanism can be realized: when a reverse voltage is applied to one side electrode of the second oxide layer, positively charged oxygen vacancies migrate from the second oxide layer to the titanium oxide layer. Since the oxygen vacancy content of the titanium oxide layer is greater than or equal to the content threshold for forming a conductive filament, a large number of oxygen vacancies can form a conductive channel, i.e., a conductive filament, by gathering locally, and the memristor exhibits low resistance; when a forward voltage is applied to one side electrode of the second oxide layer, oxygen vacancies migrate and the conductive filament breaks, and the memristor exhibits high resistance. Once the filament is formed, its conductance can remain stable after the external excitation is removed, exhibiting a long-term memory characteristic.

[0046] In the present application, the dynamic memristor and the nonvolatile memristor designed as above both use a stack and the same material, only the oxygen vacancy content of the titanium oxide layer is different. The same material makes the dynamic memristor and the nonvolatile memristor have homogeneity, thereby improving the system integration and process compatibility.

[0047] Meanwhile, the application also designs a feedback loop for reversely adjusting the bias voltage applied to the dynamic memristor in the reservoir layer according to the recognition result output by the full-connection readout layer, so as to match the time constant of the decay of the dynamic memristor with the time scale of the input time sequence signal. That is to say, the dynamic memristor receives the pulse excitation of the time sequence signal, changes the conductance under the pulse excitation, and extracts the high-dimensional time-space signal of the pulse, and after the voltage pulse of the time sequence signal is removed, the dynamic memristor enters the nonlinear decay stage according to the short-term decay memory characteristic, at this time, the bias voltage is applied to the dynamic memristor again for modulation, so as to control the decay speed and match the time constant of the decay with the time scale of the input time sequence signal.

[0048] As shown in Figure 2 The conductance change of the dynamic memristor under the joint action of the time sequence signal and the bias voltage is shown in the figure, under the excitation of the time sequence signal voltage pulse V signal , the conductance increases, after the time sequence signal voltage pulse V signal is removed, the bias voltage V bias is applied to modulate the decay constant. In the interval of the time sequence signal voltage pulse V signal , the conductance level G of the device presents a double-exponential decay process, and this phenomenon is the short-term memory characteristic of the dynamic memristor in the application. When the dynamic memristor is used to realize the reservoir, the short-term memory characteristic of the device determines the time characteristic of the reservoir. In order to more accurately describe the time characteristic of the reservoir, a specific number of square wave voltage pulses are applied to the top electrode of the device to make the device reach a specific conductance state, and then the conductance evolution curve with time is fitted by a double exponential to obtain the fast decay time constant τ1 and the slow decay time constant τ2 as well as the corresponding coefficients A1 and A2. These parameters reflect the strength of the correlation between adjacent virtual nodes of the reservoir, and thus can be used to describe the time characteristic of the reservoir. In the application, the bias voltage is applied during the decay to modulate the fast decay time constant τ1 and the slow decay time constant τ2, so as to modulate the decay speed during the period.

[0049] Generally speaking, the higher the matching degree, the more suitable the correlation between the nodes, and the more accurate the prediction (if the correlation is too strong, the information overlap occurs, the reservoir state is the linear superposition of the time sequence information, and the sequence information cannot be reflected; if the correlation is too weak, the reservoir state can only reflect the current input, and the historical information cannot be reflected). Moreover, considering that the time scale of the input time sequence signal is variable, the bias voltage required in the decay stage is different when the time scale of the input time sequence signal is different, in order to adapt to the variability, the application introduces a feedback loop, which can automatically adjust the bias voltage according to the output result of the readout layer, so as to adapt to the time sequence signal with different time scales and realize the scalability of the reservoir system in the time scale.

[0050] In one embodiment, the fully connected readout layer has multiple output terminals, and the recognition result includes the recognition of the content category of the time series signal and the recognition of the decay speed of the dynamic memristor. The bias voltage of the dynamic memristor in the reservoir layer can be reversely adjusted based on the recognition result of the decay speed of the dynamic memristor by the fully connected readout layer. If the output terminal feedback decay speed is too fast, the bias voltage is adjusted to slow down the decay speed of the dynamic memristor. If the output terminal feedback decay speed is too slow, the bias voltage is adjusted to speed up the decay speed of the dynamic memristor. It is understandable that before being put into application, the network needs to be trained, and deep learning is used to enable the network to have content recognition capabilities and the ability to recognize whether the decay speed is too fast or too slow. Taking image recognition as an example, during the training phase, two types of labels can be added to the input image: one type of label represents the image's content, and the other type of label indicates whether the decay time constant of the dynamic memristor is too fast or too slow compared to the time scale of the time series signal when the image is converted into a time series signal and input into the reservoir. Through multiple batches of training, the weights of each node in the readout layer are determined, that is, the conductance of each non-volatile dynamic memristor is determined. After training is completed, it can be directly put into use. When put into use, the state of each non-volatile dynamic memristor in the readout layer is stable. The output terminals of the readout layer will output the content and recognition results, as well as the recognition results of whether the dynamic memristor in the reservoir layer decays too fast or too slowly. The feedback loop adjusts the bias voltage based on the recognition results to match the time scale of the time series signal as closely as possible.

[0051] In one embodiment, the material used for the second oxide layer is aluminum oxide or hafnium oxide. Aluminum oxide or hafnium oxide are both materials with lower electron affinity than titanium oxide and have non-volatile resistive switching properties. When stacked with the titanium oxide layer, they can form a larger interface potential barrier, thereby enabling the dynamic memristor to have a larger resistance window and achieve richer storage states. Moreover, when stacked with the titanium oxide layer with a higher oxygen content, they can form stable conductive filaments under external voltage stimulation and have non-volatility.

[0052] In a specific embodiment, the dynamic memristor includes a substrate, a bottom electrode, a first stack, and a top electrode stacked sequentially on the substrate from bottom to top, wherein the titanium oxide layer in the first stack is located on a side close to the top electrode.

[0053] like Figure 3 The figure shows a schematic structural diagram of a dynamic memristor in one embodiment of the present invention, which comprises, from bottom to top, a substrate 1, a bottom electrode adhesion layer 2, a bottom electrode 3, an AlO y Layer 4, TiO x layer 5 and a top electrode 6.

[0054] The above dynamic memristor can be prepared by the following method:

[0055] 1. Selecting a SiO2 / Si(100) substrate with a size of 1.5 cm x 1.5 cm as a substrate, cleaning the surface and back with acetone, alcohol and deionized water respectively to remove dust particles, organic and inorganic impurities, and obtaining a substrate 1 to be deposited;

[0056] 2. Preparing a bottom electrode 3 by using a magnetron sputtering process; the material of the bottom electrode is preferably Pt, and the thickness of the bottom electrode is 80 nm to 120 nm, for example, 100 nm can be selected; the adhesion layer 2 of the bottom electrode is Ti, and the thickness range is 8 nm to 12 nm, for example, 10 nm can be selected;

[0057] 3. Preparing a TiO x / AlO y stack (AlO y layer 4 / TiO x layer 5) by using an atomic layer deposition process; the Al source precursor is trimethylaluminum (TMA), the oxygen source is H2O, the deposition temperature is 250℃-350℃, for example, 250℃ can be selected, and the deposition thickness is preferably 3 nm to 7 nm, for example, 5 nm can be selected; the Ti source precursor is tetrakis(dimethylamino) titanium (TDMAT), the oxygen source is H2O, the deposition temperature is 110℃-200℃, for example, 120℃ can be selected, and the deposition thickness is 35 nm to 45 nm, for example, 40 nm can be selected; a hard mask method is used to reserve the bottom electrode during the deposition process; the introduction of the TiO x / AlO y interface makes the dynamic memristor have adjustable short-term memory characteristics;

[0058] 4. Preparing a pattern of a top electrode on the interface variable resistance layer by photolithography, and preparing a pattern mask layer of the top electrode by uniform coating, pre-baking, pre-exposure, post-baking, post-exposure and development processes;

[0059] 5. Taking the sample after photolithography to prepare a top electrode 6 by using a sputtering process; the material of the top electrode is preferably Pt, and the thickness range is 90 nm to 110 nm, for example, 100 nm can be selected;

[0060] 6. Taking the sample after the sputtering preparation of the top electrode, soaking in an acetone solution for 20 minutes, peeling off the pattern mask layer of the top electrode, then cleaning with ethanol and deionized water, and blowing dry with a nitrogen gun, to obtain a patterned top electrode, and the size of the top electrode is preferably 50 μm, 100 μm, 150 μm, 200 μm.

[0061] Experiments show that when the dynamic memristor top electrode has a voltage pulse input, the conductance state will change accordingly. When the external voltage pulse is removed, the conductance level will gradually relax back to the initial state. Further experiments show that the short-term memory effect described above can be modulated by the bias voltage of the relaxation process. With the change of the bias voltage value, the relaxation process is accelerated or slowed down, ultimately reflecting the different decay time constants of the device.

[0062] As shown in Figure 4 Figure 3 is a schematic diagram of the decay change of the dynamic memristor under different bias voltages, from which it can be seen that the decay speed of the device is different when the applied bias voltage is different, and the decay speed is obviously accelerated when the bias voltage is increased, thereby proving the effectiveness of the present scheme in modulating the decay process of the device by adjusting the bias voltage.

[0063] In a specific embodiment, the non-volatile memristor comprises a substrate and, stacked on the substrate from bottom to top, a bottom electrode, a second stack and a top electrode, the titanium oxide layer in the second stack being located on the side close to the bottom electrode.

[0064] As shown in Figure 5 Figure 4 is a schematic diagram of the structure of the dynamic memristor in an embodiment of the present application, which comprises, from bottom to top, a substrate 7, a bottom electrode adhesion layer 8, a bottom electrode 9, a TiO x layer 10, an AlO y layer 11 and a top electrode 12.

[0065] The above non-volatile memristor can be prepared by the following method:

[0066] 1. Select a SiO2 / Si(100) substrate with a size of 1.5 cm x 1.5 cm as the substrate, clean the surface and back with acetone, alcohol and deionized water to remove dust particles, organic and inorganic impurities, and obtain a deposition substrate 7;

[0067] 2. Prepare the bottom electrode 9 by using a magnetron sputtering process; the bottom electrode material is preferably Pt, the bottom electrode thickness is 80 nm to 120 nm, for example, 100 nm can be selected; the bottom electrode adhesion layer 8 is Ti, the thickness range is 8 nm to 12 nm, for example, 10 nm can be selected;

[0068] 3. Prepare the AlO y / TiO z bulk resistance variable layer (TiO z layer 10 / AlO y layer 11) by using an atomic layer deposition process; the Ti source precursor is tetrakis(dimethylamino) titanium (TDMAT), the oxygen source is H2O, the deposition temperature is 250°C-350°C, for example, 250°C can be selected, the deposition thickness is 35 nm to 45 nm, for example, 40 nm can be selected; the Al source precursor is trimethylaluminum (TMA), the oxygen source is H2O, the deposition temperature is 250°C-350°C, for example, 250°C can be selected, the deposition thickness is preferably 3 nm to 7 nm, for example, 5 nm can be selected; a hard mask method is used to reserve the bottom electrode during the deposition process; the deposition temperature of the TiO z layer is adjusted so that it can act as an oxygen storage layer, giving the device non-volatile resistance variable characteristics;

[0069] 4. Prepare the pattern of the top electrode on the interface resistive layer by photolithography, and prepare the pattern mask layer of the top electrode by the processes of coating, pre-baking, pre-exposure, post-baking, post-exposure, and development;

[0070] 5. Take the sample after photolithography and use sputtering process to prepare the top electrode 12; the top electrode material is preferably Pt, and the thickness range is 90nm to 110nm, for example, 100nm can be selected;

[0071] 6. Take the sample prepared by sputtering the top electrode, soak it in acetone solution for 20 minutes, peel off the pattern mask layer of the top electrode, then wash it with ethanol and deionized water, and blow it dry with a nitrogen gun to obtain the patterned top electrode. The size of the top electrode is preferably 50μm, 100μm, 150μm, and 200μm.

[0072] like Figure 6 The figure shows a long-term depression cycle measurement diagram of a non-volatile memristor. The non-volatile memristor is subjected to a long-term depression (LTD) / potentiation (LTP) cycle test. The pulse form used to achieve LTD in the test is a reverse step pulse, that is, a continuous pulse with a constant amplitude change in pulse amplitude; the pulse used to achieve LTP is a single forward pulse. With the addition of the pulse, the normalized conductance is linearly modulated and shows very high cycle stability. In each cycle, the conductance of the non-volatile memristor decreases monotonically, and a total of 32 intermediate resistance states are obtained, and a one-step setting from low conductance to high conductance is achieved by a setting pulse. Linear fitting of the LTD part yields a Pearson correlation coefficient of -0.945, which shows that the conductance state has a strong linear correlation with the number of pulses.

[0073] In the above preparation method, the deposition temperature of the titanium oxide layer in the first stack and the second stack is different, the deposition temperature of the titanium oxide layer in the first stack is 110-200 DEG C, and the deposition temperature of the titanium oxide layer in the second stack is 250-350 DEG C. The titanium oxide deposited at low temperature has a low oxygen vacancy concentration and cannot participate in the formation of stable conductive filaments. The titanium oxide deposited at high temperature has a high oxygen vacancy concentration and can provide sufficient oxygen vacancies for the resistance change layer. Therefore, although the materials of the first stack and the second stack are the same, different resistance change characteristics can be achieved by controlling the oxygen vacancy concentration of the titanium oxide layer. Moreover, since the deposition temperature of the titanium oxide layer in the first stack is lower, the other oxide layer with a higher deposition temperature is prepared first, and then the titanium oxide layer is prepared, so that the influence of high temperature on the titanium oxide layer can be avoided. When the second stack is prepared, it is found that the second oxide layer is arranged on the side far from the bottom electrode, and compared with the side close to the bottom electrode, it is easier to modulate under external excitation and reach the corresponding high resistance state. The reason may be that if the second oxide layer is close to the bottom electrode, the adhesion layer Ti of the bottom electrode will adsorb the oxygen ions of the second oxide layer, so that the second oxide layer generates oxygen vacancies without titanium oxide. The obtained device always presents low resistance and is difficult to modulate. According to the above method, the obtained device can be grounded when the system is built, and the timing signal and the bias voltage are directly or after processing applied to the top electrode of the dynamic memristor, so that the dynamic memristor realizes the resistance change mechanism.

[0074] In an embodiment, the material of the top electrode and the bottom electrode is selected from Al, Pt, Au, W, Ti or TiN.

[0075] In an embodiment, the system further comprises an input layer for converting the target to be identified into a timing signal and inputting the reservoir layer.

[0076] In an embodiment, the system further comprises a control module for applying a bias voltage to the dynamic memristor only during the period when the voltage pulse without timing signal is applied to the dynamic memristor.

[0077] In operation, an initial bias voltage V bias is set first, and the timing signal is reversely applied to the top electrode of the dynamic memristor, and the bottom electrode is grounded, so that the current of the dynamic memristor is taken as the extracted feature to enter the subsequent recognition network or other classification algorithm such as ridge regression, so as to complete the classification identification or prediction of the input timing signal, and the bias voltage V bias is adjusted reversely based on the output result, and the process is repeated, so as to improve the prediction accuracy.

[0078] As Figure 7 shown is a system prediction accuracy diagram in an embodiment of the application, and the prediction mean square error MSE is less than or equal to 0.1, which indicates that the prediction accuracy is high.

[0079] As Figure 8 The figure shows the comparison of the time scale range that can achieve accurate prediction under different bias voltages, where the mixed means that the bias voltage can be adaptively adjusted, that is, the scheme of the application, from which it can be seen that the adaptive adjustment of the bias voltage significantly expands the time scale of accurate identification, and the adaptive adjustment of the decay time constant of the reservoir system through the feedback loop to match the time scale of the specific input signal can significantly improve the time scale of accurate identification of the network. The advantage of this system is that compared with the traditional reservoir, this system can process more extensive time sequence information with a wider frequency range.

[0080] In summary, the adaptive homogeneous reservoir computing system based on the oxide memristor proposed in the application uses a titanium oxide material-based stack for the reservoir layer and the readout layer, realizes a homogeneous reservoir computing system, and improves the integration potential and process compatibility of the system; a single dynamic memristor constitutes a reservoir unit, when a time sequence signal is applied to the dynamic memristor, the device current is used as the output signal of the reservoir computing result, the device current is used for subsequent network inference identification based on the virtual node method in reservoir computing, the short-term memory feature of adjustable decay time constant is realized by applying a bias voltage during the conductance state decay of the dynamic memristor, and the ion drift and diffusion phenomenon in the interface resistance switching layer is utilized, so that the time scale of the reservoir can be modulated. In summary, the application solves the problems of non-uniformity of the material system of the reservoir layer and the readout layer, and single time scale and non-adjustable of the reservoir system, and the method proposed in the application can adaptively adjust the time scale of the reservoir, so that it can better adapt to different categories of input signals, thereby improving the reservoir computing performance and completing the dynamic information identification function.

[0081] The technical features of the above-described embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described, but it should be considered that any combination of the technical features is within the scope of the present disclosure. It should be noted that the "in an embodiment of the present application", "for example", "such as" and the like in the present disclosure are intended to illustrate the present disclosure, but are not used to limit the present disclosure.

[0082] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the patent application. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application.

Claims

1. An adaptive homogeneous reservoir computing system based on oxide memristors, characterized in that: include: The reservoir layer includes a plurality of volatile dynamic memristors with nonlinear short-term decay memory characteristics, which are used to receive timing signals and extract feature information; The dynamic memristor comprises a top electrode, a bottom electrode and a first stacked layer located in the middle; a fully connected readout layer, comprising a plurality of non-volatile memristors with multi-valued long-term memory characteristics, configured to receive the characteristic information output by the reservoir layer and output a recognition result, wherein the non-volatile memristor comprises a top electrode, a bottom electrode, and a second stacked layer located in between; A feedback loop, configured to reversely adjust the bias voltage applied during the decay period of the dynamic memristor in the reservoir layer according to the recognition result output by the fully connected readout layer, so as to match the decay time constant of the dynamic memristor with the time scale of the input timing signal; The first stack and the second stack are both formed by a first oxide layer and a second oxide layer stacked in contact, the first oxide layer is a titanium oxide layer, the material of the second oxide layer is a material with an electron affinity lower than that of titanium oxide and has non-volatile resistive switching properties, and the oxygen vacancy content of the titanium oxide layer in the first stack is less than the content threshold for generating conductive filaments, and the oxygen vacancy content of the titanium oxide layer in the second stack is greater than or equal to the content threshold for generating conductive filaments.

2. The oxide memristor-based adaptive homogeneous reservoir computing system according to claim 1, wherein: The fully connected readout layer has a plurality of output terminals, and the recognition result includes recognition of the content category of the time series signal and recognition of the decay speed of the dynamic memristor; The reverse adjustment of the bias voltage of the dynamic memristor in the reservoir layer according to the recognition result output by the fully connected readout layer includes: reverse adjustment of the bias voltage of the dynamic memristor in the reservoir layer according to the recognition result of the fully connected readout layer on the decay speed of the dynamic memristor.

3. The adaptive homogeneous reservoir computing system based on oxide memristor according to claim 1, characterized in that: The second oxide layer is made of aluminum oxide or hafnium oxide.

4. The adaptive homogeneous reservoir computing system based on oxide memristor according to claim 1, characterized in that: The dynamic memristor includes a substrate and a bottom electrode, a first stack and a top electrode stacked on the substrate from bottom to top, wherein the titanium oxide layer in the first stack is located on a side close to the top electrode; the non-volatile memristor includes a substrate and a bottom electrode, a second stack and a top electrode stacked on the substrate from bottom to top, wherein the titanium oxide layer in the second stack is located on a side close to the bottom electrode.

5. The oxide memristor-based adaptive homogeneous reservoir computing system according to claim 4, wherein: The bottom electrode is adhered to the substrate through a bottom electrode adhesion layer, and the material of the bottom electrode adhesion layer is titanium.

6. The oxide memristor-based adaptive homogeneous reservoir computing system according to any one of claims 1 to 5, characterized in that: The first stack and the second stack are both prepared by atomic layer deposition process, wherein The titanium oxide layer in the first stack is deposited at a temperature of 110° C. to 200° C. using tetrakis(dimethylamino)titanium as a titanium source precursor and H 2 O as an oxygen source; The titanium oxide layer in the second stack is deposited at a temperature of 250° C. to 350° C. using tetrakis(dimethylamino)titanium as a titanium source precursor and H 2 O as an oxygen source.

7. The oxide memristor-based adaptive homogeneous reservoir computing system according to any one of claims 1 to 5, characterized in that: The thickness of the same material layers in the first stack and the second stack is the same, wherein the thickness of the titanium oxide layer is 35 nm to 45 nm, and the thickness of the second oxide layer is 4 nm to 6 nm.

8. The oxide memristor-based adaptive homogeneous reservoir computing system according to any one of claims 1 to 5, characterized in that: The materials of the top electrode and the bottom electrode are selected from Al, Pt, Au, W, Ti or TiN.

9. The oxide memristor-based adaptive homogeneous reservoir computing system according to claim 1, wherein: It also includes an input layer for converting the target to be identified into a time series signal and inputting the signal into the reservoir layer.

10. The oxide memristor-based adaptive homogeneous reservoir computing system according to claim 1, wherein: The system further includes a control module configured to apply the bias voltage to the dynamic memristor only during a period in which a voltage pulse without a timing signal is applied to the dynamic memristor.

Images