A synaptic analog method and system for a resistive switching device

Abstract

This invention relates to the field of synaptic simulation technology and discloses a method and system for synaptic simulation of resistive switching devices. The method includes: acquiring a target synaptic learning curve and determining a set of sampling points; setting an error threshold, a uniform initial value for the time interval, and an adjustment step size; cyclically outputting a sinusoidal driving signal with constant amplitude and duration at a fixed frequency sequence; applying the signal and reading the conductance, constructing a synaptic simulation curve and calculating its root mean square error compared to the target curve; if the error exceeds the threshold, determining the time interval adjustment direction based on the difference in conductance increments, updating the interval, and iteratively executing until the error meets the requirements. This invention, by using the time interval as a single adjustment variable, combined with directional adjustment guided by incremental differences and an adaptive step size mechanism, achieves repeated fitting of the target synaptic learning curve under stable excitation conditions, thereby improving the accuracy and stability of synaptic simulation of resistive switching devices.

Description

Technical Field

[0001] This invention relates to the field of synaptic simulation technology, and more specifically, to a synaptic simulation method and system for resistive switching devices. Background Technology

[0002] Neuromorphic computing simulates the learning process of biological neural networks by implementing weight storage and updates at the hardware level. Artificial synaptic devices are used to characterize synaptic weights and their plasticity changes in response to stimuli using electrical quantities. Resistive switching devices (RSDs) can generate reversible or quasi-reversible conductance modulation under applied electrical excitation, and are therefore widely used in the hardware implementation of synaptic simulation and weight updates.

[0003] However, in engineering implementation, the conductance response of resistive switching devices typically exhibits significant nonlinearity, saturation characteristics, and history dependence. Furthermore, the conductance modulation process is susceptible to factors such as initial state, device differences, cyclic drift, and noise disturbances. Open-loop driving methods based on fixed excitation parameters often fail to ensure that the conductance evolution trajectory over time stably reflects the expected weight update pattern. This can easily lead to inconsistencies between the gain change of the synaptic simulation curve and the target synaptic learning curve, accumulated deviations in local curve segments, and significant differences in fitting results between different devices or batches. Ultimately, this results in a large fitting error for the target synaptic learning curve, affecting the accuracy and repeatability of synaptic simulation and limiting the application of resistive switching devices in high-precision neuromorphic systems. Summary of the Invention

[0004] In view of this, the present invention proposes a synaptic simulation method and system for resistive switching devices, aiming to solve the problem of large fitting error of the target synaptic learning curve during the synaptic simulation of resistive switching devices.

[0005] In one aspect, the present invention proposes a synaptic simulation method for resistive switching devices, comprising: Obtain the target synaptic learning curve and determine the set of sampling points, set the error threshold, the initial value of the unified time interval and the adjustment step size; Pre-set at least two sinusoidal drive signals of different frequencies and determine the preset frequency sequence; The sinusoidal drive signal is applied according to the initial value of the unified time interval and the preset frequency sequence, and the conductance of the resistive switching device is read to obtain the conductance value sequence. A synaptic simulation curve is constructed based on the conductivity value sequence, and the root mean square error between the synaptic simulation curve and the target synaptic learning curve is calculated based on the set of sampling points to obtain the error value. When the error value is less than or equal to the error threshold, the synaptic simulation curve is output; when the error value is greater than the error threshold, the conductance increment of adjacent sampling points is calculated based on the sampling point set, and an increment difference sequence is obtained based on the difference between the conductance increment of the synaptic simulation curve and the conductance increment of the target synaptic learning curve. The average increment difference is calculated based on the increment difference sequence, and the time interval adjustment direction is determined based on the sign of the average increment difference. The unified time interval is updated according to the adjustment direction and the adjustment step size, and an update error value is determined based on the updated unified time interval; The adjustment step size is updated and the time interval adjustment direction is corrected based on the comparison result between the error value and the updated error value. When the updated error value is greater than or equal to the error value, the time interval adjustment direction is reversed and the adjustment step size is halved until the updated error value is less than the error threshold.

[0006] Furthermore, when acquiring the target synaptic learning curve and determining the set of sampling points, setting the error threshold, the initial value of the unified time interval, and the adjustment step size, the process includes: Obtain the time coordinates and amplitude coordinates of the target synaptic learning curve, and determine the time series of the sampling point set based on the time coordinates; When the sampling density of the target synaptic learning curve is inconsistent in different time periods, the target synaptic learning curve is resampled to the sampling point set based on interpolation; The average time interval is calculated based on the time interval between adjacent sampling points in the set of sampling points, and the average time interval is used as the initial value of the unified time interval. The adjustment step size is half of the initial value of the unified time interval.

[0007] Furthermore, when presetting at least two sinusoidal drive signals of different frequencies and determining the preset frequency sequence, the process includes: According to the preset frequency sequence, the sinusoidal driving signals corresponding to different frequencies are arranged in ascending order of frequency to form a first sequence, and the first sequence is arranged in descending order of frequency to form a second sequence; A symmetrical frequency sequence is obtained by splicing the first sequence and the second sequence together, and the symmetrical frequency sequence is used as the preset frequency sequence. Based on the preset frequency sequence, the position of the conductivity value after each application is marked to obtain the conductivity index corresponding to the frequency position.

[0008] Further, when applying the sinusoidal drive signal according to the initial value of the unified time interval, and reading the conductance of the resistive switching device after each application to obtain the conductance value sequence, the process includes: The sinusoidal driving signal is triggered based on the zero-crossing point of the sinusoidal driving signal; The reading time is determined to be the immediate moment after the end of the sinusoidal driving signal based on the duration of the sinusoidal driving signal, and a reading signal is applied at the reading time to obtain the reading current; The conductivity value is calculated based on the read current and the read signal, and the average of the conductivity values ​​after the same application is obtained to obtain the conductivity value corresponding to the application. The conductivity values ​​corresponding to each application are written into the conductivity value sequence according to the application order.

[0009] Further, a synaptic simulation curve is constructed based on the conductivity value sequence, and the root mean square error between the synaptic simulation curve and the target synaptic learning curve is calculated based on the sampling point set. The error value is obtained by including: The conductivity value sequence is rearranged according to the order of the sampling point set to obtain the simulated conductivity sequence corresponding to the sampling point set; The simulated conductance sequence is normalized based on the maximum and minimum values ​​of the simulated conductance sequence in the same round of iteration to obtain the synaptic simulation curve; The error sequence is obtained by subtracting the synaptic simulation curve from the target synaptic learning curve point by point based on the set of sampling points, and the root mean square error is calculated as the error value based on the sum of squares of the error sequence and the number of sampling points.

[0010] Further, when determining the time interval adjustment direction based on the sign of the average incremental difference, the process includes: Based on the set of sampling points, the conductance increment of adjacent sampling points is calculated on the target synapse learning curve to obtain the target increment sequence, and based on the set of sampling points, the conductance increment of adjacent sampling points is calculated on the synapse simulation curve to obtain the simulation increment sequence; The incremental difference sequence is obtained by subtracting the simulated incremental sequence from the target incremental sequence point by point. When there is a zero increment term in the target increment sequence, the corresponding increment difference term is removed from the averaging operation, and the average increment difference is calculated for the remaining increment difference terms. The direction of the time interval adjustment is determined based on the sign of the average incremental difference.

[0011] Further, when updating the unified time interval according to the time interval adjustment direction and the adjustment step size, and determining the update error value based on the updated unified time interval, the process includes: When the time interval adjustment direction is to increase, the updated unified time interval is obtained by adding the unified time interval to the adjustment step size; When the time interval adjustment direction is decreasing, the updated unified time interval is obtained by subtracting the adjustment step size from the unified time interval; When the updated unified time interval is less than the duration of the sinusoidal driving signal, the updated unified time interval is corrected to the duration of the sinusoidal driving signal based on the duration of the sinusoidal driving signal. Based on the updated uniform time interval and according to the preset frequency sequence, the sinusoidal drive signal is applied and the conductance of the resistive switching device is read to obtain the updated conductance value sequence. An updated synaptic simulation curve is constructed based on the updated conductivity value sequence, and the root mean square error between the updated synaptic simulation curve and the target synaptic learning curve is calculated based on the set of sampling points to obtain the update error value.

[0012] Further, when updating the adjustment step size and correcting the time interval adjustment direction based on the comparison result between the error value and the updated error value, the process includes: When the update error value is less than the error value, the updated unified time interval is used as the current unified time interval, and the update error value is used as the current error value, while maintaining the time interval adjustment direction and the adjustment step size; When the update error value is greater than or equal to the error value, maintain the same time interval as before the update and maintain the error value before the update, reverse the adjustment direction of the time interval and halve the adjustment step size; The next round of unified time interval update will proceed based on the updated unified time interval, the time interval adjustment direction, and the adjustment step size.

[0013] Further, until the update error value is less than the error threshold, it includes: Based on the uniform time interval, the update error value, and the difference between the average increment recorded in any round of iteration, an iterative record sequence is obtained; Based on the iterative record sequence, a uniform time interval that makes the update error value satisfy the error threshold is determined as the target uniform time interval; The synaptic simulation curve is constructed and output based on the conductivity value sequence corresponding to the target uniform time interval, as the synaptic simulation result.

[0014] Compared with the prior art, the beneficial effects of this invention are as follows: By using the target synaptic learning curve as a reference and determining the set of sampling points, the synaptic simulation curve and the target synaptic learning curve are compared point-to-point under a unified sampling benchmark, thus providing a consistent quantitative basis for error evaluation; by using at least two different frequency sinusoidal driving signals and cyclically outputting according to a preset frequency sequence, while keeping the frequency sequence unchanged and the amplitude and duration of the sinusoidal driving signal unchanged during the iteration process, the driving conditions are kept stable in both the frequency and energy input dimensions, thereby concentrating the main variable factors in the fitting process into a unified time interval and reducing the fitting uncertainty caused by the simultaneous change of multiple parameters; by reading the conductance of the resistive switching device to obtain the conductance value sequence and constructing the synaptic simulation curve accordingly, and then using the root mean square error to quantify the overall deviation between the synaptic simulation curve and the target synaptic learning curve, the goodness or badness of the fit can be judged by a single error index and used for convergence control; by using the error... When the error threshold is exceeded, the difference in conductance increment between adjacent sampling points is calculated to form an increment difference sequence and calculate the average increment difference. Then, the direction of time interval adjustment is determined based on the sign of the average increment difference, so that the time interval update has a directional basis and avoids inefficient iteration caused by trial updates alone. By adjusting the direction of the time interval according to the time interval and combining it with the adjustment step size to update the unified time interval, the process of applying, reading, constructing and calculating the error value is repeated to obtain the updated error value, so that the update of the unified time interval forms a closed-loop feedback iteration process, thereby gradually reducing the fitting error. By comparing the updated error value with the error value and adjusting the time interval in the opposite direction and halving the adjustment step size when the updated error value does not improve, the iteration process can automatically reduce the update amplitude and switch the search direction when overshoot or directional deviation occurs, improving convergence stability and promoting the convergence of the error to within the error threshold, thereby improving the fitting accuracy and repeatability of the target synapse learning curve in the synapse simulation of resistive switching devices.

[0015] On the other hand, this application also provides a synaptic simulation system for resistive switching devices, used to implement the above-mentioned synaptic simulation method for resistive switching devices, comprising: The parameter initialization module is configured to acquire the target synaptic learning curve and determine the set of sampling points, set the error threshold, the initial value of the uniform time interval, and the adjustment step size. The signal configuration module is configured to use at least two different frequency sinusoidal drive signals and to determine a preset frequency sequence. The conductivity data reading module is configured to apply the sinusoidal drive signal according to the initial value of the uniform time interval and the preset frequency sequence, and read the conductivity of the resistive switching device to obtain a conductivity value sequence. The curve evaluation module is configured to construct a synaptic simulation curve based on the conductivity value sequence, and calculate the root mean square error between the synaptic simulation curve and the target synaptic learning curve based on the set of sampling points to obtain the error value; The decision-making module is configured to output the synaptic simulation curve when the error value is less than or equal to the error threshold; when the error value is greater than the error threshold, calculate the conductance increment of adjacent sampling points based on the sampling point set, obtain an increment difference sequence based on the difference between the conductance increment of the synaptic simulation curve and the conductance increment of the target synaptic learning curve, calculate the average increment difference based on the increment difference sequence, and determine the time interval adjustment direction based on the sign of the average increment difference. The iterative execution module is configured to adjust the direction and the adjustment step size according to the time interval, update the unified time interval, and determine the update error value based on the updated unified time interval. The convergence control module is configured to update the adjustment step size and correct the time interval adjustment direction based on the comparison result between the error value and the updated error value. When the updated error value is greater than or equal to the error value, the time interval adjustment direction is reversed and the adjustment step size is halved until the updated error value is less than the error threshold.

[0016] It is understandable that the above-mentioned synaptic simulation method and system for resistive switching devices have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a synapse simulation method for resistive switching devices provided in an embodiment of the present invention; Figure 2 This is a functional block diagram of a synaptic simulation system for resistive switching devices provided in an embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] See Figure 1As shown, this application proposes a synaptic simulation method for resistive switching devices, comprising: S1: Obtain the target synaptic learning curve and determine the set of sampling points, set the error threshold, uniform initial value of time interval and adjustment step size; S2: Preset at least two different frequency sinusoidal drive signals and determine the preset frequency sequence; S3: Apply a sinusoidal drive signal according to the initial value of the uniform time interval and the preset frequency sequence, and read the conductance of the resistive switching device to obtain the conductance value sequence; S4: Construct a synaptic simulation curve based on the conductivity value sequence, and calculate the root mean square error between the synaptic simulation curve and the target synaptic learning curve based on the set of sampling points to obtain the error value; S5: When the error value is less than or equal to the error threshold, output the synaptic simulation curve; when the error value is greater than the error threshold, calculate the conductance increment of adjacent sampling points based on the sampling point set, and obtain the increment difference sequence based on the difference between the conductance increment of the synaptic simulation curve and the conductance increment of the target synaptic learning curve. Calculate the average increment difference based on the increment difference sequence, and determine the time interval adjustment direction based on the sign of the average increment difference. S6: Adjust the direction and step size according to the time interval, update the unified time interval, and determine the update error value based on the updated unified time interval; S7: Update the adjustment step size and correct the time interval adjustment direction based on the comparison result between the error value and the updated error value. When the updated error value is greater than or equal to the error value, reverse the time interval adjustment direction and halve the adjustment step size until the updated error value is less than the error threshold.

[0020] Specifically, the target synapse learning curve is used to characterize the reference trajectory of the desired synaptic weights evolving with stimulation or time. It can be understood as a curve with time as the horizontal axis and conductance or equivalent weight amplitude as the vertical axis. The sampling point set is a set of discrete aligned positions selected from the target synapse learning curve. It is used to compare the conductance changes measured by the resistive switching device point by point on the same benchmark, thereby ensuring the comparability of error calculations. The error threshold is used to define the acceptable range of fitting accuracy. When the calculated error does not exceed the threshold, the synaptic simulation curve is considered to have met the target approximation requirements. The initial value of the uniform time interval refers to the initial interval time used between two adjacent sinusoidal driving signals. This interval is updated as an adjustable variable during the iteration process. The adjustment step size is the step size used to update the uniform time interval. It represents the amount of change in the uniform time interval each time it is updated. It is used to control the magnitude of the uniform time interval update and affects the stability of the iteration convergence. A sinusoidal drive signal is an electrical excitation signal used to drive a resistive switching device to generate conductance modulation. It contains at least two different frequencies so that the device's response under different frequency conditions is included in the same simulation process. The conductance modulation of the resistive switching device is frequency-dependent. Different frequencies of AC excitation correspond to different action time scales, which may trigger different dominant conductance response processes. For example, under higher frequency conditions, the device response is more likely to reflect interface state charging and discharging, electron capture and release, or equivalent capacitance effect. Under lower frequency conditions, the device response is more likely to reflect slow processes such as ion migration, conductive filament growth, or interface oxygen vacancy reconstruction. By constructing a multi-timescale excitation based on at least two sinusoidal driving signals of different frequencies, the conductivity value sequence can simultaneously sample the dynamic response of the device in different frequency domains within the same iterative framework. This allows the constructed synaptic simulation curve to more comprehensively characterize the plasticity evolution of the device, reducing the risk that using only a single frequency will only effectively fit a portion of the target synaptic learning curve but fail to fit other intervals, and improving the adaptability to different target synaptic learning curve shapes. The preset frequency sequence is an arrangement of output sequences of different frequencies, used to constrain the application order of the sinusoidal driving signals and ensure the reproducibility of the driving sequence in each iteration. In some embodiments, the preset frequency sequence remains unchanged during the iteration process, so that the conductivity response is mainly affected by the update of a uniform time interval during the iteration process, thereby facilitating the comparison of the changing trends of error values ​​in each round. In some embodiments, the amplitude and duration of the sinusoidal driving signal remain unchanged, used to reduce the uncertainty introduced by the change in driving intensity, so that the error change more centrally reflects the update effect of the uniform time interval.Based on a uniform initial time interval, sinusoidal drive signals are applied to the resistive switching device according to a preset frequency sequence. After application, the conductance of the resistive switching device is read, resulting in a conductance value sequence arranged according to the application order. This sequence reflects the electrodynamic behavior of the resistive switching device under the aforementioned drive conditions and time intervals. A synaptic simulation curve is constructed based on this conductance value sequence. This simulation curve can be understood as a simulated weighted trajectory formed by organizing the conductance values ​​according to the correspondence of the sampling point set, used for alignment and comparison with the target synaptic learning curve. Subsequently, the root mean square error (RMSE) between the synaptic simulation curve and the target synaptic learning curve is calculated based on the sampling point set. The RMSE measures the overall deviation of the two curves at the sampling points. Its square operation emphasizes the influence of larger deviations, and the square root ensures that the error dimension is consistent with the original amplitude, thus obtaining the error value used to judge the fitting accuracy. When the error value is less than or equal to the error threshold, the synaptic simulation curve is directly output, indicating that the driving and device response under the current time interval have made the simulation results meet the accuracy requirements. When the error value is greater than the error threshold, the conductance increment is further introduced as a direction criterion. That is, the conductance change at adjacent sampling points is calculated based on the sampling point set, and the conductance increment of the synaptic simulation curve and the conductance increment of the target synaptic learning curve are obtained respectively. The difference between the two forms an increment difference sequence, which is used to characterize whether the simulation curve is faster or slower than the target curve in the local change trend. The average increment difference is then calculated from the increment difference sequence, and its sign is used to determine the direction of time interval adjustment. That is, the sign of the average increment difference determines whether the uniform time interval should be increased or decreased in the next round, so that the local increment trend of subsequent electrodynamics is closer to the target curve.The unified time interval is updated by adjusting the direction and step size according to the determined time interval. Based on the updated unified time interval, a sinusoidal drive signal is applied according to a preset frequency sequence, and the conductance of the resistive switching device is read to obtain an updated conductance value sequence. Based on the updated conductance value sequence, an updated synapse simulation curve is constructed, and the root mean square error is calculated to obtain the updated error value, thus forming a complete closed-loop iteration. Then, the adjustment step size and time interval adjustment direction of the unified time interval are corrected according to the comparison between the error value and the updated error value. When the updated error value is greater than or equal to the error value, it indicates that this round of update has not brought about error improvement. In this case, the time interval adjustment direction is reversed, and the adjustment step size is halved. This refers to halving the step size used to update the uniform time interval to reduce the search amplitude and suppress oscillations or overshoot until the update error value is less than the error threshold. This achieves iterative convergence fitting with the uniform time interval as the control variable and outputs a synaptic simulation curve that meets the accuracy requirements. To avoid the situation where the iteration cannot terminate when the noise disturbance is strong or the error reduction is limited, some implementations can set a maximum number of iterations as a protective termination condition. When the number of iterations reaches the maximum number of iterations, the iteration stops, and the uniform time interval corresponding to the minimum update error value is selected from the iteration record sequence as the target uniform time interval to output the synaptic simulation curve, so as to ensure the controllability and executability of the method in the engineering system.

[0021] In some embodiments of this application, the process of acquiring the target synaptic learning curve and determining the set of sampling points, setting the error threshold, the initial value of the uniform time interval, and the adjustment step size includes: Obtain the time coordinates and amplitude coordinates of the target synaptic learning curve, and determine the time series of the sampling point set based on the time coordinates; When the sampling density of the target synaptic learning curve is inconsistent in different time periods, the target synaptic learning curve is resampled to the set of sampling points based on interpolation; The average time interval is calculated based on the time interval between adjacent sampling points in the sampling point set, and the average time interval is used as the initial value of the unified time interval. The adjustment step size is half of the initial value of the unified time interval.

[0022] Specifically, the target synaptic learning curve can be understood as a reference curve used to characterize the change of the expected synaptic weight over time. Its time coordinate is used to describe the position distribution of the curve on the time axis, and the amplitude coordinate is used to describe the magnitude of the weight amplitude or conductance amplitude at the corresponding time position. When obtaining the target synaptic learning curve, the time coordinate sequence and amplitude coordinate sequence are extracted first, and the time sequence of the sampling point set is determined according to the time coordinate sequence, so that the sampling point set covers the effective change range of the target synaptic learning curve, thereby providing an alignment benchmark for subsequent error calculation. Since the target synaptic learning curve is derived from experimental data acquisition or simulation output, the sampling density may vary across different time periods, resulting in denser sampling points in some intervals and sparser ones in others. To avoid biased error calculations due to inconsistent sampling density, the target synaptic learning curve is resampled based on interpolation. This ensures that the target synaptic learning curve has a corresponding amplitude at each time position in the sampling point set, thus obtaining a target amplitude sequence that corresponds one-to-one with the sampling point set. Piecewise linear interpolation or spline interpolation can be used. Its function is to estimate the amplitude at the time position of the sampling point set based on the time-amplitude relationship between adjacent known sampling points, making the resampled target synaptic learning curve continuous on the time axis and suitable for point-by-point comparison. Subsequently, the time interval between adjacent sampling points is calculated based on the sampling point set. This involves statistically analyzing the difference between two adjacent points in the time series and averaging the time intervals between adjacent points. This average time interval is used as the initial value for a unified time interval, ensuring that the rhythm of subsequent applied driving signals matches the time scale of the target synaptic learning curve. This avoids an excessively large initial interval leading to an overly sparse response or an excessively small initial interval leading to an overly dense response. The adjustment step size is set to half of the initial value of the unified time interval, so that the adjustment step size and the initial value of the time interval are on the same order of magnitude. This makes it easier to explore the direction of error reduction with a more obvious step size in the early stage of iteration. At the same time, the search granularity is gradually refined through the step size halving mechanism in the later stage, so as to balance the convergence speed and convergence stability. In this process, an error threshold is set to limit the acceptable range of fitting accuracy. When the error value obtained by subsequent calculation does not exceed the error threshold, it can be determined that the synaptic simulation curve meets the target approximation requirements and the result is output.

[0023] In some embodiments of this application, when at least two different frequency sinusoidal drive signals are preset and the preset frequency sequence is determined, the following steps are included: According to the preset frequency sequence, the sinusoidal driving signals corresponding to different frequencies are arranged in ascending order of frequency to form a first sequence, and then arranged in descending order of frequency according to the first sequence to form a second sequence; A symmetrical frequency sequence is obtained by splicing the first and second sequences together, and the symmetrical frequency sequence is used as the preset frequency sequence. The conductivity value after each application is marked with a preset frequency sequence to obtain the conductivity index corresponding to the frequency position.

[0024] Specifically, a sinusoidal driving signal is used to apply periodic electrical excitation to the resistive switching device to trigger changes in its conductance state. Different frequencies correspond to different excitation time scales, thereby enabling the resistive switching device to produce differentiated conductance responses under dynamic conditions of varying speeds. The purpose of presetting at least two different frequencies is to enable the response characteristics under multi-frequency conditions during the synapse simulation process, avoiding the situation where the electrical conductance law obtained at only a single frequency is effective for fitting a local segment of the target synapse learning curve but insufficient for fitting other segments. A preset frequency sequence is used to specify the order in which sinusoidal drive signals of various frequencies are applied. To reduce the order bias caused by temperature drift, state drift, or cumulative effects, the sinusoidal drive signals corresponding to different frequencies are first arranged in ascending order to form a first sequence, and then the first sequence is arranged in descending order to form a second sequence. The first and second sequences are concatenated end-to-end to obtain a symmetrical frequency sequence, which is then used as the preset frequency sequence. Resistive switching devices typically exhibit historical dependence and excitation sequence effects. The cumulative effect between adjacent drives may cause a systematic drift in the device's base state, resulting in the conductance response corresponding to the later applied frequency being superimposed on the already drifted base state. This makes the response at the same frequency related to its position in the preset frequency sequence. The symmetrical frequency sequence ensures that each frequency appears once in both the ascending and descending segments, with the corresponding historical conditions being opposite. By applying the signals cyclically and using position markers to form a repeatable index structure, the responses of each frequency under different historical conditions are repeatedly sampled and comprehensively reflected, thereby weakening the systematic order bias caused by unidirectional frequency arrangement and improving the comparability of the conductance value sequence and the synaptic simulation curve during the iteration process. When applying a sinusoidal driving signal based on a preset frequency sequence, each application can be considered as a frequency position in the corresponding sequence, where the frequency position indicates which term the application belongs to within the preset frequency sequence. Marking the position of the conductance value after each application based on the preset frequency sequence means that after reading the conductance value, a mapping relationship is established between the conductance value and its corresponding frequency position, and index information is recorded. This ensures that the conductance value sequence contains structured labels consistent with the frequency positions. In this way, when constructing synaptic simulation curves subsequently, the conductance values ​​can be rearranged or aligned according to the sampling point set requirements based on the index, avoiding conductance value mismatch due to frequency overlap. Furthermore, in some embodiments, the preset frequency sequence remains unchanged during iteration, making the conductance value sequences of each iteration comparable. This allows the calculation of the root mean square error and the difference in conductance increment to stably reflect the changes brought about by the uniform time interval update.

[0025] In some embodiments of this application, when applying a sinusoidal drive signal based on a uniform initial time interval and reading the conductance of the resistive switching device after each application to obtain a conductance value sequence, the process includes: The sinusoidal drive signal is triggered based on the zero-crossing point of the sinusoidal drive signal; The reading time is determined based on the duration of the sinusoidal drive signal, which is the immediate moment after the end of the sinusoidal drive signal. A reading signal is then applied at the reading time to obtain the reading current. The conductivity value is calculated based on the read current and read signal, and the average of multiple conductivity values ​​after the same application is obtained to obtain the conductivity value corresponding to that application. Write the corresponding conductance values ​​for each application into the conductance value sequence according to the application order.

[0026] Specifically, applying a sinusoidal drive signal refers to sequentially loading sinusoidal drive signals of different frequencies onto the terminals of the resistive switching device under a unified time interval control, according to the output order determined by a preset frequency sequence, so that the resistive switching device generates an evolution of its conductance state under periodic electrical excitation. To ensure the comparability of applications of different cycles and frequencies, this embodiment triggers the sinusoidal drive signal based on the zero-crossing point. The zero-crossing point can be understood as the moment when the instantaneous voltage of the sinusoidal wave is equal to zero and the phase switches from the negative half-cycle to the positive half-cycle or from the positive half-cycle to the negative half-cycle. Using this moment as the triggering starting point can ensure that the initial phase of each application remains consistent, avoiding the difference in equivalent energy input caused by different starting phases, thereby reducing the random influence on the conductance modulation result. The conductance of the resistive switching device is read to obtain the device state after the application of the driving signal. The read time is determined to be the immediate moment after the end of the sinusoidal driving signal. The immediate moment can be understood as the immediate reading performed after the sinusoidal driving signal stops without introducing additional waiting, so as to reduce the state drift of the device during natural relaxation, thermal diffusion or spontaneous back-off, and make the reading result more representative of the conductance state at the moment the drive ends. In engineering implementation, the immediate moment can be determined by a fixed delay between the driving stop edge and the application of the read signal. This fixed delay is jointly determined by the switching delay, the sampling circuit setup time and the synchronization timing of the measurement system, and remains constant throughout the iteration process to avoid introducing new sources of error by changes in the read timing. The principle for determining the reading time is to minimize the time required for the reading circuit to reach stable sampling conditions without significantly altering the conductance state of the resistive switching device, and to ensure that this delay is reproducible in each round of application and reading, thereby guaranteeing consistent measurement conditions for the conductance value sequence and valid comparison of the root mean square error. A reading signal is applied at this reading time to obtain the reading current. The reading signal is typically a test signal with a small amplitude and short duration, used to obtain the current response without significantly altering the conductance state of the resistive switching device. The conductance value is calculated based on the reading current and the reading signal. For example, the conductance value can be obtained from the ratio of the reading current to the reading signal using the Ohm's law, thus converting the measured current into a conductance value characterizing the conductance of the resistive switching device. Considering that measurement noise, transient disturbances, or sampling jitter may cause fluctuations in a single reading, this embodiment averages multiple conductance values ​​after the same application to obtain the conductance value corresponding to that application. Multiple readings can be understood as performing several reading operations consecutively near the same reading time and obtaining conductance values ​​separately. Averaging is used to suppress the influence of random noise on the results and improve the stability of the conductance estimation.Finally, the conductivity values ​​corresponding to each application are written into the conductivity value sequence according to the application order. This means that the conductivity values ​​obtained from each drive and read are stored in the same sequence structure according to the chronological order. This makes the conductivity value sequence fully reflect the conductivity transformation process after the sinusoidal drive signal is applied under the constraint of a unified time interval and the frequency sequence is cyclically output. This provides basic data for subsequent construction of synaptic simulation curves, calculation of root mean square error, and calculation of conductivity increments at adjacent sampling points based on this sequence.

[0027] In some embodiments of this application, a synaptic simulation curve is constructed based on a conductivity value sequence, and the root mean square error between the synaptic simulation curve and the target synaptic learning curve is calculated based on a set of sampling points. The error value is obtained by: The conductivity value sequence is rearranged according to the order of the sampling point set to obtain the simulated conductivity sequence corresponding to the sampling point set; The simulated conductance sequence is normalized based on the maximum and minimum values ​​of the simulated conductance sequence in the same round of iteration to obtain the synaptic simulation curve; The error sequence is obtained by subtracting the synaptic simulation curve from the target synaptic learning curve point by point based on the set of sampling points, and the root mean square error is calculated as the error value based on the sum of squares of the error sequence and the number of sampling points.

[0028] Specifically, the conductivity value sequence is the measurement result recorded according to the order in which sinusoidal driving signals are applied, while the sampling point set is a pre-determined set of aligned positions on the time axis of the target synapse learning curve. The two may not naturally correspond one-to-one in terms of organization. Therefore, the conductivity value sequence is first rearranged according to the order of the sampling point set so that the arrangement order of the conductivity values ​​in the sequence is consistent with the time sequence of the sampling point set, thereby obtaining the analog conductivity sequence corresponding to the sampling point set. This rearrangement can be understood as mapping the original conductivity value sequence to the index space of the sampling point set according to the sampling point index corresponding to each application, so that each sampling point corresponds to an analog conductivity value, ensuring that subsequent error calculation is a point-by-point comparison rather than a misaligned comparison. Subsequently, the simulated conductance sequence within the same iteration is normalized. Normalization is performed by scaling each conductance value using the maximum and minimum values ​​in the simulated conductance sequence of that iteration, so that the transformed values ​​fall within a consistent amplitude range. This reduces the influence of the absolute conductance level difference of the resistive switching device on the curve shape comparison and highlights the relative change trend of the conductance trajectory. The normalized simulated conductance sequence constitutes the synaptic simulation curve. The synaptic simulation curve can be understood as a curve with the time series corresponding to the sampling point set as the horizontal axis and the normalized simulated conductance as the vertical axis, used to characterize the simulated learning trajectory of the resistive switching device under the current uniform time interval. Subsequently, based on the set of sampling points, the difference between the synaptic simulation curve and the target synaptic learning curve is calculated point by point to obtain an error sequence. Point-by-point difference means subtracting the amplitude of the target synaptic learning curve from the amplitude of the synaptic simulation curve at each sampling point to obtain the deviation at each sampling point. The error sequence reflects the distribution of deviations at each sampling point. On this basis, the squares of each deviation in the error sequence are summed to obtain a sum of squares. The squaring operation is used to eliminate the mutual cancellation of positive and negative deviations and to assign higher weights to larger deviations. Then, the sum of squares is averaged and squared using the number of sampling points to obtain the root mean square error. The dimension of the root mean square error is consistent with the curve amplitude, which can be used as a quantitative indicator of the overall fit. This root mean square error is used as the error value for subsequent comparison with the error threshold to determine whether to output the synaptic simulation curve or to enter the time interval adjustment process based on conductance increment. Normalization is performed based on the maximum and minimum values ​​of the same iteration to map the simulated conductance sequence to a uniform amplitude scale, thereby highlighting the relative shape change of conductance and weakening the influence of absolute conductance level differences on error calculation. To ensure fairness in cross-iteration comparisons, normalization is preferably used only when the device has not undergone extreme resets or abnormal saturation that could cause abrupt changes in the response range. This ensures that the amplitude scaling in each iteration primarily reflects the dynamic differences caused by changes in the uniform time interval rather than range changes caused by abnormal state switching. In other instances, the normalization benchmark can also be determined based on the amplitude range of the target synaptic learning curve, keeping the normalization scale consistent with the target curve to further reduce the impact of normalization scaling on the error sequence.

[0029] In some embodiments of this application, determining the time interval adjustment direction based on the sign of the average incremental difference includes: The target increment sequence is obtained by calculating the conductance increment of adjacent sampling points based on the target synapse learning curve using the sampling point set, and the simulated increment sequence is obtained by calculating the conductance increment of adjacent sampling points based on the synapse simulation curve using the sampling point set. The incremental difference sequence is obtained by calculating the difference between the simulated incremental sequence and the target incremental sequence point by point. When there is a zero increment term in the target increment sequence, the corresponding increment difference term is removed from the averaging operation, and the average increment difference is calculated for the remaining increment difference terms. The direction of time interval adjustment is determined based on the sign of the average increment difference.

[0030] Specifically, the time interval adjustment direction is used to indicate whether the uniform time interval should be increased or decreased in the next iteration. The determination is not based solely on the overall error magnitude, but rather on the directional information provided by the trend of the curve's change in a local segment. Based on the set of sampling points, the conductance increments of adjacent sampling points are calculated for both the target synapse learning curve and the synapse simulation curve. These adjacent sampling point conductance increments can be understood as the difference in amplitude between two adjacent sampling points on the same curve, characterizing the rise or fall of the curve in that local interval. The target increment sequence calculated for the target synapse learning curve reflects the expected change of the target curve in each adjacent sampling interval, while the simulated increment sequence calculated for the synapse simulation curve reflects the actual change of the device response curve in the corresponding interval under the current driving conditions. Subsequently, the difference between the simulated increment sequence and the target increment sequence is calculated point by point to obtain the increment difference sequence. Point by point, the difference is obtained by subtracting the target increment from the simulated increment in each adjacent sampling interval. The difference is used to characterize whether the change of the simulated curve in that interval is faster or slower than that of the target curve: when the difference is positive, it means that the change of the simulated curve in that interval is greater than the change of the target curve, which is reflected in the local segment gain being larger or the update being faster; when the difference is negative, it means that the change of the simulated curve in that interval is less than the change of the target curve, which is reflected in the local segment gain being smaller or the update being slower. Because the target synaptic learning curve may have flat or stable segments in certain intervals, resulting in zero-increment terms in the target increment sequence, the target change in the corresponding interval is zero. In this case, the positive or negative value of the increment difference may reflect noise or small fluctuations rather than trend deviations. Therefore, the corresponding increment difference terms are removed from the averaging calculation to avoid interference from flat segments in direction determination. When the target curve has a plateau segment, making the number of increment difference terms available for averaging after removing zero-increment terms lower than expected, the average increment difference is more sensitive to fluctuations in a single interval. In this case, the direction of the previous time interval adjustment can be kept unchanged, and the direction reversal and step size reduction can be triggered mainly by comparing the updated error value with the error value, thus maintaining closed-loop convergence. Alternatively, the increment difference terms can be screened before calculation, retaining only intervals where the absolute value of the target increment is non-zero and the change amplitude is more significant to participate in the calculation, thereby enhancing the representativeness of the direction determination for the effective change segment and improving reliability. The arithmetic mean of the remaining increment difference terms is used to obtain the average increment difference, so that the local trend deviations of different intervals are statistically summarized into an overall direction indicator.Finally, the direction of time interval adjustment is determined based on the sign of the average increment difference. The sign of the average increment difference can be understood as a comprehensive representation of the global trend deviation: a positive average increment difference indicates that the simulation curve changes faster than the target curve overall. To reduce the deviation, the device response update rhythm needs to be slowed down by adjusting the unified time interval. Therefore, the direction of time interval adjustment is determined to be increasing. A negative average increment difference indicates that the simulation curve changes slower than the target curve overall. To increase the change amplitude, the device response update rhythm needs to be accelerated by adjusting the unified time interval. Therefore, the direction of time interval adjustment is determined to be decreasing. This provides a clear adjustment direction for subsequent iterative updates of the unified time interval.

[0031] In some embodiments of this application, when updating the unified time interval based on the time interval adjustment direction and adjustment step size, and determining the update error value based on the updated unified time interval, the process includes: When the time interval adjustment direction is increasing, the updated unified time interval is obtained by adding the unified time interval to the adjustment step size; When the time interval adjustment direction is decreasing, the updated uniform time interval is obtained by subtracting the adjustment step size from the uniform time interval. When the updated uniform time interval is less than the duration of the sinusoidal drive signal, the updated uniform time interval is corrected to the duration of the sinusoidal drive signal based on the duration of the sinusoidal drive signal. Based on the updated uniform time interval, a sinusoidal drive signal is applied according to a preset frequency sequence, and the conductance of the resistive switching device is read to obtain the updated conductance value sequence. An updated synaptic simulation curve is constructed based on the updated conductivity value sequence, and the root mean square error between the updated synaptic simulation curve and the target synaptic learning curve is calculated based on the set of sampling points to obtain the update error value.

[0032] Specifically, the unified time interval is used to define the time interval between two consecutive sinusoidal drive signals, and its update process constitutes the control quantity adjustment link in the closed-loop iteration. The time interval adjustment direction is used to indicate whether the unified time interval should be increased or decreased in the next iteration. The adjustment step size is used to limit the change magnitude of the unified time interval in each update, thereby gradually approaching the unified time interval that meets the error threshold while ensuring convergence. In specific updates, when the time interval adjustment direction is to increase, the current unified time interval is added to the adjustment step size to obtain the updated unified time interval, so as to extend the waiting time between adjacent applications and change the change rhythm of subsequent electrodynamics; when the time interval adjustment direction is to decrease, the current unified time interval is subtracted from the adjustment step size to obtain the updated unified time interval, so as to shorten the waiting time between adjacent applications and make the driving action more compact. To ensure the physical feasibility of the updated uniform time interval and avoid overlapping drive timings, when the updated uniform time interval is less than the duration of the sinusoidal drive signal, it is corrected to match the duration of the sinusoidal drive signal. This ensures that the next application begins only after the previous sinusoidal drive signal is completed and that the reading operation can be performed after each drive. Therefore, the fastest simulating update rhythm, under the premise of unchanged duration, is limited by this lower bound constraint. After obtaining the updated uniform time interval, a sinusoidal drive signal is applied according to a preset frequency sequence based on the updated uniform time interval, and the conductance of the resistive switching device is read to obtain an updated conductance value sequence. An updated synaptic simulation curve is constructed based on the updated conductance value sequence, and the root mean square error between the updated synaptic simulation curve and the target synaptic learning curve is calculated based on the set of sampling points to obtain the update error value. The update error value is used to characterize the fitting effect after this round of uniform time interval update and serves as an input for subsequent comparison and convergence control to determine whether it is necessary to correct the time interval adjustment direction and adjust the adjustment step size used for updating the uniform time interval.

[0033] In some embodiments of this application, when updating the adjustment step size and correcting the time interval adjustment direction based on the comparison result of the error value and the updated error value, the following steps are included: When the update error value is less than the error value, the updated unified time interval is used as the current unified time interval, and the update error value is used as the current error value, while maintaining the time interval adjustment direction and the adjustment step size. When the update error value is greater than or equal to the error value, maintain the same time interval and error value as before the update, reverse the direction of the time interval adjustment and halve the adjustment step size. The next round of unified time interval updates will begin based on the updated unified time interval, time interval adjustment direction, and adjustment step size.

[0034] Specifically, the error value characterizes the overall deviation of the synaptic simulation curve from the target synaptic learning curve under the current uniform time interval condition, while the updated error value characterizes the change in this deviation under the updated uniform time interval condition. Therefore, the comparison between the two is used to determine whether the current time interval update is effective, and to adaptively adjust the subsequent search strategy accordingly. Specifically, when the updated error value is less than the error value, it indicates that adjusting the direction according to the predetermined time interval and updating the uniform time interval with the current adjustment step size can reduce the root mean square error, indicating that the adjustment direction is consistent with the target error reduction direction. In this case, using the updated uniform time interval as the current uniform time interval means writing the updated uniform time interval calculated in this round and replacing the uniform time interval of the previous round, so that the new interval parameter is used when applying the sinusoidal driving signal in the future. At the same time, using the updated error value as the current error value means updating the current error value to the updated error value of this round as the comparison benchmark for the next round. In addition, maintaining the time interval adjustment direction and maintaining the adjustment step size means continuing to advance in the same direction with the same amplitude under the trend of continuous error reduction, thereby accelerating the convergence to the solution that meets the error threshold. Conversely, when the update error value is greater than or equal to the error value, it indicates that the current time interval update has not brought about error improvement. This could result in either no change in error or an increase in error, both indicating that the update direction or magnitude is unsuitable for continued use. In this case, maintaining the previous uniform time interval and error value signifies rejecting the update result and reverting to the previously confirmed parameter state, ensuring that the iteration process does not deviate from the existing optimal region due to an unfavorable update. Subsequently, reversing the time interval adjustment direction means switching from increasing the uniform time interval to decreasing it, or vice versa, so that the next round searches for an error reduction path from the opposite direction. Simultaneously, halving the adjustment step size means reducing the magnitude of each adjustment based on the reversed direction, reducing overshoot risk and improving search resolution, enabling the system to find uniform time intervals that reduce error at a finer granularity. Finally, based on the updated unified time interval, time interval adjustment direction, and adjustment step size, the next round of unified time interval update is entered. This means that the currently confirmed unified time interval, the corrected direction, and the updated step size are used as input parameters for the next round of iteration. The application, reading, construction, and error calculation process are repeatedly executed, thereby forming a closed-loop convergence mechanism with error comparison as feedback. This allows the unified time interval to gradually approach the range of values ​​that make the updated error value meet the error threshold in continuous iteration.

[0035] In some embodiments of this application, the process of waiting until the update error value is less than an error threshold includes: Based on the difference between the uniform time interval, update error value and average increment of any round of iteration, the iterative record sequence is obtained; The target uniform time interval is determined based on the iterative record sequence to make the update error value meet the error threshold. The synaptic simulation curve is constructed and output based on the conductivity value sequence corresponding to the target uniform time interval, which serves as the synaptic simulation result.

[0036] Specifically, the iteration process terminates when the update error value is less than the error threshold, indicating that the fitting accuracy requirement has been met. The update error value reflects the root mean square error of the synaptic simulation curve relative to the target synaptic learning curve under the current uniform time interval. The error threshold limits the acceptable range of fitting deviation. When the update error value enters this range, the synaptic simulation curve is considered to have reached the expected approximation accuracy. To avoid infinite loops in cases of abnormal device conditions, strong noise disturbances, or when the error cannot continue to decrease, a maximum number of iterations can be set as a protective termination condition. When the maximum number of iterations is reached, the iteration stops, and the uniform time interval corresponding to the minimum update error value is selected from the iteration record sequence as the target uniform time interval for outputting the synaptic simulation curve. During the iteration process, recording the unified time interval, update error value, and average increment difference for each iteration means that after each iteration completes the application, reading, construction, and error calculation, the unified time interval parameter used in this iteration, the update error value obtained in this iteration, and the average increment difference used for direction determination are written into the same record structure to form an iteration record sequence. The unified time interval is used to identify the control value of the driving timing of this iteration, the update error value is used to identify the overall fitting effect corresponding to the value, and the average increment difference is used to identify the deviation direction of the local change trend of this iteration relative to the target curve. The combination of these three records enables subsequent tracing and verification of the iteration path and provides a basis for the final parameter selection. Determining the target unified time interval based on the iterative record sequence, where the update error value meets the error threshold, involves selecting records from the iterative record sequence whose update error value is less than the error threshold, reading the corresponding unified time interval from these records, and using this unified time interval as the final target unified time interval. If multiple records meet the condition, the first record meeting the condition can be selected according to the iteration order, allowing the iteration to end and output the result when the accuracy requirement is first met, thus balancing convergence speed and implementation complexity. Finally, a synaptic simulation curve is constructed and output based on the conductance value sequence corresponding to the target unified time interval, serving as the synaptic simulation result. This involves calling the conductance value sequence corresponding to the round of the target unified time interval, aligning and rearranging it according to the sampling point set, and performing the same normalization process as described above to generate the synaptic simulation curve. This synaptic simulation curve is then output as the final result. Since the update error value corresponding to this curve already meets the error threshold, the output curve has overall consistency with the target synaptic learning curve within a preset accuracy range, and can be used for subsequent synaptic behavior simulation or neuromorphic computation weight update scenarios.

[0037] In another preferred embodiment based on the above embodiments, see [reference] Figure 2As shown, this embodiment provides a synapse simulation system for resistive switching devices, used to implement the above-described synapse simulation method for resistive switching devices, including: The parameter initialization module is configured to acquire the target synaptic learning curve and determine the set of sampling points, set the error threshold, the initial value of the uniform time interval, and the adjustment step size. The signal configuration module is configured to use at least two different frequency sinusoidal drive signals and to determine a preset frequency sequence. The conductivity data reading module is configured to apply the sinusoidal drive signal according to the initial value of the uniform time interval and the preset frequency sequence, and read the conductivity of the resistive switching device to obtain a conductivity value sequence. The curve evaluation module is configured to construct a synaptic simulation curve based on the conductivity value sequence, and calculate the root mean square error between the synaptic simulation curve and the target synaptic learning curve based on the set of sampling points to obtain the error value; The decision-making module is configured to output a synaptic simulation curve when the error value is less than or equal to the error threshold; when the error value is greater than the error threshold, it calculates the conductance increment of adjacent sampling points based on the set of sampling points, and obtains an increment difference sequence based on the difference between the conductance increment of the synaptic simulation curve and the conductance increment of the target synaptic learning curve. It then calculates the average increment difference based on the increment difference sequence and determines the time interval adjustment direction based on the sign of the average increment difference. The iterative execution module is configured to adjust the direction and the adjustment step size according to the time interval, update the unified time interval, and determine the update error value based on the updated unified time interval. The convergence control module is configured to update the adjustment step size and correct the time interval adjustment direction based on the comparison result between the error value and the updated error value. When the updated error value is greater than or equal to the error value, the time interval adjustment direction is reversed and the adjustment step size is halved until the updated error value is less than the error threshold.

[0038] Understandably, by dividing the synaptic simulation process into functional modules such as parameter initialization, signal configuration, conductance reading, curve evaluation, decision-making, iterative execution, and convergence control, a clear data flow and control flow are formed for target curve acquisition, drive excitation output, device response acquisition, error quantification evaluation, and closed-loop parameter updating. This enables repeatable fitting control of the target synaptic learning curve. Specifically, the signal configuration module fixes a preset frequency sequence and keeps the amplitude and duration of the sinusoidal drive signal constant, ensuring that the excitation conditions in each iteration remain consistent in the frequency dimension. This reduces the additional uncertainty introduced by changes in excitation conditions, ensuring that error variations are mainly caused by uniform time intervals. Driven by updates, the convergence control module can easily determine stability. The curve evaluation module quantifies the overall deviation between the simulated synapse curve and the target synapse learning curve based on the root mean square error, providing a unified and comparable numerical scale for good and bad fitting, avoiding misjudgments based solely on local segment or single-point differences. When the error exceeds the threshold, the decision-making module further introduces the calculation of the difference in conductance increment between adjacent sampling points as a direction indicator, so that the time interval adjustment not only depends on the magnitude of the error but also utilizes the local trend of curve changes to provide the direction of increase or decrease, reducing the number of iterations caused by blind trial and error. The calculated average increment difference is used to provide the time interval adjustment with low computational complexity. The effectiveness of the direction determination criterion relies on the premise that the incremental difference sequence can represent the main change range. To reduce the risk of direction information being canceled out due to differences in the magnitude of increments in different ranges, some implementations can use a weighted average of the incremental difference sequence based on the absolute value of the target incremental sequence, so that the range with more significant changes has a higher weight in the direction determination. At the same time, when the target curve has a plateau segment that makes the number of incremental difference terms that can be used for averaging after removing zero increment terms lower than the expected number, the direction adjustment of the previous time interval can be kept unchanged, and the direction reversal and step size halving can be triggered mainly by comparing the updated error value with the error value, thereby maintaining closed-loop convergence. Alternatively, before averaging, only the intervals with non-zero target increments and more significant changes are retained for calculation to enhance the representativeness of the direction determination for the effective change segment and improve reliability; the iterative execution module updates the unified time interval according to the adjustment direction and step size, and reuses the same application, reading, construction and error calculation links to ensure that the error value after each update can be directly compared with the error value of the previous round, enhancing the consistency of closed-loop feedback; the convergence control module reverses the direction and halves the step size when the updated error value does not improve, so that the system can automatically reduce the search granularity and switch the search direction when the error decrease direction deviates or the step size is too large and causes overshoot, suppressing oscillation and improving convergence stability.Therefore, without changing the frequency sequence, signal amplitude, or duration, this system uses a uniform time interval as the sole primary adjustment variable. Through a closed-loop iterative mechanism combining error assessment and direction determination, it gradually reduces the fitting error of the target synapse learning curve, thereby improving the controllability and reproducibility of the synapse simulation results. It also helps to obtain more stable fitting outputs under different resistive switching devices or different batch conditions.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A synaptic simulation method for resistive switching devices, characterized in that, include: Obtain the target synaptic learning curve and determine the set of sampling points, set the error threshold, the initial value of the unified time interval and the adjustment step size; Preset at least two sinusoidal drive signals of different frequencies and determine the preset frequency sequence; The sinusoidal drive signal is applied according to the initial value of the unified time interval and the preset frequency sequence, and the conductance of the resistive switching device is read to obtain the conductance value sequence. A synaptic simulation curve is constructed based on the conductivity value sequence, and the root mean square error between the synaptic simulation curve and the target synaptic learning curve is calculated based on the set of sampling points to obtain the error value. When the error value is less than or equal to the error threshold, the synaptic simulation curve is output; when the error value is greater than the error threshold, the conductance increment of adjacent sampling points is calculated based on the sampling point set, and an increment difference sequence is obtained based on the difference between the conductance increment of the synaptic simulation curve and the conductance increment of the target synaptic learning curve. The average increment difference is calculated based on the increment difference sequence, and the time interval adjustment direction is determined based on the sign of the average increment difference. The unified time interval is updated according to the adjustment direction and the adjustment step size, and an update error value is determined based on the updated unified time interval; The adjustment step size is updated and the time interval adjustment direction is corrected based on the comparison result between the error value and the updated error value. When the updated error value is greater than or equal to the error value, the time interval adjustment direction is reversed and the adjustment step size is halved until the updated error value is less than the error threshold.

2. The synaptic simulation method for resistive switching devices according to claim 1, characterized in that, When acquiring the target synaptic learning curve and determining the set of sampling points, setting the error threshold, the initial value of the uniform time interval, and the adjustment step size, the following are included: Obtain the time coordinates and amplitude coordinates of the target synaptic learning curve, and determine the time series of the sampling point set based on the time coordinates; When the sampling density of the target synaptic learning curve is inconsistent in different time periods, the target synaptic learning curve is resampled to the sampling point set based on interpolation; The average time interval is calculated based on the time interval between adjacent sampling points in the sampling point set, and the average time interval is used as the initial value of the unified time interval. The adjustment step size is half of the initial value of the unified time interval.

3. The synaptic simulation method for resistive switching devices according to claim 2, characterized in that, When pre-setting at least two sinusoidal drive signals of different frequencies and determining the preset frequency sequence, the process includes: According to the preset frequency sequence, the sinusoidal driving signals corresponding to different frequencies are arranged in ascending order of frequency to form a first sequence, and the first sequence is arranged in descending order of frequency to form a second sequence; A symmetrical frequency sequence is obtained by splicing the first sequence and the second sequence together, and the symmetrical frequency sequence is used as the preset frequency sequence. Based on the preset frequency sequence, the position of the conductivity value after each application is marked to obtain the conductivity index corresponding to the frequency position.

4. The synaptic simulation method for resistive switching devices according to claim 3, characterized in that, When applying the sinusoidal drive signal according to the initial value of the uniform time interval, and reading the conductance of the resistive switching device after each application to obtain the conductance value sequence, the following steps are included: The sinusoidal driving signal is triggered based on the zero-crossing point of the sinusoidal driving signal; The reading time is determined to be the immediate moment after the end of the sinusoidal driving signal based on the duration of the sinusoidal driving signal, and a reading signal is applied at the reading time to obtain the reading current; The conductivity value is calculated based on the read current and the read signal, and the average of the conductivity values ​​after the same application is obtained to obtain the conductivity value corresponding to the application. The conductivity values ​​corresponding to each application are written into the conductivity value sequence according to the application order.

5. The synaptic simulation method for resistive switching devices according to claim 4, characterized in that, A synaptic simulation curve is constructed based on the conductivity value sequence, and the root mean square error between the synaptic simulation curve and the target synaptic learning curve is calculated based on the set of sampling points. The error value is obtained by including: The conductivity value sequence is rearranged according to the order of the sampling point set to obtain the simulated conductivity sequence corresponding to the sampling point set; The simulated conductance sequence is normalized based on the maximum and minimum values ​​of the simulated conductance sequence in the same iteration to obtain the synaptic simulation curve; The error sequence is obtained by subtracting the synaptic simulation curve and the target synaptic learning curve point by point based on the set of sampling points, and the root mean square error is calculated as the error value based on the sum of squares of the error sequence and the number of sampling points.

6. The synaptic simulation method for resistive switching devices according to claim 5, characterized in that, When determining the direction of time interval adjustment based on the sign of the average incremental difference, the following are included: Based on the set of sampling points, the conductance increment of adjacent sampling points is calculated on the target synapse learning curve to obtain the target increment sequence, and based on the set of sampling points, the conductance increment of adjacent sampling points is calculated on the synapse simulation curve to obtain the simulation increment sequence; The incremental difference sequence is obtained by calculating the difference between the simulated incremental sequence and the target incremental sequence point by point. When there is a zero increment term in the target increment sequence, the corresponding increment difference term is removed from the averaging operation, and the average increment difference is calculated for the remaining increment difference terms. The direction of the time interval adjustment is determined based on the sign of the average incremental difference.

7. The synaptic simulation method for resistive switching devices according to claim 6, characterized in that, When updating the unified time interval based on the time interval adjustment direction and the adjustment step size, and determining the update error value based on the updated unified time interval, the process includes: When the time interval adjustment direction is to increase, the updated unified time interval is obtained by adding the unified time interval to the adjustment step size; When the time interval adjustment direction is decreasing, the updated unified time interval is obtained by subtracting the adjustment step size from the unified time interval; When the updated uniform time interval is less than the duration of the sinusoidal drive signal, the updated uniform time interval is corrected to the duration of the sinusoidal drive signal based on the duration of the sinusoidal drive signal. Based on the updated uniform time interval and according to the preset frequency sequence, the sinusoidal drive signal is applied and the conductance of the resistive switching device is read to obtain the updated conductance value sequence. An updated synaptic simulation curve is constructed based on the updated conductivity value sequence, and the root mean square error between the updated synaptic simulation curve and the target synaptic learning curve is calculated based on the set of sampling points to obtain the update error value.

8. The synaptic simulation method for resistive switching devices according to claim 7, characterized in that, When updating the adjustment step size and correcting the time interval adjustment direction based on the comparison result between the error value and the updated error value, the following steps are included: When the update error value is less than the error value, the updated unified time interval is used as the current unified time interval, and the update error value is used as the current error value, while maintaining the time interval adjustment direction and the adjustment step size; When the update error value is greater than or equal to the error value, maintain the same time interval as before the update and maintain the error value before the update, reverse the adjustment direction of the time interval and halve the adjustment step size; The next round of unified time interval update will proceed based on the updated unified time interval, the time interval adjustment direction, and the adjustment step size.

9. The synaptic simulation method for resistive switching devices according to claim 8, characterized in that, Until the update error value is less than the error threshold, including: Based on the uniform time interval, the update error value, and the difference between the average increment recorded in any round of iteration, an iterative record sequence is obtained; Based on the iterative record sequence, a uniform time interval that makes the update error value satisfy the error threshold is determined as the target uniform time interval; The synaptic simulation curve is constructed and output based on the conductivity value sequence corresponding to the target uniform time interval, as the synaptic simulation result.

10. A synaptic simulation system for resistive switching devices, used to implement the synaptic simulation method for resistive switching devices as described in any one of claims 1-9, characterized in that, include: The parameter initialization module is configured to acquire the target synaptic learning curve and determine the set of sampling points, set the error threshold, the initial value of the uniform time interval, and the adjustment step size. The signal configuration module is configured to use at least two different frequency sinusoidal drive signals and to determine a preset frequency sequence. The conductivity data reading module is configured to apply the sinusoidal drive signal according to the initial value of the uniform time interval and the preset frequency sequence, and read the conductivity of the resistive switching device to obtain a conductivity value sequence. The curve evaluation module is configured to construct a synaptic simulation curve based on the conductivity value sequence, and calculate the root mean square error between the synaptic simulation curve and the target synaptic learning curve based on the set of sampling points to obtain the error value; The decision-making module is configured to output the synaptic simulation curve when the error value is less than or equal to the error threshold; when the error value is greater than the error threshold, calculate the conductance increment of adjacent sampling points based on the sampling point set, obtain an increment difference sequence based on the difference between the conductance increment of the synaptic simulation curve and the conductance increment of the target synaptic learning curve, calculate the average increment difference based on the increment difference sequence, and determine the time interval adjustment direction based on the sign of the average increment difference. The iterative execution module is configured to update the unified time interval by adjusting the direction and the adjustment step size according to the time interval, and to determine an update error value based on the updated unified time interval. The convergence control module is configured to update the adjustment step size and correct the time interval adjustment direction based on the comparison result between the error value and the updated error value. When the updated error value is greater than or equal to the error value, the time interval adjustment direction is reversed and the adjustment step size is halved until the updated error value is less than the error threshold.

Images