A method for optimizing an indium-aluminum-zinc-oxide-based optoelectronic neuromorphic array

Abstract

The application discloses an indium-aluminum-zinc oxide-based photoelectric neuromorphic array optimization method, relates to the technical field of photoelectric neuromorphic arrays, and is used for solving the problems of low reading accuracy and poor parallel operation efficiency caused by inherent crosstalk, device nonlinear distortion and integration contradictions of existing technologies in an indium-aluminum-zinc oxide-based photoelectric neuromorphic cross array. The method comprises the following steps: dividing the entire photoelectric synapse cross array into a plurality of independent sub-arrays, and providing a hierarchical virtual ground reading circuit for each sub-array; screening an update target that needs to be adjusted in conductance based on neural network gradient calculation; simultaneously applying light pulses to the update target by using a programmable light source to perform parallel coarse adjustment, and combining an adaptive auxiliary bias to suppress crosstalk; finally, performing a loop verification operation on each update target to realize high-precision conductance adjustment, effectively suppress bypass current in the cross array, and improve reading and writing accuracy and parallel operation efficiency, and the method is suitable for high-density and high-precision neuromorphic computing systems.

Description

Technical Field

[0001] This invention relates to the field of indium aluminum zinc oxide (IZO) optoelectronic neuromorphic array technology, specifically to an optimization method for IZO-based optoelectronic neuromorphic arrays. Background Technology

[0002] Indium aluminum zinc oxide (IZO) is an advanced class of amorphous oxide semiconductor thin film materials widely used as the functional layer for constructing optoelectronic artificial synaptic devices. This type of material possesses significant photosensitivity, excellent electrical properties, good flexibility and compatibility, and low power consumption. However, a single synaptic device itself does not possess complete computational capabilities; only by integrating synaptic devices and artificial neuron circuits into a high-density and regularly structured network in hardware can practically valuable neuromorphic computing be achieved. The most common organizational form is the cross-array structure: in the array, horizontal and vertical wires serve as input and output lines, respectively, and each intersection integrates an IZO optoelectronic synaptic device.

[0003] The commonly used cross-array in indium aluminum zinc-based optoelectronic neuromorphic arrays is a purely passive resistive network. When reading or writing a target cell, the current flows not only through the target cell but also through all other cells sharing the same row or column, forming complex parallel and series paths. The resistance state of non-target cells can steal or shunt some current, causing the read current value to fail to accurately reflect the weight of the target cell, thus generating calculation errors. At the same time, the device itself has non-ideal characteristics (such as nonlinear conductance transition characteristics), which further exacerbates crosstalk: a non-target cell in a low-resistance state can become a significant bypass, completely drowning out the signal from the target cell, thus aggravating crosstalk. Existing technologies typically introduce nonlinear selectors (such as 1S1R) at the device level, which makes it difficult to achieve high-precision, parallel operation in large-scale arrays while ensuring integration density. Summary of the Invention

[0004] The purpose of this invention is to solve the problems mentioned in the background art by proposing an optimization method for photoelectric neuromorphic arrays based on indium aluminum zinc oxide.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] An optimization method for photoelectric neuromorphic arrays based on indium aluminum zinc oxide (IZO) includes the following steps:

[0007] Step 1: Divide the photoelectric synaptic cross array into several sub-arrays, each sub-array integrating a hierarchical virtual ground readout circuit; calculate the weight gradient of each unit based on the neural network gradient, select the units whose conductance needs to be adjusted as update targets, and map them into a list of operation instructions for the corresponding physical devices;

[0008] Step 2: Simultaneously apply light pulses to the target position using a programmable light source for parallel coarse adjustment, combined with an adaptive auxiliary bias voltage;

[0009] Step 3: In one training iteration of the neural network, the algorithm calculates a target weight for each weight according to the gradient descent rule of the array. A global weight-conductance mapping function is pre-defined. Each target weight calculated by the algorithm is uniquely converted into its corresponding physical target conductance value through the global weight-conductance mapping function. Thus, the target conductance value of each updated target can be obtained. Based on the target conductance value of the updated target, a loop verification operation is performed on each updated target until the conductance value reaches the target conductance value or the maximum number of iterations is reached.

[0010] In a preferred embodiment of the present invention, the operation instruction list includes global physical coordinates (i, j), the original weight gradient ∇W(m, n) before mapping, and the relative update intensity scale α(i, j), where i is the row index of the update target in the subarray whose conductance needs to be adjusted, j is the column index of the update target in the subarray whose conductance needs to be adjusted, m is the row index of the entire photosynaptic cross array, and n is the column index of the entire photosynaptic cross array.

[0011] As a preferred embodiment of the present invention, the hierarchical virtual ground readout circuit includes transimpedance amplifier channels, each channel corresponding to one or more columns of multiplexing, and the input terminal is servoed to near zero potential during the readout operation.

[0012] In a preferred embodiment of the present invention, the process of selecting units whose conductance needs to be adjusted as update targets is as follows:

[0013] In one training iteration of a neural network, the weight gradient matrix ∇W of a specific fully connected layer is calculated using the backpropagation algorithm. Each element ∇W(m,n) of the weight gradient matrix ∇W represents the update direction and magnitude of the logical weights W(m,n) in the network.

[0014] The gradient matrix is ​​filtered and grouped: a gradient threshold θ is set, and the weights are considered to need to be updated in this round only when |∇W(m,n)|>θ, and the corresponding physical units are marked as target positions; all target positions are divided into two categories according to their gradient signs: positive update group and negative update group; a differential pair encoding strategy is adopted to map the update requirements of the weights into conductance increase operations of the corresponding physical devices. After mapping, all update requirements from the algorithm are converted into conductance increase operations of the physical devices.

[0015] Based on the global coordinates of the target location in the physical array, the space is further divided according to the subarray block to which it belongs, so that each target location is assigned to its corresponding subarray, thus obtaining the updated target in the subarray; for the updated target in the subarray, a structured list of operation instructions is generated.

[0016] As a preferred embodiment of the present invention, parallel coarse adjustment includes:

[0017] S2-1: Receive the update target and operation instruction list of the subarray; designate the row containing the update target as the relevant row and the column containing the update target as the relevant column, thus forming a relevant row set and a relevant column set; the row index of at least one update target in the subarray constitutes the relevant row set, and the column index of at least one update target in the subarray constitutes the relevant column set; all rows and columns in the subarray that do not contain the aforementioned relevant row set and relevant column set are designated as unrelated rows and unrelated columns of the subarray; extract electrical state parameters from the update target and its corresponding relevant row and relevant column, specifically including row load, column load, and the center conductance value of the subarray;

[0018] S2-2: Based on the adaptive calculation of electrical state parameters, the auxiliary bias voltage is used to obtain the voltage compensation terms for updating the target related row i and related column j.

[0019] S2-3: Perform optoelectronic coordinated timing control adjustment based on the voltage compensation terms of the relevant row i and relevant column j of the updated target to complete a parallel optical coarse adjustment operation.

[0020] In a preferred embodiment of the present invention, the extraction of electrical state parameters includes:

[0021] For each relevant row i, count the number of update targets on that relevant row, and denote it as the row load as Ni; similarly, for each relevant column j, count the number of update targets on that relevant column, and denote it as the column load as Nj;

[0022] Record the current conductance value of all updated targets in the subarray as G(i,j), and then calculate the geometric mean of the current conductance values ​​of all updated targets in the subarray as the center conductance value of the subarray.

[0023] In a preferred embodiment of the present invention, the calculation of the adaptive auxiliary bias voltage includes:

[0024] Based on material properties and experimental calibration, a safety baseline value related to the center conductivity is set: baseline bias value V. base To compensate for the line voltage drop caused by the simultaneous conduction of multiple update targets in the same row or column, a voltage compensation term based on its load is calculated for each relevant row i and relevant column j; the calculation formula is:

[0025] For each relevant row i: ;

[0026] For each relevant column j: ;

[0027] Where k is the compensation coefficient calibrated based on the wire resistance and the updated target on-resistance, Nmax(j) represents the maximum theoretical load number of the relevant row i, that is, the total number of columns of the subarray in the relevant row; Nmax(i) represents the maximum theoretical load number of the relevant column j, that is, the total number of rows of the subarray in the relevant column.

[0028] As a preferred embodiment of the present invention, the optoelectronic coordinated timing control adjustment includes:

[0029] For all relevant rows i in the subarray, apply their corresponding calculated voltage compensation term V. compensate (i) Apply the corresponding voltage compensation term V to all relevant columns j in the subarray. compensate (j); Simultaneously, the potentials of all unrelated rows and columns are actively clamped to a fixed reference ground potential. After the row and column bias voltages stabilize, the programmable light source integrated above the subarray is triggered to generate a two-dimensional light spot pattern corresponding to the coordinates (i, j) of the updated target. The light pulse parameters are used to illuminate only the updated target position, including wavelength, pulse width, and pulse intensity. The pulse intensity is spatially modulated according to the relative update intensity scale α(i, j) of the updated target, i.e., α(i, j) is positively correlated with the pulse intensity. After the light pulse ends, the row and column bias voltages are maintained for a fixed duration, and the row and column drivers are controlled to smoothly restore all line voltages to the standby level, completing one parallel optical coarse adjustment operation.

[0030] In a preferred embodiment of the present invention, the cyclic verification operation includes:

[0031] S3-1 uses a hierarchical virtual ground readout circuit to perform a high-fidelity readout of the updated target (i, j): a readout voltage is applied to the row containing the updated target, the column containing the updated target is connected to the virtual ground, and all non-target rows and columns are clamped to a balanced bias, i.e., balanced bias = readout voltage / 2; the current is obtained through an analog-to-digital converter, and combined with the known readout voltage and transimpedance gain, the conductance value of the updated target (i, j) after optical coarse adjustment is calculated and recorded, and the target conductance value G'(i, j) of the updated target is subtracted from it to obtain the initial error; the error tolerance δ and the maximum number of iterations are preset;

[0032] S3-2, enter a cyclic iterative electrical pulse fine-tuning from n=1 to the maximum iteration number, where n is the index of the cyclic iteration number;

[0033] S3-3, Weight Confirmation and Update: After completing the weight update of a complete network layer, a global verification read is initiated to update all successfully fine-tuned final conductance values ​​to the non-volatile hardware weight lookup table, thereby completing a complete update from algorithm requirements to hardware state in a closed loop.

[0034] As a preferred embodiment of the present invention, the cyclic iterative electrical pulse fine-tuning includes:

[0035] S3-2-1, determine the parameters of the current electrical pulse based on the sign and amplitude of the initial error. The electrical pulse parameters include pulse amplitude and pulse width. The pulse amplitude is set within the safe adjustable range of the device, and the polarity is determined by the sign of the initial error: if the initial error > 0, apply a positive pulse; if the initial error < 0, apply a negative pulse. The pulse width is adaptively adjusted according to the initial error, and the adaptive adjustment calculation formula is as follows: Where Δg(i,j) is the initial error of updating target (i,j), G'(i,j) is the target conductance value of the updated target, and t min and t max β represents the minimum and maximum pulse widths, and β is the sensitivity coefficient. When applying an electrical pulse according to the electrical pulse parameters, the pulse amplitude is applied only to row i where the updated target is located, while column j where the updated target is located remains grounded, and a balanced bias is applied to all unrelated rows and columns.

[0036] S3-2-2, after the pulse ends, wait for the stabilization time, and then use the hierarchical virtual ground reading circuit again to measure the new conductance value at the updated target. Subtract the new conductance value from the target conductance value of the updated target to obtain the new error.

[0037] S3-2-3, if the absolute value of the new error is less than or equal to the error tolerance δ, the fine-tuning of the updated target is determined to be successful, the final conductance value is recorded, and the loop is exited; if the absolute value of the new error is greater than or equal to the error tolerance δ, the initial error is set to the new error, and the loop is entered again; if the accuracy is still not met after the maximum number of loops, the fine-tuning is determined to be a failure.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] 1. By dividing a large-scale cross array into multiple independent subarrays and integrating hierarchical virtual ground readout circuits in each subarray, the inherent crosstalk problem of passive networks is effectively solved at the system architecture level. Then, by clamping the target column to the virtual ground and actively driving all non-target rows and columns to a balanced bias, the voltage difference across the crawling path is reduced to a maximum of half the readout voltage. This physically reduces the crawling current by multiples or even exponentially, ensuring that the readout signal can truly reflect the weight of the target cell, laying the foundation for high-precision calculation.

[0040] 2. By utilizing a programmable light source, spatially modulated light pulses are applied to a batch of updated targets in the subarray simultaneously, achieving highly parallel physical operations. Based on the current conductance state and line load, adaptive auxiliary bias is dynamically calculated to compensate for IR voltage drop, optimize photogenerated carrier injection efficiency, and prevent accidental writing. Simultaneously, the light intensity is spatially modulated according to the gradient amplitude to achieve differentiated and adaptive coarse adjustment of a large number of physical units. During the optical modulation process, unrelated rows and columns are actively clamped to ground potential, completely cutting off non-target paths, ensuring low crosstalk in the parallel operation environment, and guaranteeing the accuracy of coarse adjustment.

[0041] 3. After coarse tuning, a closed-loop iterative process of reading, calculating errors, applying adaptive electrical pulses, and then reading again can accurately converge the conductance value to the target value, with the error controlled within a preset tolerance. The adaptive pulse width formula used ensures fine approximation with small errors and rapid convergence with large errors, achieving an optimal balance between speed and accuracy. The final accurate conductance value is updated to a non-volatile lookup table, providing an accurate and reliable physical basis for subsequent inference or the next round of training, realizing a complete and reliable update closed loop from algorithm requirements to hardware status. Attached Figure Description

[0042] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0043] Figure 1 This is a general block diagram of the method of the present invention. Detailed Implementation

[0044] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Application Scenarios: Existing indium aluminum zinc-based optoelectronic neuromorphic arrays commonly use cross-arrays, which are purely passive resistive networks. When reading or writing a target cell, the current flows not only through the target cell but also through all other cells sharing the same row or column, forming complex parallel and series paths. The resistance state of non-target cells can steal or shunt some current, causing the read current value to fail to accurately reflect the weight of the target cell, thus generating calculation errors. Simultaneously, the device itself has non-ideal characteristics (such as nonlinear conductance transition characteristics), which further exacerbate crosstalk: a non-target cell in a low-resistance state becomes a significant bypass, completely drowning out the target cell's signal, thus intensifying crosstalk. To solve the above technical problems, this invention employs a metal-insulator-semiconductor-metal stacked design. The specific implementation involves constructing a heterojunction on the bottom electrode consisting of an IAZO (indium aluminum zinc) active layer (30 nm), an In2O3 nanowire network layer (10 nm), and an IAZO active layer (20 nm). In this structure, the interfacial barrier formed by the In2O3 nanowires and the upper and lower IAZO layers imparts key self-rectification characteristics to the device: at low readout voltages, current is significantly suppressed, reducing the creeping current flowing through non-target cells. Simultaneously, this heterojunction provides a channel for the efficient separation and injection of photogenerated carriers, and the light pulse can effectively modulate the barrier height. To further improve the response efficiency to visible light, a perovskite quantum dot or aluminum nanoparticle layer can be integrated below the top electrode as a photosensitive enhancement structure. Based on this, the following IAZO opto-neuromorphic array optimization method is implemented.

[0046] Please see Figure 1 As shown, an optimization method for photoelectric neuromorphic arrays based on indium aluminum zinc oxide includes the following steps:

[0047] Step 1: Divide the entire photosynaptic cross array into N independent subarrays. For example, a 64×64 photosynaptic cross array can be divided into 16 8×8 subarrays, where 64×64 refers to 64 rows and 64 columns, and the same applies to 8×8. Block division reduces row and column line resistance and bypass complexity, and makes the measurement matrix tend towards a block diagonal structure, facilitating matrix correction. Each subarray integrates a hierarchical virtual ground readout circuit to achieve high-fidelity, low-crosstalk signal reading. Specifically, each subarray will be equipped with an independent readout control unit, which contains several transimpedance amplifier channels. Each channel corresponds to one or more columns multiplexed. The input of the transimpedance amplifier is at the readout... During the read operation, the node is treated as a virtual ground node and servoed to near zero potential. When reading the target position, the following bias is applied: only the target row is applied with a read voltage, and the target column is connected to the virtual ground input of the corresponding transimpedance amplifier. All non-target rows and columns are actively driven to a balanced bias voltage, typically read voltage / 2. This bias scheme ensures that the voltage difference across any non-target path (e.g., from the target row through a non-target cell to a non-target column) is at most (read voltage / 2), rather than the read voltage under traditional connection, thus physically reducing creep current by a factor of two. The read signal within the subarray is converted to digital signal locally before being transmitted to the upper-level aggregator, achieving hierarchical signal processing.

[0048] The target locations where conductivity needs to be adjusted are determined based on neural network gradient calculations.

[0049] In one training iteration of a neural network, the weight gradient matrix ∇W of a specific fully connected layer is calculated using the backpropagation algorithm. Each element ∇W(m, n) of this matrix represents the update direction and magnitude of the logical weights W(m, n) in the network. If ∇W(m, n) > 0, it means that the weights need to be increased (corresponding to an increase in conductivity). If ∇W(m, n) < 0, it means that the weights need to be decreased (corresponding to a decrease in conductivity). m and n are the row index and column index of the entire array, respectively.

[0050] To improve operational efficiency and reduce power consumption, the gradient matrix is ​​filtered and grouped: a gradient threshold θ is set (e.g., θ = 0.05 * max(|∇W|)). The gradient threshold θ is set to filter out updates that contribute little to network optimization, thereby improving system energy efficiency and operating speed, and reducing unnecessary operations on synaptic devices. The value of the gradient threshold θ is usually determined dynamically based on empirical formulas (e.g., taking a fixed proportion of the maximum absolute value of the gradient matrix in this round) or based on the target sparsity (e.g., only updating the top 10% of weights in terms of gradient magnitude). Only when |∇W(m,n)| > θ When a weight is considered to need updating in this round, its corresponding physical unit is marked as the target position. All target positions are divided into two categories based on their gradient signs: positive update group and negative update group. To maximize the parallel efficiency of the system, this invention preferably adopts a differential pair encoding strategy, mapping the weight update requirement (regardless of gradient sign) to the conductance increase operation of the corresponding physical device. Specifically, if ∇W(m, n) > 0 (needs to be increased), it is mapped to the conductance increase operation of the positive device; if ∇W(m, n) < 0 (needs to be decreased), it is mapped to the conductance increase operation of the negative device. After mapping... Subsequently, all update requests from the algorithm are converted into conductance increase operations on the physical devices. To improve parallelism, the space is further divided according to the target position's global coordinates in the physical array and its corresponding subarray blocks, so that each target position is assigned to its corresponding subarray, resulting in updated targets within the subarray. For example, all updated target positions located within the same subarray are grouped into the same batch so that the local light source and circuitry of that subarray can operate in coordination. For updated targets within a subarray, a structured list of operation instructions is generated, which includes global physical... The coordinates (i, j), the original weight gradient ∇W(m, n) before mapping, and the relative update intensity scale α(i, j) are used to pass the update target in each subarray and the list of operation instructions corresponding to the update target to step 200 for coarse adjustment; where i and j are the local row and column coordinate order in the subarray, and the relative update intensity scale is the parameter α(i, j) obtained by normalizing the magnitude of the absolute value of the weight gradient of the update target |∇W(m, n)|, which is in the range of 0-1 and is used to indicate the relative light energy weight that the update target should obtain in this parallel coarse adjustment.

[0051] By dividing a large-scale cross array into multiple independent subarrays and integrating hierarchical virtual ground readout circuits in each subarray, the inherent crosstalk problem of passive networks is effectively solved at the system architecture level. Then, by clamping the target column to the virtual ground and actively driving all non-target rows and columns to a balanced bias, the voltage difference across the crawling path is reduced to a maximum of half the readout voltage. This physically reduces the crawling current by multiples or even exponentially, ensuring that the readout signal can truly reflect the weight of the target cell, laying the foundation for high-precision calculation.

[0052] Step two involves simultaneously applying light pulses to multiple target locations requiring conductance upregulation using a programmable light source integrated above or below the subarray. The wavelength of the light pulses must match the absorption spectrum of the photosensitive enhancement layer (such as perovskite quantum dots). During this process, relevant rows and columns of the target locations are selected from the subarray, and the relevant rows and columns are...

[0053] A small auxiliary bias voltage is applied to promote the injection of photogenerated carriers into the IAZO heterojunction channel, enabling rapid and parallel coarse adjustment of the conductance; the specific process is as follows:

[0054] S2-1, define the relevant rows and columns, and extract the key electrical state parameters:

[0055] The system receives the update targets and operation instruction list of the subarray. The operation instruction list includes the local coordinates (i, j) of all update targets that need to be coarsely adjusted within the subarray in this batch, their corresponding original weight gradient values ​​∇W(m, n), and the calculated relative update intensity scale α(i, j). The row containing the update target is designated as the relevant row, and the column containing the update target is designated as the relevant column, thus forming a relevant row set and a relevant column set. The row index of at least one update target in the subarray constitutes the relevant row set, and the column index of at least one update target in the subarray constitutes the relevant column set. All rows and columns in the subarray that do not contain the above-mentioned relevant row set and relevant column set are designated as unrelated rows and unrelated columns of the subarray. During the coarse adjustment operation, unrelated rows and unrelated columns will be actively managed to isolate electrical interference.

[0056] For each relevant row i, the number of update targets in that row is counted and denoted as Ni as the row load; similarly, for each relevant column j, the number of update targets in that column is counted and denoted as Nj as the column load; the row and column loads are used to evaluate the IR voltage drop of the conductor caused by multiple update targets being turned on at the same time;

[0057] The current conductance value of all updated targets in the subarray is recorded as G(i,j). Then, the geometric mean of the current conductance values ​​of all updated targets in the subarray is calculated as the center conductance value of the subarray. In practical applications, the geometric mean is not sensitive to extreme values ​​and can more robustly reflect the typical conduction state center trend of the batch of devices, providing a reliable basis for the subsequent stable calculation of auxiliary bias. The center conductance value directly affects the effective transport and injection efficiency of photogenerated carriers in the channel.

[0058] S2-2: Adaptive auxiliary bias calculation:

[0059] The calculation of the auxiliary bias voltage needs to seek the optimal solution under three constraints: (1) effectively promote the injection of photogenerated carriers; (2) be absolutely lower than the electrical write threshold voltage of the device to prevent accidental writing; (3) fully consider the influence of IR voltage drop caused by the row line load.

[0060] Based on material properties and experimental calibration, a safety baseline value related to the center conductivity is set: baseline bias value V. base For example, through prior testing, a lookup table is established: when the center conductivity value is in the low, medium, or high range, the base bias value V... base The base bias voltage was set to 0.15V, 0.12V, and 0.08V respectively. The higher the conductivity value, the lower the required bias voltage for carrier injection can be. base Relatively low; in practical applications, the following calibration procedure must be performed before commissioning or during periodic calibration to establish V. base Lookup table: Programs the test unit to a series of known, stable conductance states G test (Covering low, medium, and high conductance ranges); applying different bias voltages V at each conductance state. bias (Starting from 0V, gradually increasing, but strictly below the electrical writing threshold V), while simultaneously applying a standard optical pulse; measuring and recording the instantaneous change in conductance ΔG caused by each optical + electrical co-operation; for each conductance state G test Select a variable that produces a significant and reversible ΔG (e.g., ΔG / G). test The highest safe voltage (>5%) is used as the base bias value V corresponding to this conductance range. base And record it in the lookup table;

[0061] To compensate for the line voltage drop caused by the simultaneous conduction of multiple update targets in the same row or column, a voltage compensation term based on its load is introduced for each relevant row i and relevant column j; the calculation formula is as follows:

[0062] For each relevant row i: ;

[0063] For each relevant column j: ;

[0064] Where k is the compensation coefficient calibrated based on the wire resistance and the updated target on-resistance (in this embodiment, k is 0.1), Nmax(j) represents the maximum theoretical load number of the relevant row i, i.e., the total number of columns of the subarray in the relevant row; Nmax(i) represents the maximum theoretical load number of the relevant column j, i.e., the total number of rows of the subarray in the relevant column; the net auxiliary bias voltage obtained at both ends of the updated target (i, j) is approximately Vnet(i, j) ≈ V compensate (i)-V compensate (j) By adjusting the value of k, it can be designed to approximate the base bias value V. base ;

[0065] S2-3: Optoelectronic Coordination Timing Control Adjustment:

[0066] For all relevant rows i in the subarray, apply their corresponding calculated voltage compensation term V.compensate (i) Apply the corresponding voltage compensation term V to all relevant columns j in the subarray. compensate (j); Simultaneously, the potentials of all unrelated rows and columns are actively clamped to a fixed reference ground potential (0V), completely cutting off the creeping current path through non-target cells and suppressing crosstalk to a minimum level to eliminate its interference with the updated target (i,j); After the row and column bias voltages stabilize, the programmable light source integrated above the subarray is triggered to generate a two-dimensional light spot pattern corresponding to the coordinates (i,j) of the updated target, illuminating only the updated target position. The light pulse parameters are as follows: wavelength, pulse width, and pulse intensity; where the wavelength corresponds to the absorption peak of the device's photosensitive enhancement layer (such as perovskite quantum dots). The photon absorption and carrier generation efficiency are matched, for example, to 520±10nm; the pulse width is set according to the device's photoresponse speed, typically on the order of 1ms to 10ms; the pulse intensity is spatially modulated according to the relative update intensity scale α(i,j) of the update target, that is, for update targets with larger α(i,j) values, stronger or longer light pulses are applied to achieve differentiated coarse adjustment amplitudes, initially approximating the update ratio indicated by the gradient; after the light pulse ends, the row and column bias voltages are maintained for a period of time to ensure that all photogenerated carriers are effectively collected and the channel conductance is modulated. Subsequently, the row and column drivers are controlled to smoothly restore all line voltages to the standby level, completing one parallel optical coarse adjustment operation;

[0067] Active clamping is used to suppress array crosstalk to the greatest extent. The bidirectional, serial weight update requirements at the algorithm level are transformed into unidirectional, highly parallel physical operations at the hardware level through differential pair coding and optoelectronic coordination, which accelerates the training and adaptation process. The auxiliary bias voltage is dynamically calculated based on the real-time conductivity state and load conditions, and the light intensity is spatially modulated in combination with the gradient amplitude to achieve adaptive and differentiated fine control of a large number of synaptic units, which improves speed while ensuring adjustment quality.

[0068] Step 3: For each updated target after optical coarse adjustment, perform a localized cyclic verification operation until the error between the read or recovered conductivity value and the target value enters the tolerance range or the maximum number of attempts is reached; the specific process is as follows:

[0069] In one training iteration of the neural network, the algorithm calculates a target weight W' for each weight W(m,n) according to the gradient descent rule of the array: Where η is the learning rate; a global weight-conductance mapping function is pre-defined, which defines the correspondence between the weight value W and the conductance value of the physical device. The functional relationship is typically as follows: Here, 'a' is a scaling factor with the same dimensions as the conductance value. Essentially, it represents the baseline conductance value exhibited by the physical device when the weight W=0. More broadly, 'a' characterizes the lower limit or offset of the device's dynamic range of conductance. 'b' is a sensitivity factor, whose dimensions are the reciprocal of the weight W (i.e., the reciprocal of the dimensionless weight value). It quantitatively describes how a small change in the weight W is amplified by an exponential change in the physical conductance value. The larger the value of 'b', the more sensitive the conductance value is to changes in the weight W. Each target weight W' calculated by the algorithm is uniquely converted into its corresponding physical target conductance value through the above global weight-conductance mapping function. Thus, the target conductance value G'(i,j) for each updated target can be obtained.

[0070] For each updated target in the subarray, a high-precision cyclic verification operation is performed until its coarsely adjusted conductance value reaches the target conductance value or the maximum number of iterations is reached; the specific process of the cyclic verification operation is as follows:

[0071] S3-1, using the hierarchical virtual ground readout circuit described in step one, perform a high-fidelity readout of the updated target (i, j): apply a readout voltage (e.g., 0.2V) to the row containing the updated target, connect the column containing the updated target to virtual ground, and clamp all non-target rows and columns to a balanced bias, i.e., balanced bias = readout voltage / 2 (e.g., 0.1V); obtain the current through an analog-to-digital converter, combine the known readout voltage and transimpedance gain, calculate and record the conductance value of the updated target (i, j) after the optical coarse adjustment in step two, and subtract the target conductance value G'(i, j) of the updated target to obtain the initial error; preset the error tolerance δ and the maximum number of iterations; in this embodiment, the error tolerance δ is set as a relative proportion of the target conductance value G'(i, j) (e.g., 0.5%), which is set according to the accuracy requirements of the matching neural network calculation and the physical limits of the hardware implementation. A relative tolerance is used instead of an absolute tolerance because... Conductivity values ​​with different weights may span several orders of magnitude. The relative tolerance can adaptively unify the accuracy standard, ensuring that both high-conductivity and low-conductivity units obtain adjustment accuracy commensurate with their order of magnitude. The purpose of setting this is to avoid a surge in iterations, increased power consumption, and device wear due to the pursuit of unnecessary high precision, while ensuring the effectiveness of the algorithm. The setting of the maximum number of iterations (e.g., 20 times) is based on a comprehensive consideration of ensuring convergence reliability, preventing infinite loops, and ensuring system real-time performance. Under the control of the adaptive pulse width formula for closed-loop fine-tuning, the conductivity error should converge exponentially or approximately exponentially. Through prior device characteristic calibration and system simulation, it can be statistically shown that most normal units can meet the error tolerance within a limited number of iterations (e.g., 10-15 times). Setting an upper limit slightly higher than this statistical value (e.g., 20 times) leaves a safety margin for device fluctuations and noise, ensuring successful fine-tuning with a very high probability.

[0072] S3-2, enter a cyclic iterative electrical pulse fine-tuning process from n=1 to the maximum number of iterations:

[0073] S3-2-1, Calculate and apply the fine-tuning pulse: Determine the pulse parameters based on the sign and amplitude of the current initial error. The pulse parameters include pulse amplitude and pulse width. The pulse amplitude is set within the device's safe adjustable range, and its polarity is determined by the sign of the initial error: if the initial error > 0 (conductance needs to be adjusted upwards), apply a positive pulse; if the initial error < 0 (conductance needs to be adjusted downwards), apply a negative pulse. The pulse width is adaptively adjusted based on the initial error, and the adaptive adjustment calculation formula is: Where Δg(i,j) is the initial error of updating target (i,j), G'(i,j) is the target conductance value of the updated target, and t min and t max For the minimum pulse width and the maximum pulse width, t min It is the shortest time reference that can induce a reliable, measurable change in the conductivity of a device, ensuring the effectiveness of operation; t max It is the safety upper limit to prevent device damage, overheating, or over-adjustment, ensuring operational reliability; min and t max The safe and effective control range during the fine-tuning process is defined; β is the sensitivity coefficient, with a value of 0.001-0.005; this adaptive adjustment formula ensures that short pulses are used for fine approximation when the error is small, and long pulses are used for fast convergence when the error is large, thereby achieving the optimal balance between accuracy, speed and reliability; when applying the electrical pulse according to the electrical pulse parameters, the pulse amplitude is only applied to the row i where the updated target is located, the column j where the updated target is located is kept grounded, and a balanced bias is applied to all unrelated rows and columns to minimize interference to other cells in the array.

[0074] S3-2-2, Verification of Reading and Error Update: After the pulse ends, wait for a short stabilization time (e.g., 10µs), and then use the hierarchical virtual ground readout circuit again to accurately measure the new conductance value at the updated target. Subtract the new conductance value from the target conductance value of the updated target to obtain the new error.

[0075] S3-2-3, Loop Judgment: If the absolute value of the new error is less than or equal to the error tolerance δ, the fine-tuning of the updated target is considered successful, the final conductance value is recorded, and the loop is exited; if the absolute value of the new error is greater than or equal to the error tolerance δ, the initial error is set to the new error, and the loop continues; if the accuracy is still not met after the maximum number of loops, the fine-tuning is considered a failure, a fault flag is recorded, and a fault-tolerant strategy can be triggered (such as marking the cell, trying several low-power reverse pulses, or using interpolation of adjacent cells, etc.); the fine-tuning process of multiple updated targets within a subarray can be executed concurrently or in a high-speed time-sharing manner with the support of their independent or multiplexed control channels to improve overall efficiency;

[0076] S3-3: Weight Confirmation and Update: After completing the weight update of a complete network layer, initiate a low-frequency global verification read to ensure the integration accuracy of the overall weight matrix; update all successfully fine-tuned final conductance values ​​to the non-volatile hardware weight lookup table. This table is the authoritative basis for conducting-weight inverse mapping during all subsequent neural network forward inference or the next training iteration, thus completing a complete update from algorithm requirements to hardware status in a closed loop.

[0077] After coarse tuning, a closed-loop iterative process of reading, calculating errors, applying adaptive electrical pulses, and then reading again can accurately converge the conductance value to the target value, with the error controlled within a preset tolerance. The adaptive pulse width formula used ensures fine approximation with small errors and rapid convergence with large errors, achieving an optimal balance between speed and accuracy. The final accurate conductance value is updated to a non-volatile lookup table, providing an accurate and reliable physical basis for subsequent inference or the next round of training, realizing a complete and reliable update closed loop from algorithm requirements to hardware status.

[0078] In summary, this invention, based on the IAZO / In2O3 nanowire MISM unit, integrates an active hierarchical virtual ground readout circuit at the array level and adopts an optoelectronic co-operated read / write strategy to suppress sneak-paths, improve read / write accuracy, and obtain accurate weighted readout in a high-density cross array.

[0079] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for optimizing photoelectric neuromorphic arrays based on indium aluminum zinc oxide, characterized in that, include: Step 1: Divide the photoelectric synaptic cross array into several sub-arrays, and each sub-array integrates a hierarchical virtual ground readout circuit; The weight gradient of each unit is calculated based on the neural network gradient, and the units whose conductance needs to be adjusted are selected as update targets and mapped to a list of operation instructions for the corresponding physical devices. Step 2: Simultaneously apply light pulses to the target position using a programmable light source for parallel coarse adjustment, combined with an adaptive auxiliary bias voltage; Step 3: In one training iteration of the neural network, the algorithm calculates a target weight for each weight according to the gradient descent rule of the array. A global weight-conductivity mapping function is preset, and each target weight calculated by the algorithm is uniquely converted into its corresponding physical target conductivity value through the global weight-conductivity mapping function. Thus, the target conductivity value of each updated target can be obtained, and a loop verification operation is performed on each updated target based on the target conductivity value of the updated target until the conductivity value reaches the target conductivity value or the maximum number of iterations is reached.

2. The method for optimizing an indium aluminum zinc oxide (IZO) photoelectric neuromorphic array according to claim 1, characterized in that, The list of operation instructions includes global physical coordinates (i, j), the original weight gradient ∇W(m, n) before mapping, and the relative update intensity scale α(i, j), where i is the row index of the update target in the subarray whose conductance needs to be adjusted, j is the column index of the update target in the subarray whose conductance needs to be adjusted, m is the row index of the entire photoelectric synaptic cross array, and n is the column index of the entire photoelectric synaptic cross array.

3. The method for optimizing an indium aluminum zinc oxide (IZO) photoelectric neuromorphic array according to claim 2, characterized in that, The hierarchical virtual ground readout circuit includes transimpedance amplifier channels, each channel corresponding to one or more columns of multiplexing, and the input terminal is servoed to near zero potential during the readout operation.

4. The method for optimizing an indium aluminum zinc oxide (IZO) photoelectric neuromorphic array according to claim 3, characterized in that, The process of selecting units whose conductance needs adjustment as update targets is as follows: In one training iteration of a neural network, the weight gradient matrix ∇W of a specific fully connected layer is calculated using the backpropagation algorithm. Each element ∇W(m,n) of the weight gradient matrix ∇W represents the update direction and magnitude of the logical weights W(m,n) in the network. The gradient matrix is ​​filtered and grouped: a gradient threshold θ is set, and the weights are considered to need to be updated in this round only when |∇W(m,n)|>θ, and the corresponding physical units are marked as target positions; all target positions are divided into two categories according to their gradient signs: positive update group and negative update group; a differential pair encoding strategy is adopted to map the update requirements of the weights into conductance increase operations of the corresponding physical devices. After mapping, all update requirements from the algorithm are converted into conductance increase operations of the physical devices. Based on the global coordinates of the target location in the physical array, the space is further divided according to the subarray block to which it belongs, so that each target location is assigned to its corresponding subarray, thus obtaining the updated target in the subarray; for the updated target in the subarray, a structured list of operation instructions is generated.

5. The method for optimizing an indium aluminum zinc oxide (IZO) photoelectric neuromorphic array according to claim 4, characterized in that, Parallel coarse tuning includes: S2-1: Receive the update target and operation instruction list of the subarray; designate the row containing the update target as the relevant row and the column containing the update target as the relevant column, thus forming a relevant row set and a relevant column set; the row index of at least one update target in the subarray constitutes the relevant row set, and the column index of at least one update target in the subarray constitutes the relevant column set; all rows and columns in the subarray that do not contain the aforementioned relevant row set and relevant column set are designated as unrelated rows and unrelated columns of the subarray; extract electrical state parameters from the update target and its corresponding relevant row and relevant column, specifically including row load, column load, and the center conductance value of the subarray; S2-2: Based on the adaptive calculation of electrical state parameters, the auxiliary bias voltage is used to obtain the voltage compensation terms for updating the target related row i and related column j; S2-3: Perform optoelectronic coordinated timing control adjustment based on the voltage compensation terms of the relevant row i and relevant column j of the updated target to complete a parallel optical coarse adjustment operation.

6. The method for optimizing an indium aluminum zinc oxide (IZO) photoelectric neuromorphic array according to claim 5, characterized in that, Electrical condition parameter extraction includes: For each relevant row i, count the number of update targets on that relevant row, and denote it as the row load as Ni; similarly, for each relevant column j, count the number of update targets on that relevant column, and denote it as the column load as Nj; Record the current conductance value of all updated targets in the subarray as G(i,j), and then calculate the geometric mean of the current conductance values ​​of all updated targets in the subarray as the center conductance value of the subarray.

7. The method for optimizing an indium aluminum zinc oxide (IZO) photoelectric neuromorphic array according to claim 6, characterized in that, The calculation of adaptive auxiliary bias includes: Based on material properties and experimental calibration, a safety baseline value related to the center conductivity is set: baseline bias value V. base To compensate for the line voltage drop caused by the simultaneous conduction of multiple update targets in the same row or column, a voltage compensation term based on its load is calculated for each relevant row i and relevant column j; the calculation formula is: For each relevant row i: ; For each relevant column j: ; Where k is the compensation coefficient calibrated based on the wire resistance and the updated target on-resistance, Nmax(j) represents the maximum theoretical load number of the relevant row i, that is, the total number of columns of the subarray in the relevant row; Nmax(i) represents the maximum theoretical load number of the relevant column j, that is, the total number of rows of the subarray in the relevant column.

8. The method for optimizing an indium aluminum zinc oxide (IZO) photoelectric neuromorphic array according to claim 7, characterized in that, Optoelectronic coordinated timing control adjustment includes: For all relevant rows i in the subarray, apply their corresponding calculated voltage compensation term V. compensate (i) Apply the corresponding voltage compensation term V to all relevant columns j in the subarray. compensate (j); Simultaneously, the potentials of all unrelated rows and columns are actively clamped to a fixed reference ground potential. After the row and column bias voltages stabilize, the programmable light source integrated above the subarray is triggered to generate a two-dimensional light spot pattern corresponding to the coordinates (i, j) of the updated target. The light pulse parameters are used to illuminate only the updated target position, including wavelength, pulse width, and pulse intensity. The pulse intensity is spatially modulated according to the relative update intensity scale α(i, j) of the updated target, i.e., α(i, j) is positively correlated with the pulse intensity. After the light pulse ends, the row and column bias voltages are maintained for a fixed duration, and the row and column drivers are controlled to smoothly restore all line voltages to the standby level, completing one parallel optical coarse adjustment operation.

9. The method for optimizing an indium aluminum zinc oxide (IZO) photoelectric neuromorphic array according to claim 8, characterized in that, The loop validation operation includes: S3-1 uses a hierarchical virtual ground readout circuit to perform a high-fidelity readout of the updated target (i, j): a readout voltage is applied to the row containing the updated target, the column containing the updated target is connected to the virtual ground, and all non-target rows and columns are clamped to a balanced bias, i.e., balanced bias = readout voltage / 2; the current is obtained through an analog-to-digital converter, and combined with the known readout voltage and transimpedance gain, the conductance value of the updated target (i, j) after optical coarse adjustment is calculated and recorded, and the target conductance value G'(i, j) of the updated target is subtracted from it to obtain the initial error; the error tolerance δ and the maximum number of iterations are preset; S3-2, enter a cyclic iterative electrical pulse fine-tuning from n=1 to the maximum iteration number, where n is the index of the cyclic iteration number; S3-3, Weight Confirmation and Update: After completing the weight update of a complete network layer, a global verification read is initiated to update all successfully fine-tuned final conductance values ​​to the non-volatile hardware weight lookup table, thereby completing a complete update from algorithm requirements to hardware state in a closed loop.

10. The method for optimizing an indium aluminum zinc oxide (IZO) photoelectric neuromorphic array according to claim 9, characterized in that, Cyclic iterative electrical pulse fine-tuning includes: S3-2-1, determine the parameters of the current electrical pulse based on the sign and amplitude of the initial error. The electrical pulse parameters include pulse amplitude and pulse width. The pulse amplitude is set within the safe adjustable range of the device, and the polarity is determined by the sign of the initial error: if the initial error > 0, apply a positive pulse; if the initial error < 0, apply a negative pulse. The pulse width is adaptively adjusted according to the initial error, and the adaptive adjustment calculation formula is as follows: Where Δg(i,j) is the initial error of updating target (i,j), G'(i,j) is the target conductance value of the updated target, and t min and t max For the minimum pulse width and the maximum pulse width, β is the sensitivity coefficient. When applying an electrical pulse according to the electrical pulse parameters, the pulse amplitude is only applied to the row i where the updated target is located, the column j where the updated target is located is kept grounded, and a balanced bias is applied to all unrelated rows and columns. S3-2-2, after the pulse ends, wait for the stabilization time, and then use the hierarchical virtual ground reading circuit again to measure the new conductance value at the updated target. Subtract the new conductance value from the target conductance value of the updated target to obtain the new error. S3-2-3, if the absolute value of the new error is less than or equal to the error tolerance δ, the fine-tuning of the updated target is determined to be successful, the final conductance value is recorded, and the loop is exited; if the absolute value of the new error is greater than or equal to the error tolerance δ, the initial error is set to the new error, and the loop is entered again; if the accuracy is still not met after the maximum number of loops, the fine-tuning is determined to be a failure.

Images