Current polarization regulation and control method and system in organic spinning device

Abstract

The invention discloses a current polarization regulation and control method and system in an organic spinning device. The invention relates to the technical field of organic spintronics. Performing diffusion convolution recursion, including controlling coding input of the electric field E0 and the transverse size distribution TD, capturing non-uniformly distributed current through a diffusion convolution layer, and simultaneously distinguishing diffusion intensity along the direction of the electric field and the transverse direction; predicting a current polarization value and correcting an electric field by correcting the interface resistance and simulating a hysteresis effect by the GRU; according to the method, the transverse size distribution of the organic spinning device is accurately determined through the data engineering step, and non-uniform distribution of current on a transverse grid is captured in combination with a diffusion convolution layer. The influence of interface resistance in an actual device on electric field distribution is considered, the accuracy of electric field simulation is improved through the interface resistance correction module, and the precision of current polarization regulation and control is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of organic spin electronics, specifically to the technical direction of electric variable control of organic spin devices, and more particularly to a method and system for controlling current polarization in organic spin devices. Background Art

[0002] Compared with traditional inorganic semiconductors, the spin relaxation time in organic semiconductors is significantly prolonged (the experimentally measured spin diffusion length is up to 200 nm), which is conducive to the long-range transport of spin-polarized carriers; its carriers exist in the form of polarons (carrying 1 / 2 spin) and bipolarons (spin neutral), resulting in an essential difference in the spin transport mechanism from that of inorganic systems. As spin information carriers, the ratio of polarons directly affects the efficiency of current spin polarization, while bipolarons weaken the spin signal due to their spin neutrality; that is, as spin information carriers, the higher the ratio of polarons, the greater the current spin polarization rate; bipolarons do not carry spin, so they dilute the spin signal and reduce the polarization efficiency. Increasing the electric field strength will significantly increase the yield of polaron excited states (experiments show that at 2.6×10 5 V / cm electric field, the excited state polaron yield reaches 6%). The electric field directly increases the polaron ratio by exciting more polarons. At the same time, the external electric field regulates the carrier transport path in the semiconductor by controlling the drift velocity of electrons and holes (the velocity is proportional to the electric field strength). For example, the electric field can guide polarons to gather in specific areas while suppressing the formation of bipolarons, thereby optimizing the polaron distribution. When the electric field exceeds the threshold (2.7×10 5 V / cm), polarons disintegrate, and charges jump to the conduction band to form free electrons. Although this process may reduce the polaron ratio, the disintegration rate can be controlled by dynamic regulation of the electric field to maintain the stability of the polaron ratio. Therefore, regulating the polaron ratio by the electric field can significantly improve the current spin polarization. Experiments have shown that under matching conductivity conditions, the electric field increases the current polarization at the interface by several orders of magnitude. This shows that the electrical properties of organic semiconductors are sensitive to external fields, and regulating carrier transport paths and spin polarization behavior through electric fields has become a key strategy to improve device performance.

[0003] However, existing research has mostly focused on ferromagnetic / organic heterostructures, and systematic research on the coordinated regulation mechanism of electric field and polaron ratio is still insufficient. For example: (1) The literature "Ren, JF, Fu, JY, Liu, DS, Mei, LM, & Xie, SJ(2005). Spin-polarized current in a ferromagnetic / organic system. Journal ofApplied Physics, 98(7), 074503." and the literature "Mi, YL, Zhang, M., & Yan, H.(2008). Effect of electric-field on spin transport and spin current in anorganic semiconductor system. Physics Letters A, 372(42), 6434-6437." disclose methods for improving the current spin polarization rate. However, these methods rely on the magnetic susceptibility and conductivity of specific ferromagnetic materials, which hinders their applicability to specific organic spin devices. (2) Chinese invention patent CN118870958A (publication date 2024-10-29) discloses a spin amplification model, which mainly controls the new spin filtering effect between the spin interface layer and the bottom electrode by actively changing the molecular coverage, changing the spin-resolved state density at the interface between the bottom electrode and the spin interface layer, and realizing continuous control of the magnetoresistance signal and spin polarization at room temperature; however, it does not involve the influence of the electric field and the spin relationship of the carrier charge in the organic semiconductor. Its control method is actually an indirect response relationship, which has certain obstacles to the applicability of specific organic spin devices; (3) Literature: Dou Zhaotao, Ren Junfeng, Wang Yumei, et al. Study on the properties of current spin polarization amplification in organic devices [J]. Acta Physica Sinica, 2012, 61(08): 502-507. Literature: Csontos D, Ulloa S E. Spin polarization control by electric field gradients [J]. Physical Review Letters, 2006. Literature: Anonymous. Effect of electric-field on spin transport and spin current in anorganic semiconductor system [J]. Physics Letters A, 2008, 372(33): 5531-5535. Literature: Shi S, Sun Z, Bedoya-Pinto A, et al. Hybrid Interface States and Spin Polarization at Ferromagnetic Metal-Organic Heterojunctions: Interface Engineering for Efficient Spin Injection in Organic Spintronics [J]. Advanced Materials Interfaces, 2014, 1(5): 1400122.》, Document "Ma YN, Ren JF, Zhang YB, et al. Effect of electric field on spin polarized current in ferromagnetic / organic semiconductor systems[J]. Chinese Physics Letters, 2007, 24(6): 1697-1700.", Document "Lv Q, Fu PH, Yu XL, et al. Electrically controlled spinpolarized current in Dirac semimetals[J]. Scientific Reports, 2021, 11(1):21509.Both of them disclose current spin polarization amplification models, which control the electric field to achieve current polarization regulation. The main strategies can be divided into two categories: one is to use a two-component drift-diffusion model to control the electric field strength by controlling the position of the conductivity; the other is to use chemical interaction methods (electrical parameters such as gate voltage, chemical potential, and coupling strength) to control the spin polarization at the hybrid interface of ferromagnetic metals and organic semiconductors (or Dirac semimetals). However, there are two technical drawbacks: 1) Lateral Size Effects: In the aforementioned prior art documents, the lateral width of the devices used is much smaller than the spin diffusion length (considered a one-dimensional structure). In practice, many devices exhibit two-dimensional / three-dimensional current distributions, and the competing effects of the lateral electric field or carrier diffusion are not considered. For example, when the lateral dimensions are large, the non-uniformity of the electric field distribution may alter the E1 / E0 ratio, thereby affecting the actual distribution of J1. (J1 is the main branch current, flowing along the dominant electric field direction or the main current channel of the device design. Its spin-polarized characteristics directly affect device function. J0 is the secondary branch current, generated by competing mechanisms such as the lateral electric field, carrier diffusion, or interface effects. Its path deviates from the main current direction and may introduce additional spin relaxation or depolarization effects. The ratio of J1 to J0 can be manipulated by applying an external electric field gradient, such as through a non-uniform gate design. For example, in regions with a large lateral electric field component, the J0 ratio increases, and the negative effects of J0 need to be suppressed by optimizing the electrode geometry (e.g., using a focusing electrode).

[0004] It should also be noted that the main branch current J1 flows along the main electric field direction or optimized transport path, while the secondary branch current J0 is driven by the transverse electric field, material defects or interface scattering, and has a more random path. The main branch current J1 has a high spin polarization efficiency (the dominant device function), while the secondary branch current J0 is low (polarization loss due to diffusion or relaxation). The main branch current J1 increases linearly with the external electric field strength. The secondary branch current J0 is significant at low electric fields and is suppressed at high electric fields due to the enhancement of the dominant current J1. 2) Ignoring interface resistance: The aforementioned prior art documents assume zero interface resistance at the device interface and a continuous electrochemical potential. In reality, many device interfaces have potential barriers or scattering, causing current distribution to deviate from theoretical values. For example, interface resistance introduces additional voltage drop, causing the actual distribution of E1 and E2 to differ from the theoretical formula E0 = E1 + E2, thereby affecting the J1 / J0 ratio.

[0005] To this end, the present invention proposes a current polarization control method and system in an organic spin device. Summary of the Invention

[0006] In view of this, the present invention aims to provide a method and system for controlling current polarization in an organic spin device to solve or alleviate the technical problems existing in the prior art, namely, how to overcome the defects in the prior art of controlling current polarization in organic spin devices, which result in errors due to ignoring lateral size effects and interface resistance. Furthermore, based on solving the above technical problems, the present invention aims to provide a method and system for controlling current polarization by estimating the electric field to be controlled. The technical solution of the present invention is achieved as follows: First, methods for controlling current polarization in organic spin devices: (I) Overview: The present invention aims to capture the non-uniform distribution of current on the lateral grid by accurately determining the lateral size distribution of the device and controlling the electric field, combining diffusion convolution recursion and GRU units (Gated Recurrent Units), and introducing anisotropic diffusion coefficients to distinguish the diffusion intensity in different directions. At the same time, the influence of interface resistance on the electric field distribution is considered, and the accuracy of electric field simulation is improved through an interface resistance correction module. Based on the predicted current polarization value and the corrected control electric field, this solution can dynamically adjust the electric field to achieve efficient current polarization control. In addition, the nonlinear relationship between the electric field and current polarization is considered, and overshoot is avoided through learning rate decay.

[0007] (2) Technical solution: To achieve the above objectives, the present invention selects to perform the following steps: 2.1 Step S1, Data Engineering: Determine the lateral size distribution TD=[G,L G ], where G is a two-dimensional grid, L G is a transverse geometry, and L G =(B w ,I s ), where B w is the branch width, I s is the interface shape; and determines the control electric field E0 (scalar) and the ideal current polarization value α ideal (scalar). Where: ; Among them, the two-dimensional grid G ​​is represented by the discrete point coordinates (iΔ x ,jΔ y ) indicates that the step size Δ x , Δ y Need to be less than the spin diffusion length L s ; r is the curvature radius of the device; polarization value α ideal The value range is [0,1].

[0008] 2.2 Step S2, perform diffusion convolution recursion: It includes the encoding input of controlling the electric field E0 and the lateral size distribution TD and capturing the non-uniformly distributed current through the diffusion convolution layer, while distinguishing the diffusion intensity along the electric field direction and the lateral (i.e. vertical) direction; by correcting the interface resistance and simulating the hysteresis effect of the GRU unit, the current polarization value is predicted and the electric field is corrected.

[0009] 2.2.1 Step S200, input layer: The control electric field E0 and the lateral size distribution TD are encoded into the independent variable X and input into the diffusion convolution layer.

[0010] ; Encode(·) is the encoding function that maps input parameters into tensors that the neural network can process; FC(·) is a fully connected layer used for dimensionality increase or decrease; ⊕ is a tensor concatenation operation (along the channel dimension). X is the encoded input tensor, which serves as the input to the dilated convolution layer. R is the number of channels.

[0011] Among them, the fully connected layer ; The activation function σ is the Sigmoid activation function. There are five dimensions in total, represented by G∈R m×n×2 (grid coordinates), L G ∈R m ×n×2 (B w and I s Broadcast to grid size), E0∈R m×n×1 (Scalar broadcast) splicing. The 1×1 convolution kernel keeps the spatial dimension (m×n) unchanged and only adjusts the number of channels. It is equivalent to the fully connected layer operating independently on each position. The output is used as the input tensor of the diffusion convolution layer, with a dimension of m×n×c out , c out is the design parameter (such as 32, 64, etc.); 2.2.2 Step S201, Diffusion Convolution Layer: Use a learnable diffusion kernel (convolution kernel weight) to capture the non-uniform distribution of current on the lateral grid; Anisotropic diffusion coefficient is introduced to distinguish the diffusion intensity along the electric field direction and the horizontal direction; Correct the effect of interface resistance on electric field distribution based on the interface resistance simulation module, including resistance parameter layer and electric field redistribution; The GRU (Ground Recurrent Unit) models the temporal and spatial dependence of current polarization on the dynamics of the electric field. As a special type of recurrent neural network (RNN) unit, the GRU can process time series data and capture the hysteresis effect of the current polarization state evolving over time. In combination with the historical state (hidden layer), it captures the hysteresis effect of the spin diffusion length as the electric field changes. 2.2.2.1 Step S2010, Diffusion Convolution Recursion: ; Among them, K diff is the learnable diffusion kernel; k is the convolution kernel size, c in / c out is the number of input / output channels; J diff is the current distribution tensor after diffusion convolution, capturing the lateral non-uniformity.

[0012] Conv2D(·) is a two-dimensional convolution kernel: ; Where X∈R m×n×cin , from step S200, c in That is, the number of dimensions; convolution kernel K diff ∈R k×k×cin×cout , are all learnable parameters, k=3, c out =32; step size s=1 keeps the spatial resolution unchanged; Among them, filling , if k=1 then P=1, maintaining the output size; 2.2.2.2 Step S2011, introduce anisotropic diffusion coefficient: ; Where θ is the angle between the diffusion direction and the electric field direction (the main path direction of the spin-polarized current propagating inside the material); D(θ) is the anisotropic diffusion coefficient, along the electric field direction D ∥ and horizontal D ⊥ The diffusion intensity is different, which is consistent with the electric field-related spin diffusion length model demonstrated in the literature "Dou Zhaotao, Ren Junfeng, Wang Yumei, et al. Study on the current spin polarization amplification properties of organic devices [J]. Acta Physica Sinica, 2012, 61(08): 502-507."

[0013] 2.2.2.3 Step S2012, interface resistance correction: ; ; ; Among them, R interface is the interface resistance value, which is learned from the input parameters by the fully connected layer FC(·). E1′, E2′ are the corrected branch electric fields, considering the effect of the interface resistance on the total resistance R total The impact of R OSCi is the branch resistance; 2.2.2.4 Step S2013, GRU unit modeling hysteresis effect: The architecture composed of all GRU units: ; ; ; ; Among them, z t and r t are the update gate and reset gate; h t is the implicit state at time t, representing the encoded current polarization value; is the candidate hidden state; σ is the sigmoid activation function; W z and W r is the weight matrix in the reset gate, used to calculate the activation value of the reset gate; h t-1 is the implicit state of the previous time step; tanh(·) is the hyperbolic tangent function; ⊙ is the Hadamard product.

[0014] 2.2.3 Step S202, output layer: For the hidden state h t Perform decoding operation and output predicted current polarization value α pred And the corrected controlled electric field E0'; then the physical constraint is injected: 1) Physical consistency loss: the control electric field E0' is forced to satisfy the current continuity equation; 2) Interface resistance constraint: self-consistent with the interface resistance simulation module.

[0015] The above decoding and physical constraint methods are: ; ; Among them, the constraint term requires that the current continuity equation J0=J1+J2 be satisfied (see formula 3 in the literature "Dou Zhaotao, Ren Junfeng, Wang Yumei, et al. Study on the current spin polarization amplification properties of organic devices [J]. Acta Physica Sinica, 2012, 61(08): 502-507."); σ OSCi is the conductivity of the i-th branch (inherited from the theoretical model of this paper). At the same time, it is also required to satisfy the interface resistance constraint to ensure that the corrected electric field E0′ is consistent with the interface resistance R interface Satisfies Ohm's law.

[0016] Among them, FC α It is a fully connected layer that transforms the hidden state h t Mapped to a polarizability scalar (sigmoid compression to [0,1]). FC E It is another fully connected layer that generates the electric field correction ΔE0, which is superimposed on the original E0 to obtain the corrected control electric field E0′.

[0017] 2.3 Step S3, perform electric field control: Corrected control electric field E0′ and predicted current polarization value α captured and decoded by the diffusion convolution layer pred , determine whether the current distribution in the wide branch of the current organic spin device is dispersed and whether the polarization rate needs to be reduced; and control the electric field, thereby controlling the current polarization behavior of the organic spin device.

[0018] The method is to drive the circuit in the high voltage area (E0>E c ), adjust the corrected control electric field E0' according to the direction of the error to offset the nonlinear effect. The process is: Dispersion calculation: J based on the output of the diffusion convolution layer diff , calculate the current distribution standard deviation σ J .

[0019] Polarizability adjustment decision: If σ J >σ th , it is necessary to reduce the polarizability by lowering E0′.

[0020] High pressure area adjustment: when E0′>E c When the error direction sign(α ideal −α pred ) Dynamically correct the electric field.

[0021] Learning rate decay: The learning rate η in the high-voltage region decays exponentially as E0′ increases to avoid overshoot.

[0022] (3) Mechanism for resolving technical issues: 3.1 Solve the lateral size effect: First, through data engineering, the lateral size distribution of the organic spintronic device is determined, including detailed parameters such as the 2D grid and lateral geometry. A diffused convolutional layer is used to capture the non-uniform current distribution across the lateral grid. This step, leveraging the learning capabilities of the neural network, simulates the current distribution variations caused by lateral size effects in actual devices.

[0023] Anisotropic diffusion coefficients are introduced to differentiate the diffusion strength along the electric field direction from that in the transverse direction. This aligns with the characteristic that the spin diffusion length in organic semiconductors varies with the electric field direction, thereby more accurately simulating the physical processes in actual devices. GRUs (gated recurrent units) are used to model the temporal and spatial dependence of current dynamics on current polarization, capturing the hysteresis effect of the spin diffusion length as it changes with the electric field. This allows for the prediction and compensation of current distribution variations caused by lateral size effects.

[0024] 3.2 Solve the interface resistance problem: Considering the influence of interface resistance on electric field distribution in actual devices, the interface resistance value is learned through the fully connected layer, and the branch electric field is corrected according to this value, thereby improving the accuracy of electric field simulation.

[0025] At the output layer, physical constraints are enforced on the predicted current polarization values ​​and the corrected control electric field, including enforcing the current continuity equation and interface resistance constraints. This ensures that the predicted results are consistent with physical laws, further reducing errors.

[0026] 3.3 Estimate and control the electric field and perform the adjustment operation: Based on the corrected control electric field and predicted current polarization value captured and decoded by the diffusion convolution layer, the present invention calculates the dispersion of the current current distribution. If the dispersion exceeds the preset threshold, it is determined that the electric field needs to be adjusted to reduce the polarization rate. By introducing a dynamic learning rate and a Heaviside step function, in the high-voltage area (that is, the electric field strength exceeds the critical value), the scheme dynamically corrects the electric field according to the direction of the polarization rate error to ensure that the adjustment amount conforms to the laws of physics and can effectively reduce the error. At the same time, the nonlinear relationship between the electric field and the current polarization is also taken into account, and over-adjustment is avoided in the high-voltage area by attenuating the learning rate. This helps to ensure the stability and accuracy of the control process.

[0027] The second aspect is the current polarization control system in organic spin devices, including: (1) A data acquisition and processing module for collecting the physical parameters of organic spin devices, including a sensor array, an analog-to-digital converter (ADC), and a microcontroller (MCU).

[0028] The sensor array is responsible for collecting real-time data from the organic spin device, including electric field strength and current distribution. After obtaining the original signal, the analog-to-digital converter converts the analog signal into a digital signal. The microcontroller performs data processing, filtering, and format conversion, and passes it to the computing module.

[0029] (2) A calculation module for executing the method of steps S1 to S3, comprising: (2.1) Data encoding module responsible for encoding input data into tensor form that can be processed by the neural network; (2.2) Diffusion convolution recursive module responsible for executing the neural network algorithm: Includes a diffusion convolution module to capture the non-uniform distribution of current on the lateral grid; Anisotropic diffusion coefficient module, used to distinguish the diffusion intensity in different directions; Interface resistance correction module, used to correct the influence of interface resistance on electric field distribution; GRU module, used to model the temporal / spatial dependence of current dynamics on current polarization; (2.3) Decode the implicit state and output the predicted current polarization value and the corrected control electric field. Output decoding and physical constraint module: Apply physical consistency loss and interface resistance constraints to ensure the consistency of the prediction results with physical laws.

[0030] (2.4) The electric field control and adjustment module calculates the dispersion of the current distribution based on the predicted current polarization value and the corrected control electric field, and determines whether the electric field needs to be adjusted. In the high-voltage region, the electric field is dynamically corrected according to the direction of the polarization error, and nonlinear compensation and learning rate decay are considered to form control instructions. (3) Organic spin device control module: Including digital-to-analog converter (DAC), power supply and drive circuit (high voltage drive circuit, pulse drive circuit or programmable drive circuit); The digital-to-analog converter receives control instructions from the electric field control and adjustment module, calculates the control parameters, and then converts them into analog signals to the power supply controller. The power supply provides the required voltage output to the drive circuit, which in turn applies the corresponding electric field to the organic spin device.

[0031] Compared with the prior art, the present invention has the following beneficial effects: 1. Improving the accuracy of current polarization control: In the present invention, by considering the lateral size distribution in combination with the diffuse convolution layer, it is possible to obtain the non-uniform distribution of the captured current on the lateral grid, combined with the influence of the interface resistance on the electric field distribution in the actual device, and improve the accuracy of the electric field simulation through the interface resistance correction module, thereby further improving the precision of current polarization control.

[0032] Second, compensating for errors caused by lateral size effects and interface resistance: This invention introduces an anisotropic diffusion coefficient to differentiate the diffusion intensity along the electric field direction from the lateral direction, compensating for errors in current distribution changes caused by lateral size effects. Using GRU units to model the temporal and spatial dependence of current dynamics on current polarization, it captures the hysteresis effect of spin diffusion length changes with electric field, further reducing error accumulation.

[0033] 3. Achieving Efficient Electric Field Control: In this invention, based on the predicted current polarization value and the corrected control electric field, the dispersion of the current distribution is calculated to determine whether the electric field needs to be adjusted. In the high-voltage region, the electric field is dynamically corrected according to the direction of the polarization error, achieving efficient electric field control. Furthermore, the nonlinear relationship between the electric field and current polarization is taken into account. Overshoot is avoided in the high-voltage region through learning rate decay, ensuring the stability and accuracy of the control process.

[0034] 4. Improving the performance of organic spin devices: In the present invention, more precise current polarization control can optimize the current transport path in organic spin devices, reduce unnecessary energy loss, improve device efficiency, and enhance spin signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0036] Figure 1 Schematic diagram of the method flow of the present invention; Figure 2 Schematic diagram of the lateral branch of a T-type device; Figure 3 Schematic diagram of the relationship between the amplification rate and the current ratio of the T-type device; Figure 4 Schematic diagram of the neural network architecture of the present invention; Figure 5 Schematic diagram of the GRU unit structure of the present invention; Figure 6 Schematic diagram of the output layer architecture of the present invention; Figure 7 Schematic diagram of the convolution operation of the present invention; Figure 8 Schematic diagram of the system composition of the present invention. DETAILED DESCRIPTION

[0037] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. The following description sets forth many specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. The devices disclosed in the embodiments are described briefly because they correspond to the methods disclosed in the embodiments. For relevant details, refer to the method description.

[0038] Explanation of relevant terms: (1) Transverse size distribution TD: a set of geometric parameters describing the lateral (perpendicular to the electric field) direction of an organic spin device. TD = [G, L G ], G is the discretized representation of the device's lateral structure, used for numerical calculation of current distribution (e.g., 10×10 grid); the lateral geometric structure L G Including branch width (B w ) and interface shape (I s ), which is used to quantify the physical morphology of the device.

[0039] (2) Branch width B w : The lateral physical width of the branches (OSC0, OSC1, OSC2) in a T-type device. w Large) may cause current dispersion and reduce the polarizability ( Figure 2 ).

[0040] (3) Interface shape I s : The geometric characteristics of the interface between the electrode and the organic semiconductor, curved surface or flat surface.

[0041] (4) Controlled electric field E0: The initial electric field strength input to the device injection terminal (OSC0) (unit: V / cm). This controls the carrier transport direction and spin polarization amplification.

[0042] (5) Ideal current polarization value α ideal : The desired current spin polarization.

[0043] (6) Encoding: Convert the input parameters (E0, Bw, Is) into the tensor format of the neural network. For example, E0 → scalar, then Bw and Is → channel features of a two-dimensional grid.

[0044] (7) Non-uniform distribution: Spatial non-uniformity of current on the lateral grid. Resistance changes caused by differences in branch width and interface shape.

[0045] (8) Diffusion intensity in the electric field direction and transverse direction, including: Anisotropic diffusion (along the electric field direction): Diffusion is accelerated by the electric field, and the intensity is determined by D ∥ =D0e E0 / Ec Decide; Lateral diffusion: intensity is fixed at D ⊥ =D0.

[0046] (9) Polarizability: the ratio of spin polarized current to total current.

[0047] (10) Hysteresis effect: After the electric field changes, spin polarization takes a certain amount of time (spin relaxation time τ s ) to reach a steady state. The GRU unit memorizes the historical state and captures the temporal dependency.

[0048] (11) The influence of interface resistance on electric field distribution: changes the actual distribution of branch electric fields.

[0049] (12) Electric field redistribution: The electric fields E1 and E2 of each branch are dynamically adjusted according to the interface resistance in order to keep the total voltage consistent.

[0050] Example 1: Please refer to Figure 1 This embodiment discloses a method for controlling current polarization in an organic spin device, including the following steps.

[0051] In this embodiment, regarding step S1, data engineering: determine the lateral size distribution TD of the current organic spin device = [G, L G ], where G is a two-dimensional grid, L G is a transverse geometry, and L G =(B w ,I s ), where B w is the branch width, I s is the interface shape; and determines the control electric field E0 (scalar) and the ideal current polarization value α ideal (scalar). ; Among them, the two-dimensional grid G ​​is represented by the discrete point coordinates (iΔ x ,jΔ y ) indicates that the step size Δ x , Δ y Need to be less than the spin diffusion length L s ( L s =200nm); r is the radius of curvature; polarization value α ideal The value range is [0,1].

[0052] It should be pointed out that: (1) Two-dimensional grid G: represented by discrete point coordinates (iΔx, jΔy), where i and j are integer indices, and Δx and Δy are the step lengths in the x and y directions, respectively. The step length must be less than the spin diffusion length L. s This ensures that the grid is fine enough to describe the lateral size variations of the device. This allows the electrical and spin properties of each point on the grid to be calculated, thereby obtaining the lateral size distribution of the entire device.

[0053] (2) In the lateral geometry LG: branch width B w determines the width of the current channel in the device, while the interface shape I s This affects the contact resistance between the device and the electrode and the spin injection efficiency.

[0054] In this embodiment, regarding step S2, the diffusion convolution recursion is performed, see Figures 4-6 Due to the lateral size effect and interface resistance of the device, the current distribution is often non-uniform. The diffusion convolution layer processes the input electric field and size distribution data through convolution operations combined with a physical diffusion model to capture the non-uniform distribution characteristics of the current.

[0055] The specific approach is to control the encoding input of the electric field E0 and the lateral size distribution TD and capture the non-uniformly distributed current through the diffusion convolution layer, while distinguishing the diffusion intensity along the electric field direction and the lateral direction; by correcting the interface resistance and simulating the hysteresis effect of the GRU unit, the current polarization value is predicted and the electric field is corrected.

[0056] For details, please refer to Figure 4 Step S200: The input layer of the neural network encodes the control electric field E0 and the lateral size distribution TD into the independent variable X and inputs it into the diffusion convolution layer. The purpose of this step is to convert the parameters of the physical world into a numerical form that the computer can understand and process: ; Encode(·) is the encoding function that maps input parameters into tensors that the neural network can process; FC(·) is a fully connected layer used for dimensionality increase or decrease; ⊕ is a tensor concatenation operation (along the channel dimension). X is the encoded input tensor, which serves as the input to the dilated convolution layer. R is the number of channels.

[0057] It should be pointed out that: (1) The encoding function Encode(·) is a custom function that uses tensor concatenation and fully connected layer processing to map E0 and TD into tensors that can be processed by the neural network.

[0058] (2) Tensor splicing operation ⊕: the two-dimensional grid G ​​and the horizontal geometric structure L G (Including branch width B w and interface shape I s ) and control the electric field E0 to be spliced ​​along the channel dimension. The advantage of this is that the spatial information and characteristics of each parameter are preserved, while facilitating subsequent fully connected layer processing.

[0059] (3) Fully connected layer FC(·): used to increase or reduce the dimension of the concatenated tensor. It actually plays a role in feature extraction and conversion, converting the original physical parameters into feature representations that are more suitable for neural network processing. ; Among them, the activation function σ is the Sigmoid activation function. There are five dimensions in total, represented by G∈R m×n×2(grid coordinates), L G ∈R m×n×2 (B w and I s Broadcast to grid size), E0∈R m×n×1 (Scalar broadcast) splicing. The 1×1 convolution kernel keeps the spatial dimension (m×n) unchanged and only adjusts the number of channels. It is equivalent to the fully connected layer operating independently on each position. The output is used as the input tensor of the diffusion convolution layer, with a dimension of m×n×c out , c out is the design parameter (such as 32, 64, etc.); this output tensor will be used as the input of the dilated convolution layer.

[0060] (4) G∈R m×n×2 : A tensor representation of a two-dimensional grid G, containing the coordinate information of the grid points.

[0061] (5) L G =[B w ⊕I s ]∈R 2 (But here, in order to concatenate with G and E0, it actually needs to be broadcasted to the grid size, i.e. L G ∈R m×n×2 ): Transverse geometry L G The tensor representation of the branch width B w and interface shape I s E0∈R (also needs to be broadcast to the grid size, that is, E0∈R m×n×1 ): scalar value controlling the electric field, broadcast to the grid size to match G and L G Splicing.

[0062] It can be understood that through tensor splicing and full connection layer processing, the control electric field E0 and the lateral size distribution TD are converted into feature representations that can be processed by the neural network. This representation method not only retains the spatial information and characteristics of the original parameters, but also facilitates learning and processing by the neural network. In addition, the introduction of the activation function σ increases the nonlinear factors of the network, enabling the model to learn more complex mapping relationships. At the same time, the fully connected layer extracts and transforms the features, and the dimension of the output tensor X is m×n×c out , which is directly applicable to the input of the diffusion convolution layer. The diffusion convolution layer can perform convolution operations on this basis to capture the non-uniform distribution characteristics of the current and provide strong support for subsequent current polarization control.

[0063] It should be further pointed out that the features extracted and transformed by the fully connected layer refer to the fusion of the control electric field E0 and the lateral size distribution T through the tensor splicing operation. D The comprehensive characteristic tensor X is formed; X includes the control electric field E0 and the lateral size distribution T DAll the information is the neural network's "understanding" of the physical parameters. Through tensor splicing, physical parameters of different dimensions (scalar E0, two-dimensional grid G, geometric structure L) are combined into a G ) are unified into tensors of the same dimension, making them easier to process by the neural network. This process maps physical parameters to neural network inputs, providing the basis for subsequent current polarization control.

[0064] For details, please refer to Figure 4 Step S201: The diffuse convolution layer of the neural network: In organic spintronic devices, the current distribution is often affected by the device's lateral size, resulting in a non-uniform current distribution across the lateral grid. To capture this non-uniform distribution, the diffuse convolution layer in this step employs a learnable diffusion kernel (i.e., convolution kernel weights). These weights are automatically adjusted during training based on the input data and target output to optimally describe the current distribution across the lateral grid.

[0065] The specific operations are: using a learnable diffusion kernel (convolution kernel weight) to capture the non-uniform distribution of current on the lateral grid; introducing anisotropic diffusion coefficient to distinguish the diffusion intensity along the electric field direction and the lateral direction; correcting the influence of interface resistance on electric field distribution based on the interface resistance simulation module, including resistance parameter layer and electric field redistribution; recurrent unit (GRU) modeling the time / spatial dependence of current dynamic changes on current polarization; combining historical states (hidden layer) to capture the hysteresis effect of spin diffusion length changing with electric field changes.

[0066] Specifically, step S2010, diffusion convolution recursion: In order to capture this non-uniformity, a learnable diffusion kernel K is introduced diff Perform a diffusion convolution operation. Diffusion convolution can extract the local features of the current distribution and generate a new feature representation J by sliding the convolution kernel on the input tensor X. diff , that is, the current distribution tensor after diffusion convolution: ; Among them, k is the convolution kernel size, c in / c out is the number of input / output channels; J diff is the current distribution tensor after the diffusion convolution, which captures the non-uniformity of the current on the lateral grid. Conv2D(·) is the two-dimensional convolution kernel; Furthermore, the two-dimensional convolution operation (Conv2D) is the core step of the diffuse convolution. It generates a new feature map by sliding the convolution kernel over the input tensor and performing dot product and sum operations at each position. In this process, the weights of the convolution kernel are learnable and are automatically adjusted according to the target output during training to optimally describe the non-uniformity of the current distribution: ; Where X∈R m×n×cin , from step S200 (m and n are the grid sizes), c in That is, the number of dimensions; convolution kernel K diff ∈R k ×k×cin×cout , are all learnable parameters, k=3, c out =32; step size s=1 keeps the spatial resolution unchanged; Among them, filling , if k=1 then P=1, maintaining the output size; like Figure 7 As shown, for each position (i, j) in the input tensor X, the convolution operation is: ; Among them, p and q are the offsets of the convolution kernel, c in and c out are the number of input and output channels, respectively.

[0067] It is understandable that by performing diffuse convolution operations using a learnable diffusion kernel, the model can automatically learn the non-uniform distribution characteristics of current on the lateral grid. This self-learning ability enables the model to more accurately capture the details and changes in the current distribution, thereby improving the accuracy of the current distribution simulation. Because the weights of the diffusion kernel are learnable, the model can automatically adjust according to different input data and target outputs during training. This flexibility enables the model to generalize to different device structures and experimental conditions, improving the model's adaptability and robustness. At the same time, the diffuse convolution operation can maintain the spatial resolution of the input tensor. While capturing the current distribution characteristics, the model can preserve the integrity of spatial information. Moreover, the convolution operation itself has the ability to extract multi-scale features. By stacking multiple convolutional layers or using convolution kernels of different sizes, the model can capture the characteristics of the current distribution at different scales, thereby further improving the model's expressive power. In this step, even if a single-scale convolution kernel is used, multi-scale features can be indirectly extracted through processing in subsequent layers.

[0068] Specifically, step S2011 introduces the anisotropic diffusion coefficient. In organic spintronic devices, the diffusion behavior of current under an electric field often exhibits anisotropy, meaning that the diffusion intensity differs along the electric field direction and transverse direction. To more accurately describe this anisotropic diffusion behavior, this step introduces the anisotropic diffusion coefficient D(θ). This coefficient is defined based on the angle θ between the diffusion direction and the electric field direction, allowing the diffusion intensity along the electric field direction and transverse direction to be set separately.

[0069] ; Where θ is the angle between the diffusion direction and the electric field direction; D(θ) is the anisotropic diffusion coefficient, along the electric field direction D ∥ and horizontal D ⊥ The diffusion intensities are different, which is consistent with the electric field-dependent spin diffusion length model demonstrated in the literature "Dou Zhaotao, Ren Junfeng, Wang Yumei, et al. Study on the current spin polarization amplification properties of organic devices [J]. Acta Physica Sinica, 2012, 61(08): 502-507."

[0070] It is understandable that by introducing the anisotropic diffusion coefficient, the model can more accurately describe the anisotropic diffusion behavior of current under the action of an electric field. This accuracy is crucial for predicting the current distribution and polarization characteristics in organic spin devices because it can capture the subtle effects of the electric field on the diffusion of spin current. The diffusion coefficient D along the electric field direction is ∥ The exponential increase in the electric field strength is consistent with the electric field-dependent spin diffusion length model. This setting allows the model to reflect the electric field's facilitation of spin current diffusion, thereby more accurately simulating the current polarization control process in organic spin devices. By introducing anisotropic diffusion coefficients, the model further conforms to the physical reality of organic spin devices.

[0071] Specifically, in step S2012, interface resistance correction: In organic spin devices, interface resistance is one of the important factors affecting current distribution and polarization characteristics. In order to more accurately simulate the influence of interface resistance, an interface resistance correction step is introduced. In this step, the interface resistance value R interface It is obtained through the fully connected layer from the input parameters (initial electric field strength E0, device width related parameters B w and the spin injection current I s ) can be learned.

[0072] At the same time, in organic spin devices, current is usually split through different organic semiconductor layers (such as OSC1 and OSC2), and the interface resistance will affect the distribution of these branched electric fields. To more accurately describe this effect, the interface resistance and branch resistance (R OSC1 and R OSC2 ) to calculate the corrected branch electric fields (E1′ and E2′): ; ; ; Among them, R interface is the interface resistance value, which is learned from the input parameters by the fully connected layer FC(·). E1′, E2′ are the corrected branch electric fields, considering the effect of the interface resistance on the total resistance R total The impact of R OSCiis the branch resistance; It is understandable that by introducing the interface resistance correction step, the model can more accurately consider the impact of interface resistance on current distribution and capture the subtle effects of interface resistance on current shunting. The interface resistance correction step makes the model more consistent with the physical reality of organic spin devices. In actual situations, interface resistance does affect the shunting and polarization characteristics of current, so by introducing this step, the model is more physically reasonable and reliable. Moreover, the interface resistance value is learned from the input parameters through the fully connected layer, which allows the model to more flexibly adjust the interface resistance value during training to optimize performance. By adjusting these parameters, the model can better adapt to different device structures and experimental conditions, thereby improving its generalization ability and robustness.

[0073] Specifically, in step S2013, the GRU unit models the hysteresis effect: In the current polarization control of organic spin devices, the current distribution and polarization state are often affected by the state at the previous moment. This effect is called the hysteresis effect. In order to capture this hysteresis effect, the gated recurrent unit models the time series changes of the current polarization state. GRU is a special recurrent neural network (RNN) unit that controls the flow and forgetting of information by introducing update gates and reset gates, thereby effectively capturing long-term dependencies in time series. The architecture composed of all GRU units is: ; ; ; ; Among them, z t and r t are the update gate and reset gate; h t is the implicit state at time t, representing the encoded current polarization value; is the candidate hidden state; σ is the sigmoid activation function; W z and W r is the weight matrix in the reset gate, used to calculate the activation value of the reset gate; h t-1 is the implicit state of the previous time step; tanh(·) is the hyperbolic tangent function; ⊙ is the Hadamard product.

[0074] The update gate determines the implicit state h at the current moment t How much information comes from the hidden state h at the previous moment t-1 , and how much information comes from the candidate hidden state at the current moment .

[0075] The reset gate determines the candidate hidden state at the current moment When calculating, how much information comes from the implicit state h at the previous moment? t-1 When the reset gate is close to 0, the candidate hidden state is almost independent of the previous hidden state h t-1 , thus achieving the forgetting of information.

[0076] The candidate hidden state is a candidate value of the hidden state at the current moment, which is based on the hidden state at the previous moment after the reset gate adjustment and the input at the current moment (here is the current distribution tensor J after the diffusion convolution diff ) is calculated.

[0077] The hidden state is the final output at the current moment, which is calculated based on the hidden state at the previous moment and the candidate hidden state after adjustment by the update gate.

[0078] It is understandable that by introducing the update gate and reset gate, the GRU unit can effectively capture the time series changes of the current polarization state, especially the hysteresis effect. The update gate controls the flow of information, allowing the model to selectively retain or forget the state information of the previous moment; while the reset gate determines how much information the candidate hidden state of the current moment uses from the hidden state h at the previous moment when calculating it. t−1 , thus enabling modeling of hysteresis effects. The GRU unit, through its unique gating mechanism, is able to effectively handle long-term dependencies in time series data. This enables the model to more accurately consider state information from previous moments when predicting the current polarization state in organic spin devices, thereby improving the accuracy and stability of the prediction.

[0079] For details, please refer to Figure 6 , step S202, the output layer of the neural network: for the hidden state h t Perform decoding operation and output predicted current polarization value α pred And the corrected controlled electric field E0'; then the physical constraint is injected: 1) Physical consistency loss: the control electric field E0' is forced to satisfy the current continuity equation; 2) Interface resistance constraint: self-consistent with the interface resistance simulation module.

[0080] The above decoding and physical constraint methods are: ; ; Among them, the constraint term requires that the current continuity equation J0=J1+J2 be satisfied (see formula 3 in the literature "Dou Zhaotao, Ren Junfeng, Wang Yumei, et al. Study on the current spin polarization amplification properties of organic devices [J]. Acta Physica Sinica, 2012, 61(08): 502-507."), J0 is the total current, J1 and J2 are branch currents; σOSC i is the conductivity of the i-th branch (inherited from the theoretical model of this paper). At the same time, it is also required to satisfy the interface resistance constraint to ensure that the corrected electric field E0′ is consistent with the interface resistance R interface Ohm's law is satisfied, that is, the modified electric field should be able to generate a current that matches the interface resistance.

[0081] Among them, FC α It is a fully connected layer that transforms the hidden state h t Mapped to a polarizability scalar (sigmoid compression to [0,1]). FC E It is another fully connected layer that generates the electric field correction ΔE0, which is superimposed on the original E0 to obtain E0′.

[0082] It is understandable that by decoding the predicted current polarization value and the corrected control electric field, and injecting physical constraints, the model can learn a current distribution and polarization state that is more consistent with physical laws. This improves the model's prediction accuracy for current polarization values. The injection of physical consistency loss and interface resistance constraints makes the model's prediction results more consistent with physical laws. This enhances the physical rationality of the model, making the model's prediction results more credible and convincing. The output of the corrected control electric field E0′ provides a direct basis for subsequent current control operations. By adjusting this electric field value, effective current control can be achieved to meet different application requirements.

[0083] In this embodiment, regarding step S3, electric field control is performed: based on the modified control electric field E0′ and the predicted current polarization value α captured and decoded by the diffusion convolution layer pred , to determine whether the current distribution in the wide branch of the current organic spin device is dispersed and whether the polarizability needs to be reduced; the method is to drive the circuit in the high voltage region (E0>E c ), the corrected control electric field E0' is adjusted according to the direction of the error to offset the nonlinear effect.

[0084] Specifically, in step S300, it is determined whether the current distribution is dispersed: ; If σ J >σ th , the polarizability needs to be reduced.

[0085] Among them, σ J is the diffusion current distribution J diffThe standard deviation of , which measures the degree of dispersion. th is the dispersion threshold (e.g. σ th =0.1μ J ), if the value exceeds the threshold, it is considered as dispersed.

[0086] Specifically, in step S301, electric field adjustment: ; ; Among them, ΔE0′ is the adjustment amount, and the direction is determined by the polarizability error (α ideal −α pred ). H(⋅) is the Heaviside step function and only in the high pressure region (E0′>E c ) to activate the adjustment. is the final corrected control electric field. Sign(·) is a Boolean function used to determine the error type (needs to be prepared in advance). η is the dynamic learning rate, which decays exponentially in the high-voltage region (decay coefficient β>0) to suppress overshoot. E c is the critical electric field (E c =1293.75V / cm), distinguishing between linear and nonlinear regions.

[0087] The principle of the above scheme is: (1) According to the polarizability error (α ideal −α pred ) direction, and calculate the electric field adjustment ΔE0′. The direction of the adjustment is determined by the sign function, and the magnitude is controlled by the dynamic learning rate η.

[0088] (2) Activation adjustment: The Heaviside step function H(⋅) is used to ensure that only in the high pressure area (E0′>E c ) to activate the electric field adjustment.

[0089] (3) Update the electric field: Add the adjustment value ΔE0′ to the original corrected electric field E0′ to obtain the final corrected control electric field E0′adjusted.

[0090] (4) Adjusting the electric field in the high-voltage region can offset the nonlinear effect of the current and improve the linearity of the current control.

[0091] Specifically, in step S303, nonlinear compensation: ; ; The specific dynamic strategy is: S3030, set the initial learning rate η0; set the error threshold ϵ c; Set the learning rate adjustment factor β>0; Initialize the current error E0 and the previous moment error E−1; S3031, Error calculation: At each time step t, calculate the current error E t ; S3032, if E t >ϵ c ,but Otherwise, η t =η0; S3033, constrained learning rate: set the minimum value of the learning rate η min and the maximum value η max ; η t =max(η min ,min(η t ,η max )); Then use the updated learning rate η t To adjust the model parameters S3034, iteration: repeat steps S3031 to S3033 until a stopping condition is met (such as reaching a maximum number of iterations).

[0092] This adaptive adjustment mechanism dynamically adjusts the learning rate based on the error, thereby accelerating convergence when the error is large and maintaining stability when the error is small. By setting the minimum and maximum values ​​of the learning rate, we can prevent the learning rate from being too large or too small, ensuring the robustness of the algorithm.

[0093] Example 2: Please refer to Figure 5 Based on the first embodiment, this embodiment further discloses the execution method of the GRU unit: In this embodiment, regarding S20130, input: ; Among them, the first term is the current diffusion current distribution, which is a three-dimensional tensor that represents the diffusion distribution of the current in the device at the current time step t. Its dimension is m×n×c out , where m and n represent the spatial dimensions, c out Represents the number of channels (or features). This input captures the spatial distribution and dynamic changes of current in the device and is the basis for modeling hysteresis effects in GRU units.

[0094] The second term is the corrected branch electric field, which represents the corrected branch electric field at the current time step t. This term reflects the impact of the interface resistance and branch resistance on the total electric field. Changes in the branch electric field directly affect the current distribution and polarization state, and therefore are an important component of the GRU unit input.

[0095] The third term is the total resistance, which represents the total resistance at the current time step t. It includes the sum of the interface resistance and branch resistance. Changes in the total resistance affect the magnitude and distribution of the current, thereby affecting the current polarization state.

[0096] The fourth term is the anisotropic diffusion coefficient, which reflects the promotion effect of the electric field on the diffusion of spin current.

[0097] In this embodiment, regarding S20131, when calculating the fusion parameters, the model fuses the aforementioned input physical quantities to extract features that have a key impact on the change in the current polarization state. This typically involves linear transformations and nonlinear activations of the input tensors and scalars to generate feature representations suitable for GRU unit processing.

[0098] Calculate fusion parameters ; In this embodiment, regarding S20132, in order to maintain the consistency of the GRU unit input dimension, the GRU input dimension is maintained. , dimension d gru =54; In this embodiment, regarding S20134, propagation is performed: the GRU unit processes the input features according to its internal mechanism (update gate, reset gate) and generates the implicit state h of the current time step t This implicit state encodes the time series information of the current polarization state: ; Among them, in the first constraint, τ s is the spin relaxation time, h eq This constraint requires that the update of the implicit state should take into account the spin relaxation effect, that is, the current polarization state will gradually approach the equilibrium state over time; The second constraint is a physical constraint, which requires the implicit state h t The update of must be consistent with the anisotropic diffusion equation; The third constraint is also a physical constraint, requiring the total resistance R total The time scale of the implicit state reflects the influence of the total resistance on the current magnitude and distribution, thereby indirectly affecting the current polarization state. spin It is a spin capacitor, which is determined by the material properties.

[0099] It is understandable that by incorporating multiple key physical quantities as inputs and accounting for the effects of spin relaxation, anisotropic diffusion, and total resistance, the GRU unit is able to more accurately capture the time series variations of the current polarization state. By introducing physical constraints (spin relaxation time constraint, anisotropic diffusion equation constraint, and total resistance influence constraint), the GRU unit is able to adhere to physical laws when predicting the current polarization state. By maintaining the consistency of the GRU input dimensionality and incorporating multiple key physical quantities as inputs, the model is able to learn more versatile feature representations. This versatility enables the model to generalize to different device structures and experimental conditions, improving its adaptability and robustness.

[0100] Example 3: This example further discloses that the fully connected layer FC in step S202 in Example 1 α and FC E : ; ; ; Among them, FC α It is the implicit state after input pooling, through the weight W α and bias b α After linear transformation, the output polarization rate α is activated by Sigmoid pred ∈[0,1]. FC E Directly output the electric field correction value ΔE0∈R to generate the corrected electric field E0′; .

[0101] For the fully connected layer FC α In this step, the hidden state h t Perform global average pooling and get The purpose of this step is to compress the spatial dimension (m×n) into a 1-dimensional vector while retaining the channel dimension d gru (GRU hidden state dimension). The advantage of this is that it reduces the amount of calculation while retaining the key features in the hidden state. Then, through the weight matrix W α and the bias vector b α The pooled hidden state is linearly transformed and then activated by the Sigmoid function to obtain the predicted current polarization value α pred The Sigmoid function compresses the output to the interval [0,1], indicating the degree of current polarization.

[0102] For the fully connected layer FC E In terms of, it also receives the hidden state after pooling As input. Through the weight matrix W E and the bias vector bE A linear transformation is performed, directly outputting the electric field correction value ΔE0. This correction value is a real number representing the degree of correction to the original electric field E0. Finally, the original electric field E0 is added to the electric field correction value ΔE0 to obtain the corrected control electric field E0′. This electric field will be used for subsequent current control operations.

[0103] in, is the hidden state h t The global average pooling result compresses the spatial dimension (m×n) into a 1-dimensional vector, retaining the channel dimension d gru (GRU hidden state dimension): ; Global average pooling is performed by calculating the hidden state h t The average value over the spatial dimension (m×n) compresses the spatial dimension into a 1-dimensional vector. The advantage of this is that it reduces the complexity of subsequent calculations while retaining the global features in the hidden state. The pooling operation retains the channel dimension d gru , which means that the characteristics of each channel are averaged and preserved. This is very important for capturing the changes in the current polarization state across different channels.

[0104] It is understandable that by extracting key features from the implicit state through global average pooling and performing linear transformations and activations through fully connected layers, the model can learn more accurate current polarization values ​​and electric field corrections. Global average pooling compresses the spatial dimension into a one-dimensional vector, significantly reducing the complexity of subsequent calculations. This enables faster model convergence and improves model training efficiency. By preserving the channel dimension and extracting global features, the model can learn a more universal representation of the current polarization state. This helps the model generalize to different device structures and experimental conditions, improving its adaptability and robustness.

[0105] Example 4: Based on the neural network described in Examples 1 to 3, this example further provides a specific training scheme. The network architecture mainly includes a data encoding module, a diffusion convolution recursive module, an output decoding and physical constraint module, and an electric field control and adjustment module. Among them, the diffusion convolution recursive module is the core, including a diffusion convolution layer, an anisotropic diffusion coefficient module, an interface resistance correction module, and a GRU module.

[0106] (1) Data Preparation: A sensor array is used to collect real-time data from the organic spin device, including electric field strength and current distribution. The analog signal is converted to a digital signal using an analog-to-digital converter (ADC), and a microcontroller (MCU) performs data processing, filtering, and format conversion. The preprocessed dataset is divided into training, validation, and test sets for model training, validation, and testing.

[0107] (2) Model initialization: Randomly initialize all weights and biases in the neural network, including the convolution kernel weights of the diffuse convolution layer, the weights and biases of the GRU module, and the weights and biases of the fully connected layer. Determine the learning rate, batch size, number of iterations, and mean squared error loss function hyperparameters.

[0108] (3) Forward propagation: The input data (including the control electric field E0 and the lateral size distribution TD) is encoded into a tensor form that can be processed by the neural network through the data encoding module.

[0109] (3) Diffused convolution recursion: Diffusion convolution layer: Uses a learnable diffusion kernel to perform convolution operations to extract the non-uniform distribution characteristics of current on the horizontal grid.

[0110] Anisotropic Diffusion Coefficient Module: Introduces the anisotropic diffusion coefficient to distinguish the diffusion intensity along the electric field direction and the horizontal direction.

[0111] Interface resistance correction module: learns the interface resistance value through the fully connected layer and corrects the impact of interface resistance on electric field distribution.

[0112] GRU module: Models the temporal / spatial dependence of the dynamic changes of the electric field on the current polarization and captures the hysteresis effect.

[0113] (4) Output decoding and physical constraints: Decode the implicit state, output the predicted current polarization value and the corrected control electric field, and apply physical consistency loss and interface resistance constraints to ensure the consistency of the predicted results with physical laws.

[0114] (5) Computational loss: Calculate the difference between the predicted current polarization value and the actual current polarization value using the mean square error loss function or other suitable loss functions, while taking into account the physical consistency loss and interface resistance constraints.

[0115] (6) Backpropagation: The gradient of the loss function with respect to the model parameters is calculated through the backpropagation algorithm. The Adam optimizer is used to update the model parameters according to the gradient to minimize the loss function.

[0116] (7) Model Validation: After each training cycle, the validation set is used to evaluate the performance of the model, including prediction accuracy and generalization ability. If the validation set performance does not improve significantly over multiple cycles, training is stopped early to prevent overfitting.

[0117] (8) Model testing: After training is completed, the final performance of the model is evaluated using the test set to ensure the effectiveness and reliability of the model in practical applications.

[0118] In this embodiment, the parameters that need to be trained and learned include: 1) Convolution kernel weights of the diffuse convolution layer: The convolution kernel weights of the diffuse convolution layer are used to extract the non-uniform distribution characteristics of the current on the horizontal grid. During training, the gradient of the loss function with respect to the convolution kernel weights is calculated using the backpropagation algorithm, and the convolution kernel weights are then updated based on the gradient. Specifically, let the loss function be L and the convolution kernel weight be W. The gradient ∂W / ∂L can be calculated using the chain rule. The update formula is: ; where c is the learning rate.

[0119] 2) GRU module weights and biases: These are used to model the temporal and spatial dependence of current dynamics on current polarization and capture hysteresis effects. Similar to the convolution kernel weights in the diffuse convolution layer, the GRU module weights and biases are also updated via the backpropagation algorithm. Specifically, the gradient of the loss function with respect to the GRU module weights and biases is calculated and then updated accordingly.

[0120] 3) Fully connected layer weights and biases: These are used for tasks such as data encoding, interface resistance correction, and output decoding. The weights and biases of the fully connected layers are also updated using the backpropagation algorithm. Specifically, the gradient of the loss function with respect to the weights and biases of the fully connected layers is calculated, and then the gradient is used to update the weights and biases.

[0121] 4) Interface resistance: This value is learned from the input parameters via the fully connected layer and is used to correct the effect of interface resistance on the electric field distribution. The interface resistance value is updated during training using the backpropagation algorithm. Specifically, the interface resistance value is used as the output of the fully connected layer, and the gradient of the loss function with respect to the interface resistance value is calculated, and then the value is updated based on the gradient.

[0122] Furthermore, the neural network architecture includes: % Clear the workspace clear; clc; rng(0); % Define hyperparameters numEpochs = 100; % Number of training cycles batchSize = 32; % batch size learningRate = 0.001; % learning rate % Dataset % trainData: training set input data (N_train x numFeatures) % trainLabels: training set labels (N_train x numLabels) % testData: test set input data (N_test x numFeatures) % testLabels: test set labels (N_test x numLabels) % Neural network architecture layers = [ imageInputLayer([128 128 1]) % 128x128, single channel % Custom layer: data encoding module (tensor concatenation and fully connected layer) customLayer(@encodeFunction, 'Name', 'EncodeLayer') % Diffusion convolution layer convolution2dLayer(3, 32, 'Padding', 'same', 'Name', 'DiffusionConv') % GRU layer gruLayer(64, 'OutputMode', 'sequence', 'Name', 'GRULayer') % Fully connected layer: output decoding and physical constraints fullyConnectedLayer(2, 'Name', 'FCAlpha') % Output current polarization value and electric field correction regressionLayer('Name', 'regressionLayer') % regression layer ]; % Set training options options = trainingOptions('adam', ... 'MaxEpochs', numEpochs, ... 'MiniBatchSize', batchSize, ... 'InitialLearnRate', learningRate, ... 'Plots', 'training-progress', ... 'Verbose', false); % Training network net = trainNetwork(trainData, trainLabels, layers, options); % Test network YPred = classify(net, testData); accuracy = sum(YPred == testLabels) / numel(testLabels); fprintf('Test accuracy: %.2f%%\n', accuracy * 100); % Custom layer function: data encoding module function Z = encodeFunction(X) % X is [N x 128 x 128 x 1] Z = fullyConnectedLayer(128, 'Name', 'EncodingFC')(X); end % Deploy the network (save the network as a .mat file) save('currentPolarizationNetwork.mat', 'net'); % Load the network (load the network from the .mat file) % load('currentPolarizationNetwork.mat', 'net'); % Reasoning (input new input data inputData) % inputData = rand(1, 128, 128, 1); % output = classify(net, inputData); % disp('Predicted current polarization:'); % disp(output); In the above scenario: Input layer: receives preprocessed input data.

[0123] Custom layer: Implements a data encoding module to encode input data into a tensor form that can be processed by the neural network.

[0124] Diffusion convolution layer: Uses a learnable diffusion kernel to perform convolution operations to extract the non-uniform features of current distribution.

[0125] GRU layer: Models the temporal / spatial dependence of the dynamic changes of the electric field on the current polarization and captures the hysteresis effect.

[0126] Fully connected layer: used to output decoding and physical constraints, and output predicted current polarization values ​​and electric field corrections.

[0127] Regression layer: calculates the loss function.

[0128] The training process involves using the Adam optimizer to train and update network parameters via the backpropagation algorithm. At the end of each training cycle, the model performance is evaluated using a validation set to prevent overfitting. The trained model's performance is evaluated using a test set, and the test accuracy is calculated. The trained model is saved as a .mat file or other format for subsequent loading and use.

[0129] Finally, the loaded model is used to perform inference on new input data and output prediction results.

[0130] Embodiment 5: Based on the embodiment 1, this embodiment further provides a compensation solution based on data driving.

[0131] In organic spintronic devices, current diffusion under an electric field exhibits anisotropic behavior, with diffusion along the electric field direction (D∥) typically being significantly stronger than diffusion along the transverse direction (D⊥). However, D(θ) is not a fixed value; the current polarization state is time-dependent, meaning that the current distribution and polarization characteristics at a given moment are influenced by the state at previous moments (hysteresis).

[0132] For example, when the electric field strength suddenly increases, D∥ may briefly increase before gradually stabilizing. By leveraging historical information (such as the values ​​of D(θ) over the past few time steps), we can explore the dynamic changes in D(θ) and achieve more accurate predictions. Current diffusion behavior exhibits both temporal dependence and spatial correlation. Historical information includes the distribution characteristics of current on the horizontal grid. By modeling historical states (hidden layers) using GRU units, we can capture the temporal and spatial correlations of current diffusion, enabling more appropriate adjustments to D(θ). Relying solely on current-time information to modify D(θ) can lead to overfitting (sensitivity to noise) or underfitting (inability to capture long-term dependencies).

[0133] This embodiment aims to use historical information to modify D(θ) to compensate for this hysteresis, enabling the model to more accurately reflect the dynamic process of current diffusion. Historical information provides richer context, helping the model balance local accuracy and global stability when modifying D(θ).

[0134] In this embodiment, regarding S4, based on data-driven compensation: Specifically, S400, collect historical data sequence D': extract the anisotropic diffusion coefficient D(θ) Time window T w (For example, in the past minute), the window M in which the electric field strength E0 suddenly increases is counted d (If the electric field strength exceeds the preset threshold); for window M d At each time step in the time series, the distribution of electric fields of different intensities E0'' and their corresponding anisotropic diffusion coefficients D(θ)'' are statistically analyzed and the exponential model is used for curve fitting to obtain the velocity curve D vc : (1) Exponential model captures its nonlinear change trend: D vc (t)=a·exp(b·t)+c; Where a, b and c are the parameters to be fitted; t is the time variable; (2) Use the least squares method to minimize the squared difference y between the observed value of the anisotropic diffusion coefficient D(θ) and the value of the anisotropic diffusion coefficient D(θ) predicted by the model i sum: ; Where N is the number of observations. Use statistical software or relevant functions or libraries in programming languages ​​such as Python and R to fit the exponential model and estimate its parameters. Solve the aforementioned least squares problem to obtain the values ​​of parameters a, b, and c.

[0135] It should be further noted that the "data clusters" described in step S400 refer to the initial groups into which the historical data sequence is divided based on a preset distance threshold or similarity criterion before the ordered cluster analysis. Specifically, these groups are divided based on the time series characteristics of the anisotropic diffusion coefficient D(θ) and the electric field intensity E0.

[0136] Exemplarily, the historical data are divided into different initial groups according to the mutation point of the electric field intensity E0 or the change trend of D(θ).

[0137] Specifically, S401, ordered cluster analysis: Using an ordered clustering algorithm, data is divided into segments while maintaining order, minimizing intra-segment variance and maximizing inter-segment variance. This method uses dynamic programming to find the optimal segmentation point, avoiding disruption of the temporal continuity of the data and effectively identifying different risk groups (e.g., abnormal diffusion behavior during sudden electric field changes).

[0138] Specifically, according to the historical data sequence D' and the speed curve Dvc Cluster the data clusters to obtain different risk groups. Quantify the evaluation indicators and calculate the adjustment factor Ad just ; Preferably, this embodiment uses the Fisher optimal segmentation method to divide the speed curve D vc The n data clusters in the dataset are divided into K ordered groups so that the similarities and differences within the ordered groups are maximized. Starting from treating all the data as one segment, the number of classifications K is gradually increased. The loss function (the sum of squared deviations within the segment) of each segmentation method is calculated, and finally the segmentation scheme that minimizes the loss function is selected: S4010, objective function driven: Let G1, G2, ..., G K It is a partitioning of n data clusters into K groups: ; Among them, n i is the i-th group G i The data clusters in It is group G i The characteristic mean vector of the data cluster in , ⋅ is the characteristic mean vector of all data clusters.

[0139] S4011, the dynamic programming function L(j,i) is the objective function value of the optimal segmentation of the first j data clusters into i groups. The recursive relationship is: ; in, is the feature mean vector from data cluster s+1 to data cluster j; It represents the difference between the feature mean vector from data cluster s+1 to data cluster j and the feature mean vector of all data clusters. represents the transpose of the difference vector (in the real case, this is the vector itself; in the case of complex numbers or more general vector spaces, the conjugate operation is performed). s represents a limit on the cluster numbering, used to divide different cluster groups. L(s,i−1) represents the objective function value for the optimal partitioning of the first s clusters into i−1 groups.

[0140] S4012, initialization: Because there is only one group, there is no internal difference; so when i=1, L(j,1)=0; S4013, calculation: starting from i=2, gradually increase the number of groups until the predetermined number of groups K is reached; S4014, obtaining the optimal partition: by backtracking the dynamic programming table, obtaining the optimal partition of dividing the n data clusters into K groups; S4015, quantification: ; in, It is group G i The ratio of the number of data clusters in R to the total number of data clusters, i It is group G i risk score; (normalized) adjustment factor Ad just Reflects the risk level of the overall data cluster.

[0141] It should be further noted that the "data clusters" mentioned in step S401 refer to fine-grained groups further divided by Fisher's optimal segmentation method. These groups are divided based on the temporal continuity and feature similarity of the data to ensure that intra-segment differences are minimized and inter-segment differences are maximized.

[0142] For example, the initial “data cluster” is further divided into a high-risk group (abnormal diffusion behavior when the electric field changes suddenly) and a low-risk group (normal diffusion behavior under a stable electric field). By calculating the feature mean vector of each fine-grained “data cluster” , and define the quantitative adjustment factor Ad just , which can reflect the risk levels of different groups. These indicators provide a basis for subsequent dynamic compensation, enabling the model to adapt to complex working environments.

[0143] S402, Execute Compensation: Based on the results of the ordered cluster analysis, dynamically adjust D(θ) to better capture the difference in diffusion intensity along the electric field direction and the transverse direction. By introducing the adjustment factor Adjust, the value of D(θ) is dynamically adjusted according to the risk level of different groups in the historical data, so that the model can maintain stable and efficient control performance under complex working conditions: D′(θ)=D(θ)×Ad just ; Then update the parameters to the next time step for use in the neural network model.

[0144] It is understandable that the regulatory factor Ad just It is calculated based on the quantitative evaluation index (intra-cluster compactness and inter-cluster separation) obtained by Fisher's optimal segmentation method. It reflects the diffusion coefficient adjustment ratio of different risk groups (abnormal diffusion behavior when the electric field suddenly changes). just Combined with the original diffusion coefficient D(θ), the diffusion process can be precisely controlled. just It reflects the diffusion coefficient adjustment ratio of different risk groups. Through multiplication, the modified diffusion coefficient D′(θ) can better adapt to complex working environments.

[0145] The diffusion coefficient is dynamically adjusted by quantitative evaluation indicators, which improves the accuracy of current polarization control. When the electric field changes dynamically, dynamic compensation can adaptively adjust the diffusion coefficient to maintain the stability and efficiency of the system. just , achieving refined correction of the diffusion coefficient and improving the precision of current polarization control. Under complex operating conditions (such as sudden changes in the electric field), dynamic compensation can adaptively adjust the diffusion coefficient to maintain system stability and efficiency.

[0146] Furthermore, closed-loop optimization: by real-time monitoring of the current polarization state, feedback adjustment of the cluster analysis and model fitting steps, end-to-end closed-loop optimization is achieved.

[0147] Furthermore, the solution of this embodiment will be deduced and verified based on Python: import numpy as np from scipy.optimize import curve_fit # Exponential model definition def exponential_model(t, a, b, c): return a * np.exp(b * t) + c # Develop simulation data for this example t = np.linspace(0, 1, 100) # time variable D_theta_true = 2 * np.exp(0.5 * t) + 1 # True D(theta) value noise = 0.1 * np.random.normal(size=t.shape) # add noise D_theta_observed = D_theta_true + noise # Observed D(theta) value # Least Squares Method p0 = [1, 0.1, 0.5] # Initial parameters popt, pcov = curve_fit(exponential_model, t, D_theta_observed, p0=p0) # Fitting parameters a_fit, b_fit, c_fit = popt print(f"Fitting parameters: a = {a_fit}, b = {b_fit}, c = {c_fit}") # Verify ordered cluster analysis from sklearn.cluster import KMeans # Generate clustering data X = np.array([[1, 2], [1, 4], [1, 0], [4, 2], [4, 4], [4, 0]]) Clustering kmeans = KMeans(n_clusters=2, random_state=0) kmeans.fit(X) # Output clustering results print("Cluster Center:", kmeans.cluster_centers_) print("Cluster to which the data point belongs:", kmeans.labels_) The final output is: a = 0.805565607970687; b = 0.9301061000479448; c = 2.252175377798324; Cluster centers: [[1. 2.] [4. 2.]]; The cluster to which the data point belongs: [0 0 0 1 1 1]; The results returned by the code interpreter tool indicate that the exponential model fitting and ordered cluster analysis methods used in this step are correct. The exponential model fitting yielded reasonable parameters, and the cluster analysis results were consistent with expectations.

[0148] It is understandable that the revised D(θ) more accurately describes the anisotropy of current diffusion, making the current distribution more uniform across the lateral grid and reducing localized overcurrent or undercurrent. By dynamically adjusting D(θ), the model can more quickly respond to changes in current distribution and promptly adjust the control electric field E0′, thereby improving the efficiency of current polarization control. The historical information-based correction mechanism enables the model to more accurately predict current polarization values, helping to reduce experimental trial-and-error costs and accelerate device optimization. The revised model is more adaptable to different device structures and experimental conditions, maintaining stable and efficient control performance even in complex and changing operating environments.

[0149] Example 6: Figure 8 As shown, this embodiment further discloses a current polarization control system for an organic spin device. This system aims to address the errors caused by prior art due to neglect of lateral size effects and interface resistance. By estimating the required electric field, the system then regulates the current. The system primarily includes a data acquisition and processing module, a computing module, and an organic spin device control module.

[0150] (1) A data acquisition and processing module for collecting the physical parameters of organic spin devices, including a sensor array, an analog-to-digital converter (ADC), and a microcontroller (MCU).

[0151] Sensor array: Responsible for collecting real-time data from organic spin devices, including electric field strength and current distribution. These raw signals are the basis for subsequent processing.

[0152] Analog-to-digital converter (ADC): Converts the analog signals collected by the sensor array into digital signals for processing by the microcontroller.

[0153] Microcontroller (MCU): Processes, filters, and converts digital signals, extracts useful information, and passes it to the computing module.

[0154] (2) A calculation module for executing the method of steps S1 to S3, comprising: (2.1) Data encoding module responsible for encoding input data into tensor form that can be processed by the neural network; (2.2) Diffusion convolution recursive module responsible for executing the neural network algorithm: It includes a diffusion convolution module to capture the non-uniform distribution of current on the lateral grid and extract the characteristics of the current distribution through convolution operation.

[0155] The anisotropic diffusion coefficient module is used to distinguish the diffusion intensity in different directions and consider the differences in current propagation in different directions.

[0156] The interface resistance correction module is used to correct the influence of interface resistance on electric field distribution and improve the accuracy of electric field prediction.

[0157] The GRU module is used to model the temporal / spatial dependence of current dynamics on current polarization and capture the dynamic relationship between the electric field and current polarization.

[0158] (2.3) Decode the implicit state and output the predicted current polarization value and the corrected control electric field. Output decoding and physical constraint module: Apply physical consistency loss and interface resistance constraints to ensure the consistency of the prediction results with physical laws.

[0159] (2.4) The electric field control and adjustment module calculates the dispersion of the current distribution based on the predicted current polarization value and the corrected control electric field, and determines whether the electric field needs to be adjusted. In the high-voltage region, the electric field is dynamically corrected according to the direction of the polarization error, and nonlinear compensation and learning rate decay are considered to form control instructions. (3) Organic spin device control module: Including digital-to-analog converter (DAC), power supply and drive circuit (high voltage drive circuit, pulse drive circuit or programmable drive circuit); Digital-to-analog converter (DAC): Receives control instructions from the electric field control and adjustment module, calculates the control parameters, and then converts them into analog signals.

[0160] Power supply: Provides the required voltage output to the drive circuit.

[0161] Driving circuit: applies the corresponding electric field to the organic spin device to achieve the regulation of current polarization.

[0162] Through the precise coordination of the DAC and driver circuit, the system can achieve precise electric field control of the organic spin device. The fast response capability of the driver circuit enables the system to quickly adapt to different operating conditions, improving the system's response speed.

[0163] All of the above embodiments merely represent implementation methods of the present invention in practical applications. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the scope of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the appended claims.

[0164] For those skilled in the art, it can be further appreciated that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0165] At the same time, those skilled in the art will understand that all or part of the processes in all the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

Claims

1. A method for controlling current polarization in an organic spin device, characterized in that: include: S1, determine the lateral size distribution TD = [G, L G ], where G is a two-dimensional grid, L G is the transverse geometry, and L G =(B w ,I s ), where B w is the branch width, I s is the interface shape; And determine the control electric field E0 and the ideal current polarization value α ideal ; S2 performs diffuse convolution recursion, including controlling the encoding input of the electric field E0 and the lateral size distribution TD and capturing the non-uniformly distributed current through the diffuse convolution layer, while distinguishing the diffusion intensity along the electric field direction and the lateral direction; by correcting the interface resistance and simulating the hysteresis effect of the GRU unit, the current polarization value is predicted and the electric field is corrected; S3, based on the modified control electric field E0′ and predicted current polarization value α captured and decoded by the diffusion convolution layer pred , determine whether the current distribution in the wide branch of the current organic spin device is dispersed and whether the polarization rate needs to be reduced; and control the electric field, thereby controlling the current polarization behavior of the organic spin device.

2. The current polarization control method according to claim 1, characterized in that: In said S1: ； Among them, the two-dimensional grid G ​​is represented by the discrete point coordinates (iΔ x ,jΔ y ) indicates that the step size Δ x and Δ y Need to be less than the spin diffusion length L s ; r is the curvature radius of the device; polarization value α ideal The value range is [0,1].

3. The current polarization control method according to claim 1, characterized in that: In said S2, it includes: S200, input layer: encode the control electric field E0 and the lateral size distribution TD into the independent variable X and input it into the diffusion convolution layer; S201, Diffusion Convolution Layer: Uses a learnable diffusion kernel (convolution kernel weight) to capture the non-uniform distribution of current on the lateral grid; introduces anisotropic diffusion coefficients to distinguish the diffusion intensity along the electric field direction from the lateral direction; corrects the effect of interface resistance on electric field distribution based on the interface resistance simulation module, including resistance parameter layer and electric field redistribution; GRU unit models the temporal / spatial dependence of electric field dynamic changes on current polarization; combines historical states to capture the hysteresis effect of spin diffusion length changes with electric field changes; S202, output layer: for the hidden state h t Perform decoding operation and output predicted current polarization value α pred and the corrected controlled electric field E0'; then the controlled electric field E0' is forced to satisfy the current continuity equation and be self-consistent with the interface resistance simulation module.

4. The current polarization control method according to claim 3, characterized in that: In the S200, the input: ； Among them, Encode(·) is the encoding function that maps the input parameters to tensors that the neural network can process; FC(·) is the fully connected layer used for dimensionality increase or decrease; ⊕ is the tensor concatenation operation; X is the encoded input tensor, which serves as the input of the dilated convolutional layer; R is the number of channels.

5. The current polarization control method according to claim 3, characterized in that: In S201, the diffusion convolution recursively: ； Among them, K diff is the learnable diffusion kernel; k is the convolution kernel size, c in / c out is the number of input / output channels; J diff is the current distribution tensor after diffuse convolution, capturing lateral non-uniformity; Conv2D(·) is the two-dimensional convolution kernel; R is the number of channels.

6. The current polarization control method according to claim 5, characterized in that: In S201, the method for introducing the anisotropic diffusion coefficient includes: ； Where θ is the angle between the diffusion direction and the electric field direction; D(θ) is the anisotropic diffusion coefficient, along the electric field direction D ∥ and horizontal D ⊥ The diffusion intensity varies; Perform interface resistance correction: ； ； ； Among them, R interface is the interface resistance value, which is learned from the input parameters by the fully connected layer FC(·); E1′, E2′ are the corrected branch electric fields, considering the effect of the interface resistance on the total resistance R total The impact of R OSCi is the branch resistance.

7. The current polarization control method according to claim 6, characterized in that: The architecture composed of the GRU unit is: ； ； ； ； Among them, z t and r t are the update gate and reset gate; h t is the implicit state at time t, representing the encoded current polarization value; is the candidate hidden state; σ is the sigmoid activation function; W z and W r is the weight matrix in the reset gate, used to calculate the activation value of the reset gate; h t-1 is the implicit state of the previous time step; tanh(·) is the hyperbolic tangent function; ⊙ is the Hadamard product.

8. The current polarization control method according to claim 3, characterized in that: In S202, the decoding operation, the forcing of the current continuity equation, and the self-consistent method of the interface resistance simulation module include: ； ； Among them, σ OSCi is the conductivity of the i-th branch; FC α It is a fully connected layer that transforms the hidden state h t Mapping to polarizability scalar; FC E It is another fully connected layer that generates the electric field correction ΔE0, which is superimposed on the original E0 to obtain the corrected control electric field E0′; J0=J1+J2 is the current continuity equation.

9. The current polarization control method according to any one of claims 1 to 8, characterized in that: In S3, the driving circuit adjusts the corrected control electric field E0' according to the direction of the error to offset the nonlinear effect; The process is: Dispersion calculation: J based on the output of the diffusion convolution layer diff , calculate the current distribution standard deviation σ J ; Polarizability adjustment decision: If σ J >σ th , then it is necessary to reduce the polarizability by lowering E0′; Adjustment: When E0′>critical electric field E c When the error direction sign(α ideal −α pred ) Dynamically correct the electric field; where sign(·) is a Boolean function; Learning rate decay: The learning rate η decays exponentially as E0′ increases to avoid overshoot.

10. A control system for implementing the current polarization control method according to any one of claims 1 to 9, characterized in that: include: an organic spin device control module, including a digital-to-analog converter, a power supply, and a drive circuit; A data acquisition and processing module for collecting physical parameters of organic spin devices, including a sensor array, an analog-to-digital converter, and a microcontroller; The calculation module for executing the method of steps S1 to S3 includes: A data encoding module responsible for encoding input data into a tensor form that can be processed by the neural network; The diffusion convolution recursive module is responsible for executing the neural network algorithm; Output decoding and physical constraint module that decodes the implicit state and outputs the predicted current polarization value and the corrected control electric field; Based on the predicted current polarization value and the corrected control electric field, the dispersion of the current distribution is calculated to determine whether the electric field control and adjustment module needs to be adjusted.

Images