An amorphous oxide semiconductor device characteristic prediction method based on a physical information neural network

Abstract

The application discloses a kind of amorphous oxide semiconductor device characteristics prediction method based on physical information neural network, comprising: using TCAD to construct the simulation model of amorphous oxide semiconductor device, and generate training data set;Physical information neural network is constructed, and input layer parameter and output layer physical field quantity are defined;Compound loss function including data loss term and physical loss term is constructed;The training data set generated by TCAD simulation is used with the joint training of physical loss term equation constraint physical information neural network;The electrical characteristics of device under different geometric structure parameters or bias voltage conditions are predicted using the trained physical information neural network and the differentiable current integration layer.The application can realize high-precision, low-cost amorphous oxide semiconductor device electrical characteristic prediction under the premise of meeting semiconductor physical constraint, significantly improve simulation efficiency and model generalization ability.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device modeling and artificial intelligence-assisted manufacturing technology, specifically to a method for predicting the characteristics of amorphous oxide semiconductor devices based on a physical information neural network (PINN). Background Technology

[0002] Amorphous oxide semiconductors (AOS) have become core materials for next-generation flat panel display driver backplanes and flexible electronic devices due to their high carrier mobility, good large-area uniformity, and low-temperature process compatibility. However, AOS devices face severe reliability challenges in practical applications, especially under bias stress, illumination, or thermal fields, where the oxygen vacancy defect state inside the device undergoes dynamic evolution, leading to threshold voltage drift and current characteristic degradation. This multiphysics coupling mechanism involving charge transport (fast process) and defect dynamics (slow process) is extremely complex, posing significant difficulties for accurate device modeling and lifetime prediction.

[0003] Traditional semiconductor device modeling primarily relies on Technology Computer-Aided Design (TCAD) tools. TCAD uses the finite element method or finite volume method to discretize and solve partial differential equations such as the Poisson equation and the drift-diffusion equation. While TCAD simulations offer high accuracy, the computational load is extremely large when handling transient simulations involving defect state evolution, often taking hours or even days, making it difficult to meet the rapid iteration requirements of device models in integrated circuit design. Furthermore, TCAD simulations typically face convergence problems in solving coupled multiphysics equations and struggle to use limited experimental data to back-calibrate internal device physical parameters (such as defect generation rate and mobility), limiting its application in "virtual-real hybrid" scenarios.

[0004] In recent years, data-driven methods, represented by deep learning, have been attempted for device modeling. While traditional neural network models offer fast inference speeds, they are essentially "black box" models, requiring massive amounts of experimental data for training and lacking physical interpretability. Purely data-driven models often only fit the surface mapping of voltage-current (IV), ignoring the internal potential distribution and charge conservation law of the device, resulting in poor generalization ability. When the predicted operating conditions exceed the range of the training data, the model's predictions often violate physical principles, failing to provide effective guidance for device process optimization.

[0005] Physical Information Neural Networks (PINNs), as an emerging scientific computing paradigm, attempt to address the aforementioned problems by incorporating the residuals of physical equations into the loss function. However, existing PINN research primarily focuses on solving the steady-state Poisson equation, mostly targeting simple silicon-based devices or idealized models. For amorphous oxide semiconductor devices, existing PINN methods have not effectively addressed how to integrate the complex "oxygen vacancy defect dynamics" equations into the neural network, nor do they lack an effective design for the differentiable mapping mechanism between microscopic physical fields (field quantities) and macroscopic electrical properties (current). Therefore, there is an urgent need for an efficient simulation method that can integrate prior knowledge from TCAD data and take into account both the physical laws of microscopic defect evolution and the prediction of macroscopic properties. Summary of the Invention

[0006] To address the shortcomings of existing amorphous oxide semiconductor device (AOS) modeling techniques, such as high computational cost, difficulty in convergence of multi-physics coupling, and difficulty in simultaneously considering microscopic defect evolution and macroscopic characteristic prediction, this invention aims to provide a method for predicting the characteristics of AOS based on a Physical Information Neural Network (PINN). This method incorporates the oxygen vacancy defect dynamics equation into the physical constraint mechanism of the neural network and combines it with a differentiable current integration layer design. This enables efficient, high-precision, and physically interpretable prediction of the transient electrical characteristics and aging behavior of AOS under multi-physics conditions without relying on massive amounts of experimental data.

[0007] The technical solution of the present invention to achieve the above objectives specifically includes:

[0008] In a first aspect, the present invention provides a method for predicting the characteristics of amorphous oxide semiconductor devices based on a physical information neural network, comprising the following steps:

[0009] A simulation model of an amorphous oxide semiconductor device was constructed using TCAD simulation software, and a training dataset was generated.

[0010] Construct a physical information neural network, and define the input layer parameters and output layer physical field quantities of the physical information neural network;

[0011] A composite loss function is constructed that includes a data loss term and a physical loss term. The physical loss term is constructed based on the semiconductor physical equations and is used to constrain the network output to conform to physical laws.

[0012] The physical information neural network is jointly trained using the training dataset generated by TCAD simulation software and the physical loss term constraints, and an adaptive balancing strategy is adopted to optimize the weights of the data loss term and the physical loss term.

[0013] The electrical characteristics of amorphous oxide semiconductor devices under different geometric parameters or bias voltage conditions are predicted by using a trained physical information neural network and a differentiable current integration layer.

[0014] Preferably, the simulation model of the amorphous oxide semiconductor device includes:

[0015] The geometry of the amorphous oxide semiconductor device is constructed according to the preset gate length and gate width;

[0016] An active layer of amorphous oxide semiconductor is set in the geometric structure, and its carrier mobility model, defect state distribution model and oxygen vacancy related physical parameters are defined.

[0017] By applying bias voltage conditions to the gate, source, and drain regions, and then extracting the corresponding device output drain-source current and internal physical field distribution after simulation, a training dataset is constructed; the bias voltage conditions include the gate-source voltage and the drain-source voltage.

[0018] More preferably, the internal physical field distribution includes electric potential distribution, electron concentration distribution, and oxygen vacancy defect concentration.

[0019] Preferably, the training dataset includes:

[0020] Input vector data includes the spatial coordinates of the simulation mesh nodes, the time points of the transient simulation, the corresponding bias voltage conditions, and the geometric parameters;

[0021] Microscopic field quantity supervision data, including internal physical field distribution data, is used to construct the initial guidance and regularization constraints for physical information neural networks;

[0022] Macroscopic characteristic monitoring data, including the output drain-source current of the device under different bias voltage conditions, is used to calibrate the integral characteristics of the model output.

[0023] Preferably, the input layer parameters of the physical information neural network include at least: input vector data of the training dataset; the output layer physical field quantities include at least microscopic field quantity supervision data. The microscopic field quantity supervision data includes at least internal physical field distribution data.

[0024] Preferably, the data loss term of the physical information neural network adopts a data loss term based on the simulated output drain-source current, which is used to measure the deviation between the predicted output drain-source current and the simulated output drain-source current of the physical information neural network; the physical loss term includes a physical residual loss term based on the Poisson equation, a physical residual loss term based on the drift-diffusion equation, and a physical residual loss term based on oxygen vacancy trap dynamics. Among them, the physical residual loss term based on the Poisson equation is used to constrain the relationship between the potential distribution and the space charge density inside the device, the physical residual loss term based on the drift-diffusion equation is used to constrain the relationship between the transient carrier concentration distribution and the corresponding current density distribution and its time rate of change, and the physical residual loss term based on oxygen vacancy trap dynamics is used to describe the dynamic process of the generation, recombination and migration of oxygen vacancy concentration as the bias stress, electric field or temperature changes.

[0025] Preferably, the implementation process of the differential current integration layer is as follows:

[0026] The spatial derivatives of electrostatic potential and electron concentration are obtained by differential calculation based on electrostatic potential and electron concentration, respectively.

[0027] The current density along the channel direction was calculated based on the drift-diffusion theory.

[0028] The predicted terminal current is obtained by numerically integrating the current density along the thickness direction of the amorphous oxide semiconductor device and multiplying it by the gate width.

[0029] Preferably, the physical information neural network adopts a multi-layer feedforward fully connected neural network structure, including an input layer, at least three hidden layers, and an output layer; the hidden layers are connected by weighted connections and use continuously differentiable nonlinear activation functions;

[0030] Training is performed using the backpropagation algorithm and gradient-based optimization methods to minimize the composite loss function, which is a weighted sum of data loss and physical loss terms.

[0031] In a second aspect, the present invention provides an amorphous oxide semiconductor device characteristic prediction system for implementing the above method, comprising the following steps:

[0032] The data acquisition module is used to acquire the bias voltage conditions and geometric parameters of the device;

[0033] The prediction module is used to predict the electrical characteristics of amorphous oxide semiconductor devices under different geometric parameters or bias voltage conditions based on the geometric parameters and bias voltage conditions by calling a trained physical information neural network and a differentiable current integration layer. Specifically, the prediction module generates spatial sampling coordinates and time sampling points based on the geometric parameters, obtains the internal physical field distribution of the device at the spatial sampling coordinates and time sampling points through the physical information neural network, and further calculates the electrical characteristics of the amorphous oxide semiconductor device under the corresponding conditions through the differentiable current integration layer.

[0034] Thirdly, the present invention provides a computing device, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the above-described method.

[0035] The beneficial effects of the present invention include at least the following:

[0036] This invention introduces the governing equations into neural network training in the form of "soft constraints" by constructing a composite loss function that includes both data and physical loss terms. After training, the PINN model can directly output the physical field distribution and electrical properties with only one forward propagation, eliminating the time-consuming iterative solution process in traditional methods. This improves the simulation speed by several orders of magnitude. At the same time, the physical loss term ensures the inherent physical consistency of the prediction results, avoiding the problem of insufficient accuracy in purely data-driven models.

[0037] This invention directly incorporates the physical residual loss term based on the Poisson equation, drift-diffusion equation, and oxygen vacancy trap dynamics into the loss function, forcing the neural network to learn and obey fundamental physical laws such as charge conservation, carrier transport, and defect evolution during training. This mechanism enables the model not only to fit data but also to understand and embed the physical principles of semiconductor device operation. Therefore, when faced with new geometric parameters or bias voltage conditions, it can still make reasonable predictions consistent with physical common sense, significantly improving the model's generalization ability and reliability.

[0038] This invention introduces a physical residual loss term based on oxygen vacancy trap dynamics, using the dynamic equations describing defect generation, recombination, and migration as constraints. This enables the trained PINN model to simultaneously and coupledly predict fast processes (potential, carrier transport) and slow processes (defect state evolution), thereby accurately simulating key reliability phenomena such as threshold voltage drift and current degradation. This provides a highly efficient tool for device lifetime prediction that is difficult to achieve with previous pure TCAD or pure data-driven methods.

[0039] This invention integrates a differentiable current integrator layer at the output of a PINN. This layer utilizes automatic differentiation technology to directly calculate the current density based on the potential distribution and carrier concentration distribution at the network output, using drift-diffusion theory, and then obtains the port current through integration. This design ensures that the prediction of the macroscopic current originates entirely from the evolution of the microscopic physical field, guaranteeing that the prediction chain from internal mechanisms to external characteristics is physically and mathematically self-consistent and differentiable. This unifies the data loss term (based on current error) and the physical loss term (based on field equations) within an optimizable framework, ensuring the physical self-consistency of the port characteristic prediction.

[0040] The training data of this invention originates from a training dataset generated by TCAD simulation, containing rich internal physical field distribution data. This is equivalent to providing "physical foreknowledge" supervision for the neural network using a high-fidelity simulation simulator. Compared to data-driven methods that require a large amount of time-consuming and expensive experimental test data, this invention only requires relatively limited TCAD simulation data, combined with strong physical constraints, to train a high-performance model. This provides a powerful tool for device design and characteristic prediction in the early stages of R&D when experimental data is scarce. In summary, this invention proposes a method for predicting the characteristics of amorphous oxide semiconductor devices by deeply integrating physical mechanisms and neural network architecture. This method combines high accuracy, high efficiency, strong generalization ability, and physical interpretability, effectively solving the common problems of high computational cost and poor physical consistency of traditional modeling methods and data-driven methods in this field. Attached Figure Description

[0041] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of a method for predicting the characteristics of amorphous oxide semiconductor devices based on a physical information neural network, provided in an embodiment of the present invention.

[0043] Figure 2 This is a schematic diagram of the structure of the amorphous oxide semiconductor device characteristic prediction system provided in an embodiment of the present invention. Detailed Implementation

[0044] The present invention will be further analyzed below with reference to specific embodiments.

[0045] This embodiment provides a method for predicting the characteristics of amorphous oxide semiconductor devices based on physical information neural networks. It is primarily applied in the field of AI-assisted intelligent manufacturing of integrated circuits. (See appendix.) Figure 1 This includes the following steps:

[0046] S1. Construct a simulation model of an amorphous oxide semiconductor device using TCAD simulation software and generate a training dataset;

[0047] S2. Construct a Physical Information Neural Network (PINN), and define the input layer parameters and output layer physical field quantities of the Physical Information Neural Network;

[0048] S3. Construct a composite loss function that includes a data loss term and a physical loss term. The physical loss term is constructed based on the semiconductor physical equations and is used to constrain the network output to conform to physical laws.

[0049] S4. The physical information neural network is jointly trained using the training dataset generated by TCAD simulation software and the semiconductor physical equation (i.e., physical loss term) constraints, and an adaptive balancing strategy is adopted to optimize the weights of the data loss term and the physical loss term.

[0050] S5. Utilize a trained physical information neural network and a pre-trained differential current integrator layer to predict the electrical characteristics of amorphous oxide semiconductor devices under different geometric parameters or bias voltage conditions.

[0051] This method enables end-to-end modeling, from physical principles to data-driven approaches, and from microscopic field quantities to macroscopic properties.

[0052] Specifically, step S1 is as follows:

[0053] In the device modeling step, a simulation model of the amorphous oxide semiconductor device is constructed based on TCAD. The simulation model includes:

[0054] The device geometry is constructed according to the preset gate length L and gate width W;

[0055] An amorphous oxide semiconductor active layer is set in the geometric structure, and its carrier mobility model, defect state distribution model and oxygen vacancy related physical parameters are defined.

[0056] Applying external bias voltages, including gate-source voltages, to the gate, source, and drain regions. With drain-source voltage Perform numerical simulation calculations to extract the corresponding device output drain-source current. And the internal physical field distribution (such as electric potential distribution, electron concentration distribution, oxygen vacancy defect concentration), thereby constructing a training dataset for model training;

[0057] In one embodiment, the training dataset includes multi-dimensional physical field data and device port characteristic data, specifically including:

[0058] Input vector data, including the spatial coordinates of the simulated mesh nodes. The time point of transient simulation Corresponding bias voltage conditions and geometric parameters ;

[0059] Microscopic field quantity monitoring data, including internal physical field distribution data, i.e., the electrostatic potential distribution at the stated spatial coordinates. Electron concentration distribution and oxygen vacancy defect concentration , used for initial guidance and regularization constraints in constructing physical information neural networks;

[0060] Specifically, step S2 is as follows:

[0061] The Physics-Informed Neural Network (PINN) adopts a multi-layer feedforward fully connected neural network structure, including an input layer, at least three hidden layers, and an output layer.

[0062] The hidden layers are connected by weighted connections and employ continuously differentiable nonlinear activation functions to enhance the network's ability to express nonlinear physical relationships.

[0063] During model training, the device's geometric parameters, bias voltage conditions, spatial coordinates, and transient simulation time points are used as network inputs, while physical quantities such as the device's internal electrostatic potential distribution, electron concentration, and oxygen vacancy defect concentration are used as network outputs.

[0064] The input vector parameters received by the input layer of the PINN model At least including: spatial coordinates Transient simulation time Gate-source voltage Drain-source voltage and device geometry parameters (gate length) (and gate width W); physical field vector output by the output layer At least including: electrostatic potential distribution electron concentration and oxygen vacancy defect concentration .

[0065] The PINN model is trained using backpropagation and gradient-based optimization to minimize the total loss function, which is a weighted sum of data and physical loss terms. The training objective of the PINN model is to minimize this total loss function, which is composed of a data loss term constructed from simulation data and a physical residual loss term constructed from the Poisson equation, drift-diffusion equation, and defect-state dynamics equation. The model parameters are iteratively updated using backpropagation combined with gradient-based optimization until the loss function converges, thus obtaining a device characteristic prediction model that satisfies physical consistency.

[0066] Specifically, step S3 is as follows:

[0067] The loss function of the PINN model includes a data loss term and a physical loss term, specifically:

[0068] (1)

[0069] in, To output drain-source current based on simulation Data loss items, These are the weighting coefficients for the data loss terms; For physical loss items, This is the physical residual loss term based on the Poisson equation. This is the physical residual loss term based on the drift-diffusion equation. This is the physical residual loss term based on oxygen vacancy trap dynamics. , , These are the weighting coefficients for the corresponding physical residual loss terms;

[0070] Data loss items It is used to measure the deviation between the drain-source current output by the model and the simulation data, and its form is:

[0071] (2)

[0072] in, The drain-source current data output by the simulation model. The drain-source current values ​​predicted by the PINN model and the differentiable current integration layer are used together with the data loss term and the physical loss term to train the PINN model, so as to improve the accuracy of the electrical characteristic prediction of amorphous oxide semiconductor devices.

[0073] The PINN model prediction calculation process, i.e., the data loss term. The construction process involves a differentiable current integration layer.

[0074] The specific steps for implementing the differential current integration layer are as follows:

[0075] (1) Based on electrostatic potential With electron concentration The electrostatic potential was obtained by differential calculation. With electron concentration spatial derivative and subscript This indicates calculating the gradient with respect to spatial coordinates;

[0076] (2) Based on the drift-diffusion theory, combined with and Calculate the current density along the channel direction. ;

[0077] (3)

[0078] in, The amount of electron charge. For electron mobility, For electron concentration, The electron diffusion coefficient;

[0079] (3) Regarding the current density The predicted terminal current is obtained by numerical integration along the thickness direction of the amorphous oxide semiconductor device and multiplication by the gate width W. That is, the drain-source current value obtained by the PINN model and the prediction calculation of the differentiable current integration layer.

[0080] Finally, through the predicted terminal current The value is the drain-source current output by the simulation model obtained in step S1. The mean squared error between data points is used as a data loss term. .

[0081] Physical residual loss term based on Poisson equation Used to constrain the potential distribution inside the device. The relationship with space charge density (including electrons, holes, and charged defects);

[0082] For example, the residual of the Poisson equation is defined as:

[0083] (4)

[0084] in, electric potential distribution The Laplace operator (i.e.) ), Represents the total net charge density. Hole concentration, Let (x, y) represent the dielectric constant of the amorphous oxide semiconductor device material, and (x, y) represent the spatial coordinates within the cross-section of the amorphous oxide semiconductor device. This indicates the concentration of oxygen vacancy defects.

[0085] Physical residual loss term based on Poisson equation Represented as:

[0086] (5)

[0087] in, This represents the total number of physical sampling points. Indicates the first Spatial location and temporal information of each sampling point For the first Total net charge density at each sampling point.

[0088] Physical residual loss term based on drift-diffusion equation Used to constrain transient carrier concentration distribution With the corresponding current density distribution The relationship between them and their rate of change over time;

[0089] For example, the residual of the drift-diffusion equation is defined as:

[0090] (6)

[0091] in, This represents the rate of change of electron concentration over time. The divergence of electron current density (in the 1-dimensional case is...) In the 2D case, it is ); It represents the net recombination rate, which includes the process by which electrons are captured or released by oxygen vacancy defects, as well as other generation-recombination mechanisms.

[0092] The physical residual loss term based on the drift-diffusion equation Represented as:

[0093] (7)

[0094] in, This represents the total number of physical sampling points. Indicates the first The rate of change of electron concentration at each sampling point over time Indicates the first The divergence of electron current density at each sampling point Indicates the first Net composite rate of each sampling point.

[0095] Physical residual loss term based on oxygen vacancy trap dynamics Used to describe oxygen vacancy concentration The dynamic processes of generation, recombination, and migration that occur in response to changes in bias stress, electric field, or temperature;

[0096] For example, the oxygen vacancy kinetic residual is defined as:

[0097] (8)

[0098] in, The defect concentration is the rate of change over time. The defect diffusion coefficient is denoted as . This represents the rate of change in defect concentration per unit volume due to the diffusion of oxygen vacancies under the influence of a spatial concentration gradient. Represents the net generation function, which describes the oxygen vacancies at a temperature of Local electron concentration and local potential State transition rate under excitation;

[0099] The physical residual loss term based on oxygen vacancy trap dynamics Represented as:

[0100] (9)

[0101] in, This represents the total number of physical sampling points. Indicates the first The rate of change in defect concentration per unit volume at each sampling point due to the diffusion of oxygen vacancies under the influence of the spatial concentration gradient. Indicates the first The net generating function for each sampling point.

[0102] The physical residual loss terms mentioned above are all calculated from the output or intermediate variables of the PINN model using automatic differentiation.

[0103] Specifically, step S4 is as follows:

[0104] A physical sampling point set is constructed, and physical sampling points are selected within the spatiotemporal solution domain of the device using Latin hypercube sampling or gradient-based adaptive sampling methods. Used to calculate the physical residual loss term , and ;

[0105] Perform joint training, and reduce data loss terms. With each physical residual loss item , , We calculate the weighted summation to construct the overall objective function, and then use a gradient-based optimization algorithm to update the neural network parameters.

[0106] Implement adaptive weight balancing for the loss terms, dynamically adjusting the weighting coefficients of each loss term during training iterations. , , and The adjustment strategy specifically involves: calculating the gradient norm of each loss term with respect to the neural network parameters; and calculating dynamic balancing weights based on the relative magnitude or statistical distribution of the gradient norms of each loss term, so that the gradient contributions of each loss term during backpropagation are on the same order of magnitude, thereby eliminating the training non-convergence problem caused by different physical dimensions.

[0107] Specifically, step S5 is as follows:

[0108] In one embodiment, the implementation process of the pre-trained differentiable current integrator layer is as follows:

[0109] The spatial derivatives of the electrostatic potential and electron concentration output by the automatic differential calculation network are obtained by differential calculation based on the electrostatic potential and electron concentration.

[0110] The current density along the channel direction is calculated based on the drift-diffusion theory; the current density is numerically integrated along the thickness direction of the amorphous oxide semiconductor device and multiplied by the gate width to obtain the predicted terminal current.

[0111] After the model training converges, during the prediction phase, the user only needs to provide new input conditions: L, W, , Using a trained physical information neural network and a differentiable current integrating layer, the L, W, and W values ​​of amorphous oxide semiconductor devices under normal operating conditions are predicted. , Corresponding .

[0112] See appendix Figure 2 This embodiment also provides an amorphous oxide semiconductor device characteristic prediction system, including the following steps:

[0113] The data acquisition module is used to acquire the bias voltage conditions and geometric parameters of the device;

[0114] The prediction module is used to predict the electrical characteristics of amorphous oxide semiconductor devices under different geometric parameters or bias voltage conditions based on the geometric parameters and bias voltage conditions by calling a trained physical information neural network and a differentiable current integration layer. Specifically, the prediction module automatically generates spatial sampling coordinates and time sampling points based on the geometric parameters, obtains the internal physical field distribution of the device at the spatial sampling coordinates and time sampling points through the physical information neural network, and further calculates the electrical characteristics of the amorphous oxide semiconductor device under the corresponding conditions through the differentiable current integration layer.

[0115] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for predicting the characteristics of amorphous oxide semiconductor devices based on physical information neural networks, characterized in that, Includes the following steps: A simulation model of an amorphous oxide semiconductor device was constructed using TCAD simulation software, and a training dataset was generated. Construct a physical information neural network, and define the input layer parameters and output layer physical field quantities of the physical information neural network; A composite loss function is constructed, which includes a data loss term and a physical loss term, wherein the physical loss term is constructed based on the semiconductor physical equations. The physical information neural network is jointly trained using a training dataset generated by TCAD simulation software and physical loss term constraints. The electrical characteristics of amorphous oxide semiconductor devices under different geometric parameters or bias voltage conditions are predicted by using a trained physical information neural network and a differentiable current integration layer.

2. The method for predicting the characteristics of amorphous oxide semiconductor devices based on a physical information neural network according to claim 1, characterized in that, The simulation model of the amorphous oxide semiconductor device includes: The geometry of the amorphous oxide semiconductor device is constructed according to the preset gate length and gate width; An active layer of amorphous oxide semiconductor is set in the geometric structure, and its carrier mobility model, defect state distribution model and oxygen vacancy related physical parameters are defined. By applying bias voltage conditions to the gate, source, and drain regions, and then extracting the corresponding device output drain-source current and internal physical field distribution after simulation, a training dataset is constructed; the bias voltage conditions include the gate-source voltage and the drain-source voltage.

3. The method for predicting the characteristics of amorphous oxide semiconductor devices based on a physical information neural network according to claim 2, characterized in that, The internal physical field distribution includes electric potential distribution, electron concentration distribution, and oxygen vacancy defect concentration.

4. The method for predicting the characteristics of amorphous oxide semiconductor devices based on a physical information neural network according to claim 1 or 2, characterized in that, The training dataset includes: Input vector data includes the spatial coordinates of the simulation mesh nodes, the time points of the transient simulation, the corresponding bias voltage conditions, and the geometric parameters; Microscopic field quantity supervision data, including internal physical field distribution data, is used to construct the initial guidance and regularization constraints for physical information neural networks; Macroscopic characteristic monitoring data, including the output drain-source current of the device under different bias voltage conditions, is used to calibrate the integral characteristics of the model output.

5. The method for predicting the characteristics of amorphous oxide semiconductor devices based on a physical information neural network according to claim 1, characterized in that, The input layer parameters of the physical information neural network include at least: input vector data of the training dataset; the output layer physical field quantities include at least microscopic field quantity supervision data. The microscopic field quantity supervision data includes at least internal physical field distribution data, specifically including potential distribution, electron concentration, and oxygen vacancy defect concentration.

6. The method for predicting the characteristics of amorphous oxide semiconductor devices based on a physical information neural network according to claim 1, characterized in that, The data loss term of the physical information neural network adopts a data loss term based on the simulated output drain-source current, which is used to measure the deviation between the predicted output drain-source current and the simulated output drain-source current of the physical information neural network. The physical loss term includes a physical residual loss term based on the Poisson equation, a physical residual loss term based on the drift-diffusion equation, and a physical residual loss term based on oxygen vacancy trap dynamics. Among them, the physical residual loss term based on the Poisson equation is used to constrain the relationship between the potential distribution and the space charge density inside the device, the physical residual loss term based on the drift-diffusion equation is used to constrain the relationship between the transient carrier concentration distribution and the corresponding current density distribution and its time rate of change, and the physical residual loss term based on oxygen vacancy trap dynamics is used to describe the dynamic process of the generation, recombination and migration of oxygen vacancy concentration as the bias stress, electric field or temperature changes.

7. The method for predicting the characteristics of amorphous oxide semiconductor devices based on a physical information neural network according to claim 1, characterized in that, The implementation process of the differentiable current integration layer is as follows: The spatial derivatives of electrostatic potential and electron concentration are obtained by differential calculation based on electrostatic potential and electron concentration, respectively. The current density along the channel direction was calculated based on the drift-diffusion theory. The predicted terminal current is obtained by numerically integrating the current density along the thickness direction of the amorphous oxide semiconductor device and multiplying it by the gate width.

8. The method for predicting the characteristics of amorphous oxide semiconductor devices based on a physical information neural network according to claim 1, characterized in that, The physical information neural network adopts a multi-layer feedforward fully connected neural network structure, including an input layer, at least three hidden layers, and an output layer; the hidden layers are connected by weighted connections and use continuously differentiable nonlinear activation functions; Training is performed using the backpropagation algorithm and gradient-based optimization methods to minimize the composite loss function, which is a weighted sum of data loss and physical loss terms.

9. A system for predicting the characteristics of amorphous oxide semiconductor devices according to any one of claims 1-8, characterized in that, Includes the following steps: The data acquisition module is used to acquire the bias voltage conditions and geometric parameters of the device; The prediction module is used to predict the electrical characteristics of amorphous oxide semiconductor devices under different geometric parameters or bias voltage conditions by calling a trained physical information neural network and a differential current integration layer, based on the geometric parameters and bias voltage conditions.

10. A computing device, comprising a memory and a processor, characterized in that, The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-8.

Images