Two-dimensional semiconductor device space physical quantity prediction method, device and medium

Abstract

The invention provides a two-dimensional semiconductor device space physical quantity prediction method and device and a medium, and relates to the technical field of semiconductor physical quantity calculation, and the method comprises the steps: inputting parameters of a two-dimensional semiconductor device into a physical coupling representation network model for prediction, and obtaining preliminary space physical quantity distribution data; taking the initial spatial physical quantity distribution data as a high-quality initial solution, inputting the initial solution into a technical computer aided design tool for simulation reasoning, and obtaining final spatial physical quantity distribution data; wherein the physical coupling representation network model is obtained by adopting an alternate training strategy and training of a mixed loss function based on physical information constraint. According to the method, the problem of difficulty in convergence of a numerical solver in a technical computer aided design tool is effectively avoided, the number of simulation iteration steps is remarkably reduced, the simulation calculation time is greatly shortened, and finally, accurate final space physical quantity distribution data is efficiently obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor physical quantity calculation, in particular to a two-dimensional semiconductor device spatial physical quantity prediction method, device and medium. BACKGROUND

[0002] Technical computer-aided design tools, such as TCAD tools, are the core of two-dimensional semiconductor device research and design. They solve control physical equation groups such as Poisson equation and drift-diffusion equation through numerical methods, and can self-consistently calculate the distribution of key spatial physical quantities such as device internal potential and carrier concentration, providing a theoretical basis for device optimization design.

[0003] However, with the continuous evolution of semiconductor device structure and the continuous improvement of complexity, the TCAD simulation calculation cost has increased significantly. The numerical solvers such as Newton-Raphson, which are relied on by TCAD, often face convergence difficulties, prolonged simulation time, or even cannot converge when dealing with strongly coupled partial differential equation groups, which restricts the device design optimization period. SUMMARY

[0004] The problem solved by the present application is the convergence difficulty, prolonged simulation time or even inability to converge of TCAD simulation.

[0005] To solve the above problems, the present application provides a two-dimensional semiconductor device spatial physical quantity prediction method, device and medium.

[0006] In a first aspect, the present application provides a two-dimensional semiconductor device spatial physical quantity prediction method, comprising: inputting parameters of a two-dimensional semiconductor device into a physical coupling representation network model for prediction to obtain preliminary spatial physical quantity distribution data; inputting the preliminary spatial physical quantity distribution data as a high-quality initial solution into a technical computer-aided design tool for simulation reasoning to obtain final spatial physical quantity distribution data; wherein the physical coupling representation network model is trained by using an alternating training strategy and a hybrid loss function based on physical information constraints.

[0007] Optionally, the physical coupling representation network model includes an electron physical property branch, a hole physical property branch and a doping physical property branch arranged in parallel; the electron physical property branch is used to generate electron concentration data explicitly according to the parameters of the two-dimensional semiconductor device; the hole physical property branch is used to generate hole concentration data explicitly according to the parameters of the two-dimensional semiconductor device; the doping physical property branch is used to generate doping concentration data explicitly according to the parameters of the two-dimensional semiconductor device; The preliminary spatial physical quantity distribution data is obtained according to the electron concentration data, the hole concentration data and the doping concentration data.

[0008] Optionally, the alternating training strategy comprises: setting a total number of training cycles and a fixed training round for each network branch in each training cycle; based on the fixed training round, using the hybrid loss function based on the physical information constraint to independently train the electron physical property branch, the hole physical property branch and the doping physical property branch according to the constructed training set and test set, and when updating the network parameters of one of the network branches, the network parameters of the remaining network branches remain unchanged; cycling the above training process until the total number of training cycles is reached to obtain the physical coupling representation network model.

[0009] Optionally, the hybrid loss function based on the physical information constraint comprises a data supervision loss function and a physical equation loss function. The data supervision loss function uses mean square error to determine the data supervision loss function value of the preliminary spatial physical quantity distribution data and the real value of the training set. The physical equation loss function comprises a Poisson equation loss function, a drift-diffusion loss function and a boundary loss function, which are used to constrain the model output to satisfy the semiconductor physical law according to the Poisson loss value of the Poisson equation loss function, the drift loss value of the drift-diffusion loss function and the boundary loss value of the boundary loss function, wherein the Poisson loss value, the drift loss value and the boundary loss value are added and fused to obtain a physical equation loss value.

[0010] Optionally, the hybrid loss value of the hybrid loss function based on the physical information constraint is obtained based on a scaling coefficient from the data supervision loss function value and the physical equation loss value. The hybrid loss function is used to update the network parameters of the electron physical property branch, the hole physical property branch or the doping physical property.

[0011] Optionally, the scaling coefficient determination method comprises: determining a first gradient of the data supervision loss function with respect to the network model parameters and a second gradient of the physical equation loss function with respect to the network model parameters at the current fixed training round; extracting the maximum gradient norm of the first gradient and the average gradient of the second gradient, and determining a scaling factor suggestion value according to the ratio of the maximum gradient norm and the average gradient; Adopting a moving average updating strategy, the scaling factor is obtained according to the scaling factor suggestion value and a historical scaling coefficient in the current fixed training round.

[0012] Optionally, the electronic physical property branch includes a trunk network and a plurality of output networks arranged in parallel, the trunk network includes a plurality of groups of sequentially connected first fully connected layers and first Relu activation layers, the output network includes a plurality of groups of sequentially connected second fully connected layers and second Relu activation layers, and an output end of the last first Relu activation layer is connected to an input end of the first second fully connected layer of each output network.

[0013] Optionally, the construction method of the training set includes: A structure model of a two-dimensional semiconductor device is constructed by using the technical computer-aided design tool; The working condition of the device is set, and the structure model is simulated under different bias voltages to generate a plurality of groups of test distribution data of internal space physical quantities of the device; According to the bias voltage range, the test distribution data is divided to construct the training set and the test set.

[0014] In a second aspect, the present application provides an electronic device, including a memory and a processor; The memory is used for storing a computer program; The processor is used for implementing the two-dimensional semiconductor device space physical quantity prediction method when the computer program is executed.

[0015] In a third aspect, the present application provides a computer readable storage medium, and the storage medium stores a computer program, and when the computer program is executed by a processor, the two-dimensional semiconductor device space physical quantity prediction method is implemented.

[0016] The two-dimensional semiconductor device space physical quantity prediction method, device and medium have the following advantages: The parameters of the two-dimensional semiconductor device are input into the physical coupling representation network model trained by using the alternating training strategy and the hybrid loss function based on physical information constraint for prediction, since the physical information constraint in the hybrid loss function can ensure that the prediction result conforms to the semiconductor physical law, and the alternating training strategy guarantees the stability of model training, so that the model can quickly output the preliminary spatial physical quantity distribution data with physical self-consistency and high precision, thereby providing a reliable basis for subsequent simulation; further, the high-quality preliminary spatial physical quantity distribution data is input into the technical computer-aided design tool as an initial solution for simulation reasoning, since the initial solution is close to the real convergent solution, the convergence difficulty of the numerical solver in the technical computer-aided design tool is effectively avoided, the simulation iteration steps are significantly reduced, the simulation calculation time is greatly shortened, and finally the accurate final spatial physical quantity distribution data is efficiently obtained. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A flowchart of a two-dimensional semiconductor device spatial physical quantity prediction method provided by an embodiment of the present application is shown in the figure. Figure 2 A structure diagram of an electronic physical property branch provided by an embodiment of the present application is shown in the figure. Figure 3 A potential distribution diagram of a MOSFET device in an initial state provided by an embodiment of the present application is shown in the figure. Figure 4 A hole concentration prediction result diagram provided by an embodiment of the present application is shown in the figure. Figure 5 A doping concentration prediction result diagram provided by an embodiment of the present application is shown in the figure. Figure 6 A potential prediction result diagram provided by an embodiment of the present application is shown in the figure. Figure 7 An electron concentration prediction result diagram provided by an embodiment of the present application is shown in the figure. Figure 8 A doping concentration spatial extrapolation error diagram provided by an embodiment of the present application is shown in the figure. Figure 9 A potential spatial extrapolation error diagram provided by an embodiment of the present application is shown in the figure. Figure 10 An electron concentration spatial extrapolation error diagram provided by an embodiment of the present application is shown in the figure. Figure 11 A hole concentration spatial extrapolation error diagram provided by an embodiment of the present application is shown in the figure. Figure 12 A structure diagram of an electronic device provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0018] In order to make the above objectives, features and advantages of the present application more clear and comprehensible, specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms and should not be interpreted as being limited to the embodiments set forth herein, but rather these embodiments are provided for a more thorough and complete understanding of the present application. It should be understood that the drawings and embodiments of the present application are merely for illustrative purposes and are not intended to limit the scope of the present application.

[0019] It should be understood that each of the steps recited in the method embodiments of the present application can be performed in different orders and / or in parallel. In addition, the method embodiments can include additional steps and / or omit the steps shown. The scope of the present application is not limited in this respect.

[0020] As used herein, the term "includes" and its variants are open-ended, meaning "includes but is not limited to"; the term "based on" means "based, at least in part, on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments." Related terms will be understood analogously. It should be noted that reference numerals in the description and drawings indicate embodiments of the application and do not necessarily indicate any priority or dependency between the embodiments. It should be noted that the concepts of "first", "second", etc. mentioned in the present application are merely used to distinguish different apparatuses, modules or units, and do not imply the order or interdependence of the functions performed by these apparatuses, modules or units.

[0021] It should be noted that the modification of "one" or "more" mentioned in the present application is illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise explicitly stated in the context, it should be understood as "one or more".

[0022] The names of the messages or information exchanged between the plurality of apparatuses in the embodiments of the present application are merely for illustrative purposes and are not intended to limit the scope of the messages or information.

[0023] To solve the convergence difficulty, prolong the simulation time, and even the situation that cannot converge, the related technology proposes a pure data-driven machine learning surrogate model, which can quickly predict device characteristics, but requires a large amount of training data, and lacks physical knowledge constraints, and the prediction result may violate the semiconductor physical law, and the physical self-consistency and reliability are insufficient. The related technology also proposes a physical information machine learning (PIML) technology, which integrates the control physical equation and boundary condition as a constraint into neural network training, which not only reduces the dependence on large-scale labeled data, but also guarantees the physical consistency of the prediction result, and shows great potential. However, the PIML technology needs to calculate the high-order spatial derivative of the predicted quantity through automatic differentiation, and the network parameters are not converged at the beginning of training, and the prediction accuracy of the physical quantity is poor, resulting in significant noise in the high-order derivative. These noisy physics-driven gradients will conflict with the data supervision loss function gradient, making the model training unstable, and it needs hundreds of thousands of iterations to converge to a higher precision, which greatly limits its practicality in TCAD acceleration tasks that require fast iteration.

[0024] To solve the problems in the above related technology, the embodiment provides a two-dimensional semiconductor device space physical quantity prediction method, device and medium.

[0025] As shown in Figure 1 The two-dimensional semiconductor device space physical quantity prediction method provided by the embodiment of the application comprises: The parameters of the two-dimensional semiconductor device are input into a physical coupling representation network model for prediction to obtain preliminary spatial physical quantity distribution data.

[0026] Specifically, the related parameters (such as spatial coordinates, bias voltage, etc.) of the two-dimensional semiconductor device are input into a trained physical coupling representation network model, and the model directly predicts the preliminary spatial physical quantity distribution data, such as the distribution of key physical fields such as electric potential, electron concentration and hole concentration in the device. The physical coupling representation network model is not an ordinary neural network, but a semiconductor physical law is fused in the structure design and training strategy, the output has clear physical meaning, and the internal coupling relationship between different physical quantities can be maintained.

[0027] The physical coupling representation network model is trained by using an alternating training strategy and a hybrid loss function based on physical information constraints.

[0028] Specifically, the physical coupling representation network model is obtained through joint optimization using an alternating training strategy and a hybrid loss function based on physical information constraints. The alternating training strategy is based on the Gummel iteration idea in semiconductor device simulation, which updates the parameters of different physical branches sequentially during training to avoid optimization conflicts caused by multi-physics coupling. The hybrid loss function simultaneously includes data supervision terms (i.e., the error between the predicted value and the TCAD true value) and physical equation residual terms (such as the degree of violation of the Poisson equation, drift-diffusion equation, and boundary conditions), and balances the gradient magnitudes of the two through dynamically adjusted scaling factors to ensure that the model achieves a good balance between data fitting ability and physical consistency.

[0029] The preliminary spatial physical quantity distribution data is used as a high-quality initial solution and input into a computer-aided design tool for simulation and reasoning to obtain the final spatial physical quantity distribution data.

[0030] Specifically, the preliminary spatial physical quantity distribution data obtained from the above predictions are used as a high-quality initial solution and input into a Technical Computer-Aided Design (TCAD) tool to activate its internal numerical solver (such as the Newton-Raphson iteration method) for simulation inference, thereby quickly converging to obtain high-precision final spatial physical quantity distribution data. Since the initial solution is highly close to the true solution, the number of iterations and computation time of the TCAD simulation are significantly reduced, effectively alleviating the convergence difficulties caused by poor initial guesses in traditional simulations.

[0031] In this embodiment, the parameters of the two-dimensional semiconductor device are input into a physically coupled representation network model trained using an alternating training strategy and a hybrid loss function based on physical information constraints for prediction. Since the physical information constraints in the hybrid loss function ensure that the prediction results conform to the laws of semiconductor physics, and the alternating training strategy ensures the stability of the model training, the model can quickly output preliminary spatial physical quantity distribution data with physical consistency and high accuracy, providing a reliable foundation for subsequent simulations. Furthermore, this high-quality preliminary spatial physical quantity distribution data is used as the initial solution and input into a computer-aided design tool for simulation inference. Since the initial solution is close to the true convergent solution, the problem of convergence difficulty of numerical solvers in computer-aided design tools is effectively avoided, significantly reducing the number of simulation iteration steps and greatly shortening the simulation calculation time, ultimately obtaining accurate final spatial physical quantity distribution data efficiently.

[0032] Optionally, the physical coupling characterization network model includes parallel branches for electronic physical characteristics, hole physical characteristics, and doping physical characteristics; The electronic physics branch is used to explicitly generate electron concentration data based on the parameters of the two-dimensional semiconductor device; The hole physical property branch is configured to explicitly generate hole concentration data according to parameters of the two-dimensional semiconductor device; The doping physical property branch is configured to explicitly generate doping concentration data according to parameters of the two-dimensional semiconductor device. The preliminary spatial physical quantity distribution data is obtained according to the electron concentration data, the hole concentration data and the doping concentration data.

[0033] Specifically, explicit means that each dimension in the output vector of the neural network branch corresponds to a specific physical quantity or its derivative quantity with a clear physical meaning and can be directly explained, rather than an abstract feature (without a clear physical meaning and only used for model internal fitting) of a “black box”. These physical quantities are directly outputted without the need for indirect acquisition through subsequent complex derivation (such as automatic differentiation). The physical coupling representation network model adopts a three-branch parallel architecture, including an electron physical property branch, a hole physical property branch and a doping physical property branch. The three branches share the same input parameters, i.e., parameters of the two-dimensional semiconductor device, such as spatial coordinates and bias voltage. The electron physical property branch is specially designed for modeling electron-related physical properties. Through feature extraction and mapping of the network, electron concentration data is explicitly outputted, which directly reflects the spatial distribution state of electrons inside the device. The hole physical property branch focuses on hole physical properties. Based on the same input parameters, hole concentration data is explicitly generated to accurately represent the spatial distribution of holes. The doping physical property branch is designed for the doping characteristics of the device. Based on the input device parameters, doping concentration data is explicitly outputted to reflect the spatial distribution of doping inside the device. The electron concentration data includes an electron concentration auxiliary potential function and its first-order and second-order spatial partial derivatives, an electron concentration and its first-order and second-order spatial partial derivatives, and the Laplacian of the electron concentration auxiliary potential function is equal to the electron concentration. The hole concentration data includes a hole concentration auxiliary potential function and its first-order and second-order spatial partial derivatives, a hole concentration and its first-order and second-order spatial partial derivatives, and the Laplacian of the hole concentration auxiliary potential function is equal to the hole concentration. The doping concentration data includes a doping concentration auxiliary potential function and its first-order and second-order spatial partial derivatives, a doping concentration and its first-order and second-order spatial partial derivatives, and the Laplacian of the doping concentration auxiliary potential function is equal to the electron doping. In addition, the preliminary spatial physical quantity distribution data includes an electric potential, an electric field and a Laplacian of the electric potential. The preliminary spatial physical quantity distribution data is obtained by substituting and combining the electron concentration data, the hole concentration data and the doping concentration data, i.e., the electric potential is composed of the auxiliary potential functions outputted by the branches, the electric field is composed of the first-order partial derivatives of the auxiliary potential functions outputted by the branches, and the Laplacian of the electric potential is composed of the second-order partial derivatives of the auxiliary potential functions outputted by the branches.

[0034] In this embodiment, the acquisition of the preliminary spatial physical quantity distribution data is based on the output results of the three branches. By correlating and integrating the electron concentration data, hole concentration data and doping concentration data, preliminary data containing core physical quantity distribution information is directly obtained. The design of the three parallel branches makes the modeling of each physical quantity independent and correlated with each other. Independent modeling ensures the pertinence of the representation of each physical quantity, and the correlation integration provides a foundation for the subsequent guarantee of physical self-consistency. By explicitly generating each key concentration data, the ambiguity of the physical quantity representation by the black box model is avoided. At the same time, the three-branch parallel architecture can fully adapt to the characteristics of different physical quantities, improve the accuracy of data generation, and lay a reliable data foundation for the subsequent application of physical constraints and model training.

[0035] Optionally, the alternating training strategy comprises: setting a total number of training cycles and a fixed training round for each network branch in each training cycle; based on the fixed training round, using the hybrid loss function based on the physical information constraint to independently train the electron physical property branch, the hole physical property branch and the doping physical property branch according to the constructed training set and test set, and when updating the network parameters of one of the network branches, the network parameters of the remaining network branches remain unchanged; repeating the above training process until the total number of training cycles is reached to obtain the physical coupling representation network model.

[0036] Specifically, first, according to the model training requirements and device characteristics, the total number of training cycles is preset, and the fixed training rounds in each training cycle are set for the electron physical characteristic branch, the hole physical characteristic branch and the doping physical characteristic branch respectively. The setting of the fixed training rounds needs to take into account the modeling complexity and training efficiency of each branch. Subsequently, based on the constructed training set and test set, a hybrid loss function based on physical information constraints is used to independently train the three network branches: when training the electron physical characteristic branch, the network parameters of the hole physical characteristic branch and the doping physical characteristic branch are frozen, and only the parameters of the electron physical characteristic branch are calculated and updated. After completing the fixed training rounds of the electron physical characteristic branch, the branch parameters are frozen, and the hole physical characteristic branch and the doping physical characteristic branch are trained in turn according to the same logic. Finally, the above independent training process is executed in a loop until the preset total number of training cycles is reached, and the trained physical coupling representation network model is finally obtained. In this training strategy, the fixed training rounds provide a clear execution standard for the independent training of each branch, and the parameter freezing mechanism ensures the independence of each branch during training, avoiding parameter interference between different branches. The beneficial effects are that by referring to the core idea of the Gummel method, the precise training of each branch is realized through alternating optimization, which not only ensures that each branch can focus on modeling the corresponding physical quantity, but also indirectly cooperates between branches through cyclic updating, effectively avoiding the parameter conflict that may be caused by simultaneous training of multiple branches, and improving the stability and convergence efficiency of model training.

[0037] Optionally, the hybrid loss function based on physical information constraints includes a data supervised loss function and a physical equation loss function. The data supervised loss function adopts mean square error, and is used to determine the data supervised loss function value of the preliminary spatial physical quantity distribution data and the training set true value. The physical equation loss function includes a Poisson equation loss function, a drift diffusion loss function and a boundary loss function, and is used to constrain the model output to satisfy the semiconductor physical law according to the Poisson loss value of the Poisson equation loss function, the drift loss value of the drift diffusion loss function and the boundary loss value of the boundary loss function, wherein the Poisson loss value, the drift loss value and the boundary loss value are added and fused to obtain a physical equation loss value.

[0038] Specifically, the hybrid loss function based on physical information constraints is composed of a data supervision loss function and a physical equation loss function. The data supervision loss function adopts mean square error as the calculation method, and specifically calculates the mean square error between the preliminary spatial physical quantity distribution data output by the physical coupling representation network and the corresponding true value in the training set. The error is the data supervision loss function value, and its core function is to ensure the consistency of the model prediction result and the true data. The physical equation loss function is further divided into a Poisson equation loss function, a drift diffusion loss function and a boundary loss function. The Poisson equation loss function calculates the residual norm of the model output data substituted into the Poisson equation, for example, the residual norm of the Laplacian of the potential and the electron concentration, to obtain the Poisson loss value. The drift diffusion loss function calculates the residual norm of the model output data substituted into the drift diffusion equation, for example, the residual norm of the electron concentration data, the hole concentration data, the doping concentration data and the Laplacian of the potential, to obtain the drift loss value. The boundary loss function calculates the normal current density of the electron and hole of the model output data under the device boundary condition, for example, the normal current density of the electron concentration data, the hole concentration data, the doping concentration data and the preliminary spatial physical quantity distribution data, and constrains it to be zero to obtain the boundary loss value. The Poisson loss value, the drift loss value and the boundary loss value are directly added to obtain the physical equation loss value. The core function of the loss value is to constrain the model output to strictly follow the semiconductor physical law. The data supervision loss function and the physical equation loss function cooperate with each other. The data supervision loss function ensures the fitting accuracy of the model, and the physical equation loss function ensures the physical compliance of the model. The two together constitute the core constraint of the model training. Its beneficial effects are that it effectively makes up for the defects of the pure data-driven model lacking physical constraints, and at the same time avoids the problem of insufficient fitting accuracy of the pure physical-driven model, so that the model learns the data law while strictly following the physical law, significantly improving the accuracy and physical self-consistency of the prediction result.

[0039] Illustratively, the data supervision loss function adopts four mean square errors to respectively punish the deviation between the potential, electron concentration, hole concentration and doping concentration in the foregoing and the true value.

[0040] Optionally, the hybrid loss value of the hybrid loss function based on physical information constraints is obtained based on a scaling coefficient from the data supervision loss function value and the physical equation loss value. The hybrid loss function is used to update the network parameters of the electron physical property branch, the hole physical property branch or the doping physical property.

[0041] Specifically, the calculation of the hybrid loss value based on the physical information constraint needs to introduce a scaling coefficient, specifically by adding the data supervised loss function value and the physical equation loss value after multiplying them by the scaling coefficient respectively, wherein the value of the scaling coefficient needs to be dynamically adjusted according to the gradient in the training process. The hybrid loss value is the core basis for network parameter update, and when training the electronic physical property branch, the hole physical property branch or the doping physical property branch, the hybrid loss value is taken as the target, the gradient is calculated through the back propagation algorithm, and the network parameters of the corresponding branch are updated. The scaling coefficient plays a key balancing role in the fusion process of the two losses, which is directly related to the calculation of the hybrid loss value, and the hybrid loss value directly guides the update direction of the network parameters, so that the parameter update can respond to the demand of data fitting and comply with the requirements of physical constraints. Its beneficial effects are that the reasonable fusion of the data supervised loss and the physical equation loss is realized through the scaling coefficient, the model bias caused by the dominance of a single loss in the training process is avoided, the model is ensured to consider both data rule learning and physical law compliance in the training process, and the prediction accuracy and stability of the model are improved.

[0042] Optionally, the method for determining the scaling coefficient comprises: determining a first gradient of the data supervised loss function with respect to the network model parameters and a second gradient of the physical equation loss function with respect to the network model parameters under the current fixed training round, respectively; extracting the maximum gradient norm of the first gradient and the average gradient of the second gradient, respectively, and determining a scaling factor suggestion value according to the ratio of the maximum gradient norm and the average gradient; using a sliding average update strategy to obtain the scaling coefficient under the current fixed training round according to the scaling factor suggestion value and the historical scaling coefficient.

[0043] Specifically, first, the first gradient of the data supervision loss function with respect to the network model parameters and the second gradient of the physical equation residual loss with respect to the network model parameters are calculated under the current fixed training round, and the gradient calculation is based on the back propagation algorithm to ensure the accuracy of the gradient information. Subsequently, the maximum gradient norm is extracted from the first gradient, which is taken as the data item gradient strength indicator, which can intuitively reflect the influence strength of the data supervision loss on parameter update; at the same time, the average gradient is extracted from the second gradient, which is taken as the physical item gradient strength indicator, representing the influence strength of the physical equation loss on parameter update, and then the ratio of the data item gradient strength indicator to the physical item gradient strength indicator is calculated, which is the scaling factor suggestion value of the current iteration step. Finally, the sliding average update strategy is adopted, and the preset momentum decay factor is combined to weight and sum the scaling factor suggestion value of the current iteration step and the historical scaling coefficient, to obtain the final scaling coefficient under the current fixed training round. In the next fixed training round, the above steps are repeated to obtain the scaling coefficient of the next fixed training round. In this process, gradient calculation provides direct data support for scaling factor suggestion value, and sliding average update strategy realizes the smooth update of scaling coefficient by weighted fusion of historical scaling coefficient and scaling factor suggestion value of the current iteration step, avoiding the coefficient mutation caused by single gradient fluctuation, and ensuring the smoothness and stability of scaling coefficient update, each step is progressive and logically related, ensuring that the scaling coefficient can dynamically adapt to the training process. The beneficial effects are that by dynamically calculating the scaling coefficient, the gradient strength of the data supervision loss and the physical equation loss can be balanced in real time, effectively avoiding the gradient conflict in the training process, keeping the model training in a stable direction, and significantly improving the convergence speed of the model.

[0044] Optionally, as shown in Figure 2 The electronic physical property branch includes a trunk network and a plurality of parallel output networks, the trunk network includes a plurality of groups of sequentially connected first fully connected layers and first Relu activation layers, the output network includes a plurality of groups of sequentially connected second fully connected layers and second Relu activation layers, and the output end of the last first Relu activation layer is connected with the input end of the first second fully connected layer of each output network.

[0045] Specifically, the structure design of the electronic physical property branch has a clear hierarchical division, specifically including a trunk network and a plurality of parallel output networks. The trunk network is composed of a plurality of sequentially connected first fully connected layers and first Relu activation layers, each group of first fully connected layers is responsible for linear transformation of features, and the first Relu activation layer introduces a nonlinear mapping capability, and through a plurality of sequentially connected groups, deep feature extraction of input parameters (spatial coordinates and bias voltage) is realized; the number of output networks is set according to actual output requirements, and each output network is composed of a plurality of sequentially connected second fully connected layers and second Relu activation layers, which are used to further map the core features extracted by the trunk network into specific output data. The output end of the last first Relu activation layer in the trunk network is connected to the input end of the first second fully connected layer of each output network one by one, forming a feature transmission path of "trunk extraction-branch generation". The structure design of the trunk network and the output network cooperate with each other, and the deep feature extraction of the trunk network provides a basis for the accurate generation of the output network, and the parallel setting of the output network meets the demand of multi-dimensional output. Its beneficial effects lie in that through the hierarchical progressive feature extraction and generation mechanism, the representation ability of the electronic physical property branch for complex physical laws is significantly improved, and multi-dimensional data related to electron concentration can be accurately output, providing high-quality basic data for subsequent combination of physical quantities and loss calculation.

[0046] Exemplarily, the structures of the electronic physical property branch, the hole physical property branch and the doping physical property branch are the same, all including a trunk network and a plurality of parallel output networks, the trunk network including a plurality of sequentially connected first fully connected layers and first Relu activation layers, the output network including a plurality of sequentially connected second fully connected layers and second Relu activation layers, and the output end of the last first Relu activation layer is connected to the input end of the first second fully connected layer of each output network, but the network parameters of the electronic physical property branch, the hole physical property branch and the doping physical property branch are different to realize different output data.

[0047] Optionally, the construction method of the training set comprises: adopting the technical computer-aided design tool to construct a structure model of a two-dimensional semiconductor device; setting the working conditions of the device, and performing scanning simulation on the structure model under different bias voltages to generate a plurality of groups of test distribution data of internal spatial physical quantities of the device; dividing the test distribution data according to the bias voltage range to construct the training set and the test set.

[0048] Specifically, first, an industry-standard technical computer-aided design (TCAD) tool is used to model the structure of a target two-dimensional semiconductor device, and the geometric structure (such as a silicon substrate, an oxide layer, a source, a gate, a drain, and a substrate electrode) and the doping distribution (such as Gaussian doping of a P-type substrate and N-type source and drain regions) of the device are accurately defined during the modeling process to ensure consistency between the structural model and the actual device. Subsequently, the working conditions of the device (such as a fixed drain voltage) are set, and a simulation is performed within a range of different bias voltages. A plurality of sets of experimental distribution data of internal spatial physical quantities of the device are generated through the simulation, covering spatial distribution information of key physical quantities such as electric potential, electron concentration, hole concentration, and doping concentration. Finally, the generated experimental distribution data is divided into a training set and a test set according to a preset bias voltage range division rule, for example, experimental data in a low voltage range is used as the training set, and experimental data in a high voltage range is used as the test set, to effectively verify the bias voltage extrapolation capability of the model. The construction of the structural model provides a reliable foundation for subsequent scanning simulation, and the experimental data generated by the scanning simulation provides an abundant data source for the construction of the data set. The division method of the bias voltage range directly determines the effectiveness of the training set and the test set. The beneficial effects are that the experimental data generated by the TCAD tool has high accuracy and authenticity, and combined with the scientific division of the bias voltage range, a training set and a test set covering different working conditions and having representativeness and diversity can be constructed, providing solid data support for the training and verification of the model, and effectively improving the generalization ability and voltage extrapolation performance of the model.

[0049] Exemplarily, to verify the effectiveness of the two-dimensional semiconductor device spatial physical quantity prediction method based on physical information, a two-dimensional MOSFET device is taken as the research object, and a device model is constructed by using a Silvaco ATLAS TCAD tool. The device model includes a silicon substrate, an oxide layer, a source, a gate, a drain, and a substrate electrode, has a length of 1.5 microns along the X-axis and a length of 2.00 microns along the Y-axis, uses a uniform grid spacing of 0.02 microns, and has a doping distribution of a P-type substrate and N-type source and drain regions. Thirty groups of simulation samples are generated by fixing the drain voltage and scanning the gate voltage from 1.0 V to 4.0 V at a step of 0.1 V, of which 21 groups of samples from 1.0 V to 3.0 V are used as the training set, and 9 groups of samples from 3.1 V to 4.0 V are used as the test set. Each group of samples includes spatial distribution data of electric potential, electron concentration, hole concentration, and doping concentration at grid points in the device. The electric potential distribution of the MOSFET device in the initial state is shown in FIG. 1, Figure 3 Figure 3 where the horizontal axis is the X-axis, the vertical axis is the Y-axis, and Potential is the electric potential. The extrapolation prediction capability of the model for unseen bias voltages is evaluated by the test set, and the voltage extrapolation prediction results of the model for the four spatial physical quantities are shown in FIG. 2, Figures 4 to 7 ​As shown in the table, the experimental results show that the root mean square error and the normalized root mean square error of the model on all physical quantity predictions are low, proving that the method has high spatial physical quantity prediction accuracy in the case of voltage extrapolation; further, a high-resolution test set with a grid spacing of 0.01 microns in the X direction (higher than the training data of 0.02 microns) is used at a voltage of 3.1V to evaluate the generalization ability of the model to unseen spatial coordinates (different grid resolutions), and the results are as shown in the table Figures 8 to 11 As shown in the table, the prediction error of the physically coupled representation network is small when facing finer grids, showing excellent stability and spatial generalization ability, and this result also proves that the method of the application truly learns the internal relationship and gradient information of the physical field in the continuous spatial domain through its physically coupled architecture and mixed loss function, rather than just fitting discrete grid point data.

[0050] As shown in the table, the prediction error of the physically coupled representation network is small when facing finer grids, showing excellent stability and spatial generalization ability, and this result also proves that the method of the application truly learns the internal relationship and gradient information of the physical field in the continuous spatial domain through its physically coupled architecture and mixed loss function, rather than just fitting discrete grid point data. Figure 12 As shown in the table, the prediction error of the physically coupled representation network is small when facing finer grids, showing excellent stability and spatial generalization ability, and this result also proves that the method of the application truly learns the internal relationship and gradient information of the physical field in the continuous spatial domain through its physically coupled architecture and mixed loss function, rather than just fitting discrete grid point data.

[0051] Alternatively, an electronic device includes a memory and a processor coupled to the memory; the memory is configured to store a computer program; the processor is configured to execute the computer program to perform the following operations: inputting parameters of the two-dimensional semiconductor device into a physically coupled representation network model for prediction to obtain preliminary spatial physical quantity distribution data; inputting the preliminary spatial physical quantity distribution data as high-quality initial solution into a technical computer-aided design tool for simulation reasoning to obtain final spatial physical quantity distribution data; The physically coupled representation network model is trained by using an alternating training strategy and a mixed loss function based on physical information constraints.

[0052] The computer readable storage medium provided by the embodiment of the application stores a computer program, and when the computer program is executed by a processor, the two-dimensional semiconductor device spatial physical quantity prediction method described above is realized.

[0053] Alternatively, a non-volatile computer readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor performs the following operations: inputting parameters of the two-dimensional semiconductor device into a physically coupled representation network model for prediction to obtain preliminary spatial physical quantity distribution data; The preliminary spatial physical quantity distribution data is input into a technical computer-aided design tool as a high-quality initial solution to perform simulation reasoning to obtain final spatial physical quantity distribution data. The physical coupling representation network model is trained by using an alternating training strategy and a hybrid loss function based on physical information constraints.

[0054] Electronic devices that can be used as servers or clients of the application will now be described, which are examples of hardware devices that can be applied to aspects of the application. The electronic devices are intended to represent a broad range of digital electronic computing devices, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computing devices. The electronic devices can also represent a broad range of mobile devices, such as personal digital assistants, cellular telephones, smart phones, wearable devices, and other similar computing devices. The components, their connections, and their functions, as described herein, are meant to be examples only, and are not intended to limit the implementations of the application described and / or claimed in this document.

[0055] The electronic device includes a computing unit that can perform various appropriate actions and processes in accordance with a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). Various programs and data required for device operation can also be stored in the RAM. The computing unit, the ROM, and the RAM are connected to each other by a bus. An input / output (I / O) interface is also connected to the bus.

[0056] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiment methods. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM), or a random access memory (RAM). In this application, the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment of the application. In addition, the functional units in each embodiment of the application can be integrated into one processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0057] Although the present application has been disclosed with reference to the above embodiments, the scope of the present application is not limited to the above. Various changes and modifications can be made to the present application without departing from the spirit and scope thereof, and such changes and modifications are intended to fall within the scope of the present application.

Claims

1. A method of predicting a spatial physical quantity of a two-dimensional semiconductor device, characterized by, The method comprises the following steps: inputting parameters of a two-dimensional semiconductor device into a physical coupling characterization network model for prediction to obtain preliminary spatial physical quantity distribution data; inputting the preliminary spatial physical quantity distribution data as a high-quality initial solution into a technical computer-aided design tool for simulation reasoning to obtain final spatial physical quantity distribution data; wherein the physical coupling characterization network model is trained by using an alternating training strategy and a hybrid loss function based on physical information constraints.

2. The two-dimensional semiconductor device space physical quantity prediction method according to claim 1, characterized by, The physical coupling characterization network model comprises an electron physical property branch, a hole physical property branch and a doping physical property branch arranged in parallel; The electron physical property branch is used to explicitly generate electron concentration data according to the parameters of the two-dimensional semiconductor device; The hole physical property branch is used to explicitly generate hole concentration data according to the parameters of the two-dimensional semiconductor device; The doping physical property branch is used to explicitly generate doping concentration data according to the parameters of the two-dimensional semiconductor device; wherein the preliminary spatial physical quantity distribution data is obtained according to the electron concentration data, the hole concentration data and the doping concentration data.

3. The two-dimensional semiconductor device space physical quantity prediction method according to claim 2, characterized by, The alternating training strategy comprises: setting the total number of training cycles and the fixed training rounds for each network branch in each training cycle; based on the fixed training rounds, using the hybrid loss function based on physical information constraints to independently train the electron physical property branch, the hole physical property branch and the doping physical property branch according to the constructed training set and test set, and when updating the network parameters of one of the network branches, the network parameters of the remaining network branches remain unchanged; the above training process is repeated until the total number of training cycles is reached to obtain the physical coupling characterization network model.

4. The two-dimensional semiconductor device space physical quantity prediction method according to claim 3, characterized by, The hybrid loss function based on physical information constraints comprises a data supervision loss function and a physical equation loss function; The data supervision loss function uses mean square error to determine the data supervision loss function value of the preliminary spatial physical quantity distribution data and the real value of the training set; The physical equation loss function comprises a Poisson equation loss function, a drift-diffusion loss function and a boundary loss function, which are used to constrain the model output to satisfy the semiconductor physical law according to the Poisson loss value of the Poisson equation loss function, the drift loss value of the drift-diffusion loss function and the boundary loss value of the boundary loss function, wherein the Poisson loss value, the drift loss value and the boundary loss value are added and fused to obtain a physical equation loss value.

5. The two-dimensional semiconductor device space physical quantity prediction method according to claim 4, characterized by, The hybrid loss value of the hybrid loss function based on physical information constraints is obtained based on a scaling coefficient from the data supervision loss function value and the physical equation loss value; The hybrid loss function is used to update the network parameters of the electron physical property branch, the hole physical property branch or the doping physical property.

6. The two-dimensional semiconductor device space physical quantity prediction method according to claim 5, characterized by, The method for determining the scaling coefficient comprises: determining a first gradient of the data supervision loss function with respect to the network model parameters and a second gradient of the physical equation loss function with respect to the network model parameters at the current fixed training round; Extract the maximum gradient norm of the first gradient and the average gradient of the second gradient respectively, and determine the suggested value of the scaling factor based on the ratio of the maximum gradient norm to the average gradient. A sliding average update strategy is adopted to obtain the scaling factor under the current fixed training round based on the proposed scaling factor value and the historical scaling factor.

7. The two-dimensional semiconductor device space physical quantity prediction method according to claim 2, characterized by, The electronic physical characteristic branch includes a backbone network and multiple parallel output networks. The backbone network includes multiple sets of sequentially connected first fully connected layers and first ReLU activation layers. The output network includes multiple sets of sequentially connected second fully connected layers and second ReLU activation layers. The output terminal of the last first ReLU activation layer is connected to the input terminal of the first second fully connected layer of each output network.

8. The two-dimensional semiconductor device space physical quantity prediction method according to claim 3, characterized by, The methods for constructing the training set include: The aforementioned technology is used to construct structural models of two-dimensional semiconductor devices using computer-aided design tools. The operating conditions of the device are set, and the structural model is scanned and simulated under different bias voltages to generate multiple sets of experimental distribution data of the internal spatial physical quantities of the device. Based on the bias voltage range, the experimental distribution data is divided to construct the training set and the test set.

9. An electronic device, comprising: Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the method for predicting spatial physical quantities of two-dimensional semiconductor devices as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the spatial physical quantity prediction method for two-dimensional semiconductor devices as described in any one of claims 1 to 7.

Images