An acceleration method for high-precision two-dimensional potential simulation of semiconductor devices
By combining TCAD and neural network methods, a two-dimensional potential prediction model of self-attention mechanism or deep reverse projection network is established, which solves the problems of long simulation time and insufficient accuracy in traditional simulation technology, and achieves fast and high-precision simulation of semiconductor devices , meet industrial application needs.
Patent Information
- Application Number
- CN202411324053.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-23
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-09-23
AI Technical Summary
Traditional semiconductor device simulation technology has problems of long simulation time and high computing cost when dealing with complex devices. Although the deep learning neural network method accelerates the simulation speed, it is difficult to meet the needs of high-precision industrial applications, especially in complex device design and advanced materials. Application.
Combining traditional TCAD simulation methods and neural network prediction methods, by establishing a two-dimensional potential prediction model of deep convolutional neural network or deep reverse projection network that introduces self-attention mechanism, the TCAD numerical method is used to iterate the high-precision two-dimensional potential distribution. , realize fast and high-precision simulation.
It achieves accurate two-dimensional potential distribution in a short time, significantly speeding up the design and optimization process of semiconductor devices, faster speed and higher accuracy, and can meet the needs of high standards of industrial applications.
Smart Images

Figure CN118839622B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of semiconductor device electrical characteristic simulation, and in particular relates to an acceleration method for high-precision two-dimensional electric potential simulation of semiconductor devices. Background Art
[0002] In today's era of rapid technological development, semiconductor technology undoubtedly plays an important role. As the cornerstone of the modern electronics industry, semiconductor technology has penetrated deeply into various fields, such as communications, computing, energy, transportation, and medical care. It has not only promoted the development of information technology, but also brought unprecedented convenience to human life.
[0003] With the continuous development of semiconductor technology, the electrical characteristics of semiconductor devices have played an increasingly important role in the design and optimization of semiconductor devices. Designers hope to obtain the electrical characteristics of semiconductor devices in the early stages of design, so as to adjust the structural parameters and process parameters to achieve the design goals of the device. Among them, the two-dimensional potential distribution of semiconductor devices is an important basis for designers to design device structural parameters and process parameters.
[0004] Traditionally, computer-aided design tools (Technology Computer Aided Design, TCAD for short) are widely used in the design and simulation process of semiconductor devices. However, despite the significant progress made in the past few decades, TCAD technology still faces problems such as long simulation time and high computing cost when dealing with complex semiconductor device simulations, which is particularly prominent in the current scientific research and industrial environment that pursues rapid iteration and high efficiency.
[0005] In recent years, neural network technology based on deep learning has provided a new solution for the simulation of semiconductor devices. Neural network technology can predict the potential distribution of semiconductor devices and achieve fast simulation. Although this method greatly speeds up the simulation speed, its accuracy is usually difficult to meet the high-standard industrial application requirements, especially in complex device design and advanced material applications. Summary of the invention
[0006] Purpose of the invention: In order to solve the above-mentioned deficiencies in the prior art, the present invention proposes an acceleration method for high-precision two-dimensional potential simulation of semiconductor devices. This method successfully combines the high precision of the traditional TCAD simulation method and the high efficiency of the neural network prediction method, which can greatly improve the simulation speed while maintaining the prediction accuracy comparable to the traditional method. Through this method, an accurate two-dimensional potential distribution can be obtained in a short time, thereby greatly accelerating the design and optimization process of semiconductor devices.
[0007] Technical solution: A method for accelerating high-precision two-dimensional potential simulation of a semiconductor device of the present invention comprises the following steps:
[0008] Step 1: Obtaining data sets: According to the type of semiconductor device, determine the value space of the simulation model and device structure parameters, and obtain the data set based on TCAD simulation;
[0009] Step 2: Establish a two-dimensional potential prediction model: Based on the data set obtained in step 1, establish a two-dimensional potential prediction model of a deep convolutional neural network or a deep back projection network that introduces a self-attention mechanism; the two-dimensional potential prediction model can predict the two-dimensional potential based on the structural parameters of the semiconductor device;
[0010] Step 3: predicting the approximate value of the two-dimensional electric potential: inputting the structural parameters of the semiconductor device to be simulated into the two-dimensional electric potential prediction model established in step 2, and predicting the approximate value of the two-dimensional electric potential corresponding to the semiconductor device to be simulated;
[0011] Step 4: High-precision two-dimensional electric potential simulation: The two-dimensional electric potential approximation predicted in step 3 is used as the initial release into the TCAD software, and the numerical method is iterated to obtain the high-precision two-dimensional electric potential corresponding to the semiconductor device to be simulated.
[0012] Furthermore, the simulation model includes a drift-diffusion model, a fluid dynamics model and a Monte Carlo model electrical performance model. The grid types of the simulation model include a uniform grid and a non-uniform grid. The semiconductor device is a field effect transistor. For a MOSFET field effect transistor, its structural parameters include: gate length , oxide layer thickness , source and drain doping depth , source-drain doping concentration , substrate doping concentration , gate voltage , drain voltage .
[0013] Furthermore, in step 1, obtaining a data set based on TCAD simulation includes the following steps:
[0014] Step 11, determining structural parameters and their value space: according to the type of semiconductor device to be simulated, selecting structural parameters related to its two-dimensional potential, and determining the value space of the structural parameters;
[0015] Step 12: Determine the mesh type of the simulation model: set the mesh in TCAD and choose to set a uniform mesh or a non-uniform mesh;
[0016] Step 13, obtaining a two-dimensional potential distribution: obtaining a two-dimensional potential distribution of semiconductor devices with different structural parameters through TCAD simulation;
[0017] Step 14, obtaining a two-dimensional electric potential training label: interpolating the two-dimensional electric potential distribution grid obtained in step 13 into a 64*64 two-dimensional electric potential training grid by an interpolation sampling method;
[0018] Step 15, forming a data set: the structural parameters of each semiconductor device in step 11 are matched one by one with the two-dimensional potential training grid obtained in step 14, so as to form a data set from structural parameters to two-dimensional potential distribution.
[0019] Furthermore, in step 2, establishing a two-dimensional potential prediction model specifically includes the following steps:
[0020] Step 21, data set classification: the data set obtained in step 1 is divided into a training set, a validation set and a test set according to a set ratio; wherein the training set is used to train the two-dimensional electric potential prediction model; the validation set is used to adjust the weight value in the training process; and the test set is used to test the two-dimensional electric potential prediction model;
[0021] Step 22, data processing: standardize or normalize the data in the training set, validation set, and test set;
[0022] Step 23, establishing a two-dimensional potential prediction model: designing a deep convolutional neural network or a deep back projection network that introduces a self-attention mechanism to establish a two-dimensional potential prediction model; the output parameter of the two-dimensional potential prediction model is the two-dimensional potential distribution matrix of the semiconductor device to be simulated; the input parameter of the two-dimensional potential prediction model is the structural parameter of the semiconductor device to be simulated;
[0023] Step 24, model training: using the data in the training set in step 22 to train the two-dimensional electric potential prediction model established in step 23, to obtain parameters of the two-dimensional electric potential prediction model;
[0024] Step 25, model verification and testing: using the data in the verification set or test set in step 22, the two-dimensional potential prediction model trained in step 24 is verified or tested.
[0025] Furthermore, in step 2, the deep convolutional neural network of the self-attention mechanism is specifically as follows: a deep convolutional neural network is added in the form of a self-attention layer, and a channel number of the self-attention layer is given Two-dimensional potential input with length H and width W , three convolutional layers with a kernel size of 1×1 are used to obtain , , :
[0026] ;
[0027] in, , yes The weight tensor of the convolution operation; the 2D potential of the output The two-dimensional electric potential with the input The same size, the number of channels is , whose value is:
[0028] ;
[0029] in, It's location Where Spatial neighborhood, The location The output value of Represents the transpose of the query vector, which is used to calculate the dot product attention weight; The location The key vector and value vector of ;
[0030] The self-attention mechanism can capture the distance dependency of the input two-dimensional potential through the attention weight and extract its features.
[0031] Furthermore, in step 2, the deep back projection network is specifically: the deep back projection network consists of a linear layer, two convolutional layers, five up-projection layers and four down-projection layers: the linear layer consists of 64 neurons, and the input one-dimensional MOSFET structure parameters The convolution kernel size of the first convolution layer is 1×1, which further learns the 8×8 two-dimensional feature information. The following five up-projection layers and four down-projection layers can learn the mapping relationship from the small-scale 8×8 two-dimensional potential to the large-scale 64×64 two-dimensional potential. The last 1×1 convolution layer converts all the input 64×64 large-scale feature information into a high-precision 64×64 two-dimensional potential.
[0032] The upper projection layer projects the small-scale two-dimensional potential characteristic information to the large-scale two-dimensional potential characteristic information. Its input is a two-dimensional potential distribution with a length and width of 8. , a two-dimensional potential distribution with a length and width of 64 is obtained through the following process :
[0033] ;
[0034] in, Represents large-scale two-dimensional electric potential features; Represents the small-scale two-dimensional electric potential characteristics; They represent the weight tensors of three different up and down convolution operations. Here, the convolution kernel is set to 12, the step size is 8, and the padding is 2;
[0035] The lower projection layer projects the large-scale two-dimensional potential characteristic information to the small-scale two-dimensional potential characteristic information, and its input is a two-dimensional potential distribution with a length and width of 64. , a two-dimensional potential distribution with a length and width of 8 is obtained through the following process :
[0036] ;
[0037] in, Represents the small-scale two-dimensional electric potential characteristics; Represents large-scale two-dimensional electric potential features; They represent the weight tensors of three different up and down convolution operations. Here, the convolution kernel is set to 12, the step size is 8, and the padding is 2;
[0038] The deep back projection network learns the mapping relationship from small-scale 8×8 two-dimensional potential to large-scale 64×64 two-dimensional potential through multiple projection layers, extracts multiple sets of effective feature information, and finally inputs all the effective feature information into the 1×1 convolutional layer to obtain high-precision two-dimensional potential.
[0039] Furthermore, in step 3, the predicting of the approximate value of the two-dimensional potential specifically includes the following steps:
[0040] Step 31, data processing: standardizing or normalizing the structural parameters of the semiconductor device to be simulated;
[0041] Step 32, predicting and obtaining the approximate value of the two-dimensional electric potential: inputting the structural parameters of the semiconductor device to be simulated after the data processing in step 31 into the two-dimensional electric potential prediction model established in step 2, and predicting the approximate value of the two-dimensional electric potential corresponding to the structural parameters.
[0042] Furthermore, in step 4, in the high-precision two-dimensional electric potential simulation, the iterative process of the numerical method includes two types: Newton iteration method and Gummel iteration method; the accuracy conditions for judging convergence include two types: X norm and RHS norm, which represent the iterative error of the independent variable and the iterative error of the nonlinear equation, respectively.
[0043] Furthermore, in step 4, in the high-precision two-dimensional potential simulation, in the numerical solution process of TCAD, a simulation model that conforms to physical reality is set, including a bandgap narrowing model, an avalanche effect model, a mobility model, and a composite model.
[0044] Furthermore, in step 4, the high-precision two-dimensional electric potential simulation method comprises the following steps:
[0045] Step 41, putting the two-dimensional potential approximation value into TCAD: putting the two-dimensional potential approximation value corresponding to the structural parameters of the semiconductor device to be simulated obtained in step 32 back into TCAD as an initial release;
[0046] Step 42, obtaining the carrier concentration corresponding to the initial solution of the two-dimensional potential: after setting the physical model in TCAD, the corresponding approximate value of the carrier concentration is obtained by the two-dimensional potential approximation value of step 41;
[0047] Step 43, numerical method iteration: according to the two-dimensional potential approximation value of step 41 and the carrier concentration approximation value of step 42, a convergence accuracy threshold is set, a physical simulation model is set, and iteration is performed through a numerical method;
[0048] Step 44, obtaining high-precision two-dimensional electric potential: When the iterative process of step 43 reaches the threshold, convergence is successful, and the high-precision two-dimensional electric potential of the semiconductor device to be simulated can be obtained.
[0049] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:
[0050] 1. The present invention utilizes a two-dimensional potential prediction model and a TCAD numerical simulation method to realize rapid simulation of the two-dimensional potential of semiconductor devices according to the structural parameters of the semiconductor devices. Compared with TCAD simulation tools, the present invention has better convergence, faster speed and can perform large-scale device simulation.
[0051] 2. Based on the simulation model in TCAD that conforms to physical reality and the reliable numerical method for solving the electric potential, the simulated two-dimensional electric potential is closer to reality and has higher accuracy. Compared with the simulation method based on deep learning, it is more practical and can meet the high-standard industrial application requirements.
[0052] 3. The present invention establishes two high-precision two-dimensional electric potential prediction models. Compared with deep convolutional neural networks, the predicted two-dimensional electric potential has higher accuracy, faster speed and better convergence in the numerical iteration process, and has good simulation performance.
[0053] 4. The deep back-projection network learns the mapping relationship from the small-scale 8×8 two-dimensional potential to the large-scale 64×64 two-dimensional potential through multiple projection layers, and extracts multiple sets of effective feature information. Finally, all the effective feature information is input into the 1×1 convolution layer to obtain high-precision two-dimensional potential. This process not only extracts a large amount of feature information through multiple projection layers, but also realizes multiple reuse of feature information, improves the utilization rate of feature information, and has better two-dimensional potential prediction performance.
[0054] 5. The self-attention mechanism can capture the distance dependence of the input two-dimensional potential through the attention weight, including the area below the MOSFET gate and the source-drain junction doping diffusion area where the potential changes too fast. Further feature extraction is performed on these areas to achieve higher-precision prediction of the two-dimensional potential. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 The figure is a flow chart of the acceleration method for high-precision two-dimensional potential simulation of semiconductor devices according to the present invention.
[0056] Figure 2 It is a schematic diagram of the structure of a MOSFET field effect tube in an embodiment of the present invention.
[0057] Figure 3 It is a schematic diagram of a simulation grid of a MOSFET field effect transistor in an embodiment of the present invention.
[0058] Figure 4 It is a schematic diagram of the structures of two two-dimensional potential prediction models in the two-dimensional potential simulation of the MOSFET field effect tube according to an embodiment of the present invention.
[0059] Figure 5 It is a schematic diagram of the predicted two-dimensional potential and error distribution of the MOSFET field effect tube according to an embodiment of the present invention.
[0060] Figure 6 It is a schematic diagram of a two-dimensional potential simulation of a MOSFET field effect tube according to an embodiment of the present invention. DETAILED DESCRIPTION
[0061] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The acceleration method of the semiconductor device high-precision two-dimensional potential simulation of the present invention is applied to the two-dimensional potential simulation of MOSFET field effect tubes for description.
[0062] like Figure 1 As shown, the present invention is applied to two-dimensional potential simulation of MOSFET field effect tube, comprising the following steps:
[0063] Step 1: Get the dataset;
[0064] Step 2: Establish a two-dimensional potential prediction model using a deep convolutional neural network or a deep back-projection network with a self-attention mechanism;
[0065] Step 3: Predict the approximate value of the 2D potential;
[0066] Step 4: High-precision two-dimensional electric potential simulation.
[0067] Example:
[0068] Taking the two-dimensional potential simulation of MOSFET field effect tube as an example for analysis, the two-dimensional potential simulation process of the present invention is as follows:
[0069] Step 1: Obtain a data set: Determine the value space of the simulation model and device structure parameters according to the type of semiconductor device, and obtain a data set based on TCAD simulation.
[0070] Step 11: Determine the structural parameters and their value space: Figure 2 As shown, for MOSFET field effect tube, the selected structural parameters include gate length , oxide layer thickness , source and drain doping depth , source-drain doping concentration , substrate doping concentration , gate voltage , drain voltage Table 1 gives the value range of the MOSFET field effect tube structure parameters in this example. All other device process structure parameters are fixed values.
[0071] Table 1 Value ranges of MOSFET field effect tube structural parameters
[0072]
[0073] Step 12: Determine the simulation model mesh type: Figure 3 As shown, the grid is set as a non-uniform grid in the Medici software.
[0074] Step 13, obtaining a two-dimensional potential distribution: according to the structural parameters in Table 1, the two-dimensional potential distribution of MOSFETs with different structural parameters is obtained by simulation using Medici software;
[0075] Step 14, obtaining a two-dimensional electric potential training label: interpolating the two-dimensional electric potential non-uniform grid obtained in step 13 into a 64*64 two-dimensional electric potential training grid by an interpolation sampling method;
[0076] Step 15, forming a data set: the structural parameters of each semiconductor device in step 11 are matched one by one with the two-dimensional potential training grid obtained in step 14, so as to form a data set from structural parameters to two-dimensional potential distribution.
[0077] Step 2: Establish a two-dimensional potential prediction model: Establish a two-dimensional potential prediction model based on the data set obtained in step 1; the two-dimensional potential prediction model can predict the two-dimensional potential based on the structural parameters of the MOSFET.
[0078] 21. Dataset classification: The data set obtained in step 1 is divided into a training set, a validation set and a test set according to a ratio of 6:2:2; the training set is used to train the two-dimensional electric potential prediction model; the validation set is used to adjust the weight value during the training process; and the test set is used to test the two-dimensional electric potential prediction model;
[0079] 22. Data processing: Standardize or normalize the data in the training set, validation set, and test set;
[0080] 23. Establish a two-dimensional potential prediction model: Establish a two-dimensional potential prediction model of a deep convolutional neural network or a deep back projection network that introduces a self-attention mechanism; the output parameter of the two-dimensional potential prediction model is the two-dimensional potential distribution matrix of the MOSFET to be simulated; the input parameter of the two-dimensional potential prediction model is the structural parameters of the MOSFET to be simulated;
[0081] Figure 4 The structural schematic diagrams of two two-dimensional potential prediction models for MOSFET are given.
[0082] For the potential distribution of MOSFET, since the potential in the area below the MOSFET gate and the source-drain junction doping diffusion area changes too quickly, the potential prediction accuracy in these areas is difficult to meet the conditions for being put into TCAD software as the initial solution, resulting in poor convergence performance and cannot be used as an efficient and usable direct prediction method.
[0083] In order to improve the above problems and enhance the prediction accuracy of two-dimensional electric potential, this method established two feasible and effective two-dimensional electric potential prediction models: a deep convolutional neural network with self-attention mechanism and a deep back-projection network.
[0084] The self-attention mechanism Figure 4 In (a), a deep convolutional neural network is added in the form of a self-attention layer, and the self-attention layer is given a two-dimensional potential input with a length of H and a width of W. , three convolutional layers with a kernel size of 1×1 are used to obtain , , :
[0085]
[0086] in, , yes The weight tensor for the convolution operation.
[0087] Output 2D electric potential The two-dimensional electric potential with the input The dimensions are the same, and their values are:
[0088]
[0089] in, It's location Where Spatial neighborhood, The location The output value of Represents the transpose of the query vector, which is used to calculate the dot product attention weight; The location The key vector and value vector of ;
[0090] The self-attention mechanism can capture the distance dependency of the input two-dimensional potential through the attention weight and extract its features.
[0091] The self-attention mechanism can capture the distance dependence of the input two-dimensional potential through the attention weight, including the area below the MOSFET gate and the source-drain junction doping diffusion area where the potential changes too quickly. Further feature extraction is performed on these areas to achieve higher-precision prediction of the two-dimensional potential.
[0092] Deep back-projection network Figure 4 As shown in (b), it consists of one linear layer, two convolutional layers, five up-projection layers and four down-projection layers: its linear layer consists of 64 neurons, which can convert the input one-dimensional MOSFET structure parameters Converted into 8×8 two-dimensional feature information; the convolution kernel size of the first convolution layer is 1×1, which can further learn the 8×8 two-dimensional feature information; the following five up-projection layers and four down-projection layers can learn the mapping relationship from small-scale 8×8 two-dimensional potential to large-scale 64×64 two-dimensional potential; the last 1×1 convolution layer can convert all the input 64×64 large-scale feature information into a high-precision 64×64 two-dimensional potential.
[0093] The upper projection layer can project the small-scale two-dimensional potential characteristic information to the large-scale two-dimensional potential characteristic information. Its input is a two-dimensional potential distribution with a length and width of 8. , a two-dimensional potential distribution with a length and width of 64 is obtained through the following process :
[0094]
[0095] in, Represents large-scale two-dimensional electric potential features; Represents the small-scale two-dimensional electric potential characteristics; They represent the weight tensors of three different up and down convolution operations. Here, the convolution kernel is set to 12, the step size is 8, and the padding is 2;
[0096] The lower projection layer can project the large-scale two-dimensional potential characteristic information to the small-scale two-dimensional potential characteristic information. It inputs a two-dimensional potential distribution with a length and width of 64. , a two-dimensional potential distribution with a length and width of 8 is obtained through the following process :
[0097]
[0098] in, Represents the small-scale two-dimensional electric potential characteristics; Represents large-scale two-dimensional electric potential features; They represent the weight tensors of three different up and down convolution operations. Here, the convolution kernel is set to 12, the step size is 8, and the padding is 2;
[0099] The deep back-projection network learns the mapping relationship from small-scale 8×8 two-dimensional potential to large-scale 64×64 two-dimensional potential through multiple projection layers, and extracts multiple sets of effective feature information. Finally, all effective feature information is input into the 1×1 convolution layer to obtain high-precision two-dimensional potential. This process not only extracts a large amount of feature information through multiple projection layers, but also realizes multiple reuse of feature information, improves the utilization rate of feature information, and has better two-dimensional potential prediction performance.
[0100] 24. Model training: using the data in the training set in step 22 to train the two-dimensional electric potential prediction model established in step 23, to obtain the parameters of the two-dimensional electric potential prediction model;
[0101] 25. Model verification and testing: Use the data in the verification set or test set in step 22 to verify or test the two-dimensional potential prediction model trained in step 24.
[0102] Step 3: Predict the approximate value of the two-dimensional potential: Input the structural parameters of the MOSFET to be simulated into the two-dimensional potential prediction model established in step 2, and predict the approximate value of the two-dimensional potential corresponding to the MOSFET to be simulated.
[0103] 31. Data processing: standardize or normalize the structural parameters of the semiconductor device to be simulated;
[0104] In this embodiment, for MOSFET, an example of structural parameters is given in Table 2. The structural parameters are input into the 2D potential prediction model to predict the 2D potential distribution. Figure 5 A schematic diagram of the predicted two-dimensional potential and its error distribution of the MOSFET field effect tube according to an embodiment of the present invention is given.
[0105] Table 2 Structural parameters of MOSFET to be simulated
[0106]
[0107] 32. Predict the approximate value of the two-dimensional potential: input the structural parameters of the MOSFET to be simulated after the data processing in step 31 into the two-dimensional potential prediction model established in step 2, and predict the approximate value of the two-dimensional potential corresponding to the structural parameters.
[0108] Step 4: High-precision two-dimensional electric potential simulation: The approximate two-dimensional electric potential predicted in step 3 is used as the initial release and input into the Medici software, and Newton iteration method is performed to obtain the high-precision two-dimensional electric potential corresponding to the semiconductor device to be simulated.
[0109] 41. Put the two-dimensional potential approximation into TCAD: Put the two-dimensional potential approximation corresponding to the structural parameters of the MOSFET to be simulated obtained in step 32 back to the Medici software as the initial release.
[0110] 42. Obtain the carrier concentration corresponding to the initial solution of the two-dimensional potential: After setting the model parameters in the Medici software, only choose to solve the Poisson equation. The corresponding carrier concentration approximation can be obtained through the two-dimensional potential approximation in step 41.
[0111] 43. Numerical method iteration: According to the two-dimensional potential approximation in step 41 and the carrier concentration approximation in step 42, set the mobility model to the constant mobility model CONMOB, the electric field dependent mobility model FLDMOB, and the surface scattering mobility model SRFMOB2, and set the X norm condition to , perform numerical iteration, and when the iterative error of the potential is less than V, convergence is successful.
[0112] 44. Obtain high-precision two-dimensional electric potential: When the iterative process in step 43 converges successfully, Figure 6 As shown, a high-precision two-dimensional potential of the MOSFET device to be simulated is obtained.
[0113] In this embodiment, for the MOSFET corresponding to the structural parameters in Table 2, Table 3 shows the simulation results in three different simulation processes.
[0114] Table 3 Comparison of the effects of different simulation methods
[0115]
Claims
1. A method for accelerating high-precision two-dimensional potential simulation of semiconductor devices, characterized in that: The steps include: Step 1: Obtaining data sets: According to the type of semiconductor device, determine the value space of the simulation model and device structure parameters, and obtain the data set based on TCAD simulation; Step 2: Establish a two-dimensional potential prediction model: Based on the data set obtained in step 1, establish a two-dimensional potential prediction model of a deep convolutional neural network or a deep back projection network that introduces a self-attention mechanism; the two-dimensional potential prediction model can predict the two-dimensional potential based on the structural parameters of the semiconductor device; Step 3: predicting the approximate value of the two-dimensional electric potential: inputting the structural parameters of the semiconductor device to be simulated into the two-dimensional electric potential prediction model established in step 2, and predicting the approximate value of the two-dimensional electric potential corresponding to the semiconductor device to be simulated; Step 4: High-precision two-dimensional electric potential simulation: The two-dimensional electric potential approximation value predicted in step 3 is used as the initial release into the TCAD software, and the numerical method is iterated to obtain the high-precision two-dimensional electric potential corresponding to the semiconductor device to be simulated; The simulation model includes a drift-diffusion model, a fluid dynamics model and a Monte Carlo model electrical performance model, and the grid types of the simulation model include a uniform grid and a non-uniform grid; The semiconductor device is a field effect transistor. For a MOSFET field effect transistor, its structural parameters include: gate length , oxide layer thickness , source and drain doping depth , source-drain doping concentration , substrate doping concentration , gate voltage , drain voltage ; In step 2, the deep convolutional neural network of the self-attention mechanism is specifically as follows: a deep convolutional neural network is added in the form of a self-attention layer, and a two-dimensional potential input with a length of H and a width of W is given to the self-attention layer. , three convolutional layers with a kernel size of 1×1 are used to obtain , , : ; in, , yes The weight tensor of the convolution operation; the 2D potential of the output The two-dimensional electric potential with the input The dimensions are the same, and their values are: ; in, It's location Where Spatial neighborhood, The location The potential value, value; The location Where Value and value; The self-attention mechanism can capture the distance dependency of the input two-dimensional potential through the attention weight and extract its features; In step 2, the deep back projection network is specifically: the deep back projection network consists of a linear layer, two convolutional layers, five up-projection layers and four down-projection layers: the linear layer consists of 64 neurons, and the input one-dimensional MOSFET structure parameters The convolution kernel size of the first convolution layer is 1×1, which further learns the 8×8 two-dimensional feature information. The following five up-projection layers and four down-projection layers can learn the mapping relationship from the small-scale 8×8 two-dimensional potential to the large-scale 64×64 two-dimensional potential. The last 1×1 convolution layer converts all the input 64×64 large-scale feature information into a high-precision 64×64 two-dimensional potential. The upper projection layer projects the small-scale two-dimensional potential characteristic information to the large-scale two-dimensional potential characteristic information. Its input is a two-dimensional potential distribution with a length and width of 8. , a two-dimensional potential distribution with a length and width of 64 is obtained through the following process : ; in, Represents large-scale two-dimensional electric potential features; Represents the small-scale two-dimensional electric potential characteristics; They represent the weight tensors of three different up and down convolution operations. Here, the convolution kernel is set to 12, the step size is 8, and the padding is 2; The lower projection layer projects the large-scale two-dimensional potential characteristic information to the small-scale two-dimensional potential characteristic information, and its input is a two-dimensional potential distribution with a length and width of 64. , a two-dimensional potential distribution with a length and width of 8 is obtained through the following process : ; in, Represents the small-scale two-dimensional electric potential characteristics; Represents large-scale two-dimensional electric potential features; They represent the weight tensors of three different up and down convolution operations. Here, the convolution kernel is set to 12, the step size is 8, and the padding is 2; The deep back projection network learns the mapping relationship from small-scale 8×8 two-dimensional potential to large-scale 64×64 two-dimensional potential through multiple projection layers, extracts multiple sets of effective feature information, and finally inputs all the effective feature information into the 1×1 convolutional layer to obtain high-precision two-dimensional potential.
2. The acceleration method for high-precision two-dimensional potential simulation of semiconductor devices according to claim 1, characterized in that: In step 1, a data set is obtained based on TCAD simulation, including the following steps: Step 11, determining structural parameters and their value space: according to the type of semiconductor device to be simulated, selecting structural parameters related to its two-dimensional potential, and determining the value space of the structural parameters; Step 12: Determine the mesh type of the simulation model: set the mesh in TCAD and choose to set a uniform mesh or a non-uniform mesh; Step 13, obtaining a two-dimensional potential distribution: obtaining a two-dimensional potential distribution of semiconductor devices with different structural parameters through TCAD simulation; Step 14, obtaining a two-dimensional electric potential training label: interpolating the two-dimensional electric potential distribution grid obtained in step 13 into a 64*64 two-dimensional electric potential training grid by an interpolation sampling method; Step 15, forming a data set: the structural parameters of each semiconductor device in step 11 are matched one by one with the two-dimensional potential training grid obtained in step 14, so as to form a data set from structural parameters to two-dimensional potential distribution.
3. The acceleration method for high-precision two-dimensional potential simulation of semiconductor devices according to claim 1, characterized in that: In step 2, establishing a two-dimensional potential prediction model specifically includes the following steps: Step 21, data set classification: the data set obtained in step 1 is divided into a training set, a validation set and a test set according to a set ratio; wherein the training set is used to train the two-dimensional electric potential prediction model; the validation set is used to adjust the weight value in the training process; and the test set is used to test the two-dimensional electric potential prediction model; Step 22, data processing: standardize or normalize the data in the training set, validation set, and test set; Step 23, establishing a two-dimensional potential prediction model: designing a deep convolutional neural network or a deep back projection network that introduces a self-attention mechanism to establish a two-dimensional potential prediction model; the output parameter of the two-dimensional potential prediction model is the two-dimensional potential distribution matrix of the semiconductor device to be simulated; the input parameter of the two-dimensional potential prediction model is the structural parameter of the semiconductor device to be simulated; Step 24, model training: using the data in the training set in step 22 to train the two-dimensional electric potential prediction model established in step 23, to obtain parameters of the two-dimensional electric potential prediction model; Step 25, model verification and testing: using the data in the verification set or test set in step 22, the two-dimensional potential prediction model trained in step 24 is verified or tested.
4. The acceleration method for high-precision two-dimensional potential simulation of semiconductor devices according to claim 1, characterized in that: In step 3, the predicting of the approximate value of the two-dimensional potential specifically comprises the following steps: Step 31, data processing: standardizing or normalizing the structural parameters of the semiconductor device to be simulated; Step 32, predicting and obtaining the approximate value of the two-dimensional electric potential: inputting the structural parameters of the semiconductor device to be simulated after the data processing in step 31 into the two-dimensional electric potential prediction model established in step 2, and predicting the approximate value of the two-dimensional electric potential corresponding to the structural parameters.
5. The acceleration method for high-precision two-dimensional potential simulation of semiconductor devices according to claim 1, characterized in that: In step 4, in the high-precision two-dimensional electric potential simulation, the iterative process of the numerical method includes two methods: Newton iteration method and Gummel iteration method; the accuracy conditions for judging convergence include two methods: X norm and RHS norm, which represent the iterative error of the independent variable and the iterative error of the nonlinear equation, respectively.
6. The acceleration method for high-precision two-dimensional potential simulation of semiconductor devices according to claim 1, characterized in that: In step 4, in the high-precision two-dimensional potential simulation, in the numerical solution process of TCAD, a simulation model that conforms to physical reality is set, including a bandgap narrowing model, an avalanche effect model, a mobility model, and a composite model.
7. The acceleration method for high-precision two-dimensional potential simulation of semiconductor devices according to claim 4, characterized in that: In step 4, the high-precision two-dimensional electric potential simulation method comprises the following steps: Step 41, putting the two-dimensional potential approximation value into TCAD: putting the two-dimensional potential approximation value corresponding to the structural parameters of the semiconductor device to be simulated obtained in step 32 back into TCAD as an initial release; Step 42, obtaining the carrier concentration corresponding to the initial solution of the two-dimensional potential: after setting the physical model in TCAD, the corresponding approximate value of the carrier concentration is obtained by the two-dimensional potential approximation value of step 41; Step 43, numerical method iteration: according to the two-dimensional potential approximation value of step 41 and the carrier concentration approximation value of step 42, a convergence accuracy threshold is set, a physical simulation model is set, and iteration is performed through a numerical method; Step 44, obtaining high-precision two-dimensional electric potential: When the iterative process of step 43 reaches the threshold, convergence is successful, and the high-precision two-dimensional electric potential of the semiconductor device to be simulated can be obtained.
