Global modeling method based on hybrid expert neural network and related device

Abstract

The application discloses a global modeling method based on a hybrid expert neural network and related equipment, relates to the technical field of semiconductor device modeling, and comprises the following steps: acquiring sample data of a semiconductor device to be modeled; based on the sample data, performing stage-by-stage training on a preset hybrid expert neural network model; inputting target physical structure parameters and target working state parameters corresponding to the semiconductor device to be modeled into the trained hybrid expert neural network model, outputting a weight distribution vector by a gating network, outputting an intermediate prediction value by an expert network, and performing weighted fusion on the weight distribution vector and the intermediate prediction value to obtain a global modeling result of the semiconductor device to be modeled. The application realizes the global modeling effect of taking into account both high fitting accuracy and regional boundary continuity.

Description

Technical Field

[0001] This application relates to the field of semiconductor device modeling technology, and in particular to a global modeling method and related equipment based on hybrid expert neural networks. Background Technology

[0002] In the design and manufacturing of semiconductor devices, establishing a model that can accurately describe the electrical characteristics of the device is a key step in circuit simulation and design verification. As device feature sizes continue to shrink, the physical effects of advanced structural devices such as all-around gate field-effect transistors are becoming increasingly complex, and their electrical characteristics exhibit significant nonlinearity and non-monotonicity with changes in geometric parameters.

[0003] In related technologies, the following two schemes are mainly used for global modeling of scalable devices.

[0004] Firstly, a single global neural network model is used for modeling. This method inputs device data under different geometric dimensions and bias conditions into a single deep neural network for training, attempting to fit the device characteristics across the entire geometric parameter space with a single set of network parameters. However, due to non-ideal factors such as short-channel effects, the physical characteristics of devices vary greatly across different size ranges. When faced with a wide range of geometric parameters, a single neural network model struggles to simultaneously ensure fitting accuracy across all regions, easily leading to underfitting or overfitting in some areas. Blindly increasing the network size to improve accuracy, on the other hand, will significantly reduce the model's inference speed.

[0005] Secondly, a binning method is used for modeling. This method pre-divides the geometric space of the device into multiple independent regions and extracts a set of model parameters for each region. In the prediction phase, it first determines which region the geometric parameters of the target device fall into, and then calls the corresponding model parameters for that region for calculation. Although the binning method can effectively improve the modeling accuracy within each region, at the boundaries of adjacent regions, the discontinuous switching of model parameters will cause significant jumps in the predicted electrical characteristic values, seriously affecting the reliability and convergence of the simulation.

[0006] Therefore, in related technologies, global modeling of scalable devices faces the technical problem of not being able to simultaneously ensure modeling accuracy and the continuity of region boundaries. Summary of the Invention

[0007] The main purpose of this application is to provide a global modeling method and related equipment based on a hybrid expert neural network, which aims to solve the technical problem in the related technology that global modeling of scalable devices cannot simultaneously take into account modeling accuracy and regional boundary continuity.

[0008] To achieve the above objectives, this application proposes a global modeling method based on a hybrid expert neural network, the method comprising: Obtain sample data of the semiconductor device to be modeled; Based on the sample data, a preset hybrid expert neural network model is trained in stages. The hybrid expert neural network model includes a gating network and multiple expert networks. In the first training stage, only the expert networks are trained. In the second training stage, the expert networks are fixed and the gating network is trained. The target physical structure parameters and target operating state parameters corresponding to the semiconductor device to be modeled are input into the trained hybrid expert neural network model. The gating network outputs a weight allocation vector, the expert network outputs an intermediate prediction value, and the weight allocation vector and the intermediate prediction value are weighted and fused to obtain the global modeling result of the semiconductor device to be modeled.

[0009] In one embodiment, the step of performing phased training on a preset hybrid expert neural network model based on the sample data includes: The sample data is normalized. Based on the distribution of physical structure parameters in the sample data, the normalized sample data is divided into multiple data subsets; The expert networks corresponding to each of the aforementioned data subsets are pre-trained in parallel using the data in the hybrid expert neural network model. The gating network is configured to receive the physical structure parameters, and the expert network is configured to receive the physical structure parameters and the working state parameters in the sample data. The network parameters of each of the expert networks after the parallel pre-training are fixed; The hybrid expert neural network model is trained as a whole using the complete sample data. During the overall training process, only the network parameters of the gating network are updated, so that the gating network learns to assign smooth transition weight coefficients to each expert network according to the input physical structure parameters, thereby achieving continuous modeling at the boundaries of different physical structure parameter regions.

[0010] In one embodiment, the step of dividing the normalized sample data into multiple data subsets based on the distribution of physical structure parameters in the sample data includes: Based on the critical size threshold for the occurrence of short-channel effect in the semiconductor device, the parameter space composed of the physical structure parameters is divided into multiple non-overlapping regions with continuous boundaries. The sample data belonging to the same region are grouped into the same data subset, such that each data subset uniquely corresponds to a specific sub-region in the parameter space; The partitioning operation enables each expert network to focus on learning the changes in the device's electrical characteristics within its corresponding sub-region during the parallel pre-training phase. After the division operation, adjacent regions are seamlessly connected in parameter space and completely cover the entire preset range of variation of the physical structure parameters.

[0011] In one embodiment, the step of performing parallel pre-training of the respective expert networks in the hybrid expert neural network model using each of the data subsets further includes: The gated network is constructed by sequentially connecting an input layer, at least two hidden layers, and an output layer. The hidden layers use the hyperbolic tangent function as the activation function. The number of neurons in the output layer is equal to the number of neurons in the expert network, and the output layer uses the Softmax function as the activation function. Construct the plurality of expert networks, each of which consists of an input layer, at least two hidden layers and an output layer connected in sequence. The hidden layers of the expert network use the SigmaRhodes function as the activation function, and the output layer of the expert network uses the hyperbolic tangent function as the activation function. The number of neurons in the hidden layer of the expert network is less than the number of neurons in the hidden layer of a single global neural network model constructed for the same modeling task. The number of expert networks, the number of hidden layers and the number of neurons in each expert network are set according to the variation range of the physical structure parameters of the semiconductor device and the preset modeling accuracy requirements.

[0012] In one embodiment, the step of training the hybrid expert neural network model using the complete sample data, wherein only the network parameters of the gating network are updated during the overall training process, includes: The physical structure parameters in the complete sample data are input into the gating network to obtain the weight allocation vectors corresponding to each expert network output by the gating network. The physical structure parameters and working state parameters from the complete sample data are input into each of the expert networks whose parameters have been fixed, to obtain the intermediate prediction values ​​output by each of the expert networks. The intermediate predicted values ​​are weighted and fused according to the weight allocation vector to obtain the global electrical characteristic predicted value. Using the actual electrical characteristic values ​​in the complete sample data as the supervision signal, the error between the global electrical characteristic prediction value and the actual electrical characteristic value is calculated by minimizing a preset composite loss function, and the network parameters of the gated network are updated only based on the backpropagation of the error. The composite loss function is composed of a weighted sum of a first error term, a second error term, a third error term, and a fourth error term. The first error term is the mean square error between the predicted and actual values ​​of the electrical characteristic data. The second error term is the mean square error between the predicted and actual values ​​of the logarithm of the electrical characteristic data (base 10). The third error term is the mean square error between the predicted and actual values ​​of the partial derivative of the electrical characteristic data with respect to the first bias voltage in the operating state parameters. The fourth error term is the mean square error between the predicted and actual values ​​of the partial derivative of the electrical characteristic data with respect to the second bias voltage in the operating state parameters.

[0013] In one embodiment, the step of inputting the target physical structure parameters and target operating state parameters corresponding to the semiconductor device to be modeled into the trained hybrid expert neural network model, and having the gating network output a weight allocation vector, includes: In the gated network, an initial weight allocation vector is generated based on the target physical structure parameters; A predetermined number of weight coefficients with the largest values ​​are retained in the initial weight allocation vector, and the remaining weight coefficients in the initial weight allocation vector are set to zero to form the weight allocation vector; When performing the weighted fusion calculation on the weight allocation vector and the intermediate predicted value, only the expert network corresponding to the retained weight coefficients participates in the calculation, while the expert network corresponding to the weight coefficients set to zero does not participate in the calculation during the current prediction process.

[0014] Furthermore, to achieve the above objectives, this application also proposes a global modeling device based on a hybrid expert neural network, the global modeling device based on the hybrid expert neural network comprising: The acquisition module is used to acquire sample data of the semiconductor device to be modeled; The training module is used to train a preset hybrid expert neural network model in stages based on the sample data. The hybrid expert neural network model includes a gating network and multiple expert networks. In the first training stage, only the expert networks are trained. In the second training stage, the expert networks are fixed and the gating networks are trained. The modeling module is used to input the target physical structure parameters and target operating state parameters corresponding to the semiconductor device to be modeled into the trained hybrid expert neural network model. The gating network outputs a weight allocation vector, the expert network outputs intermediate prediction values, and the weight allocation vector and the intermediate prediction values ​​are weighted and fused to obtain the global modeling result of the semiconductor device to be modeled.

[0015] Furthermore, to achieve the above objectives, this application also proposes a global modeling device based on a hybrid expert neural network, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the global modeling method based on a hybrid expert neural network as described above.

[0016] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the global modeling method based on a hybrid expert neural network as described above.

[0017] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the global modeling method based on a hybrid expert neural network as described above.

[0018] One or more technical solutions proposed in this application have at least the following technical effects: Compared to related technologies that employ a single global neural network model, resulting in insufficient accuracy in local areas, or that use binning methods, leading to abrupt changes in electrical characteristics at regional boundaries, this application constructs a hybrid expert neural network model comprising a gating network and multiple expert networks, and optimizes it using a phased training strategy. Understandably, this application involves acquiring sample data of the scalable device to be modeled; based on the sample data, a pre-defined hybrid expert neural network model is trained in phases, wherein the hybrid expert neural network model includes a gating network and multiple expert networks. In the first training phase, only the expert networks are trained; in the second training phase, the expert networks are fixed while the gating network is trained; the target physical structure parameters and target operating state parameters corresponding to the scalable device to be modeled are input into the trained hybrid expert neural network model; the gating network outputs a weight allocation vector, and the expert networks output intermediate predicted values; the weight allocation vector and the intermediate predicted values ​​are then weighted and fused to obtain the global modeling result of the semiconductor device to be modeled. When global modeling of scalable devices with large geometric parameter spans is required, the first training phase trains only the expert networks, enabling each expert network to focus on learning the electrical characteristics of the device within a specific geometric parameter subspace, achieving high-precision fitting of the device behavior within that subspace. The second training phase fixes the expert networks and trains the gating network, allowing the gating network to learn to assign smooth transition weight coefficients to each expert network based on the input physical structure parameters, thus achieving continuous transition of prediction results at the boundaries of different geometric parameter regions. Therefore, based on the architecture design of the hybrid expert neural network model and the phased training strategy, the global modeling task can be decomposed into multiple expert networks focused on local regions, dynamically and softly combined by the gating network. This allows for the determination of device electrical characteristic descriptions that maintain high accuracy and smooth boundaries throughout the complete geometric parameter space, ultimately achieving global modeling of scalable devices that balances modeling accuracy and regional boundary continuity. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0021] Figure 1This is a flowchart illustrating an embodiment of the global modeling method based on hybrid expert neural networks in this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the global modeling method based on hybrid expert neural networks provided in this application; Figure 3 This is a flowchart illustrating Embodiment 3 of the global modeling method based on hybrid expert neural networks in this application; Figure 4 This is a schematic diagram of the module structure of the global modeling device based on a hybrid expert neural network according to an embodiment of this application; Figure 5 This is a schematic diagram of the device structure of the hardware operating environment involved in the global modeling method based on hybrid expert neural networks in the embodiments of this application.

[0022] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0024] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0025] The main solution in this application's embodiments is: Obtain sample data of the semiconductor device to be modeled; Based on the sample data, a preset hybrid expert neural network model is trained in stages. The hybrid expert neural network model includes a gating network and multiple expert networks. In the first training stage, only the expert networks are trained. In the second training stage, the expert networks are fixed and the gating network is trained. The target physical structure parameters and target operating state parameters corresponding to the semiconductor device to be modeled are input into the trained hybrid expert neural network model. The gating network outputs a weight allocation vector, the expert network outputs an intermediate prediction value, and the weight allocation vector and the intermediate prediction value are weighted and fused to obtain the global modeling result of the semiconductor device to be modeled.

[0026] In this embodiment, the application uses a global modeling device based on a hybrid expert neural network as the execution subject. For ease of description, it will be referred to as "device" in detail below.

[0027] The relevant technologies use a single global neural network model, which leads to insufficient accuracy in local areas, or use a binning method, which causes abrupt changes in electrical characteristics at the boundaries of the regions.

[0028] This application provides a solution for global modeling of scalable devices with large geometric parameter spans. In the first training phase, only the expert networks are trained, allowing each network to focus on learning the electrical characteristics of the device within a specific geometric parameter subspace, achieving high-precision fitting of the device behavior within that subspace. In the second training phase, the expert networks are fixed while the gating network is trained, enabling the gating network to learn to assign smooth transition weight coefficients to each expert network based on the input physical structure parameters, thus achieving continuous transition of prediction results at the boundaries of different geometric parameter regions. Therefore, based on the architecture design of the hybrid expert neural network model and the phased training strategy, the global modeling task can be decomposed into multiple expert networks focused on local regions, dynamically and softly combined by the gating network. This allows for the determination of a device electrical characteristic description that maintains high accuracy and smooth boundaries throughout the complete geometric parameter space, ultimately achieving global modeling of scalable devices that balances modeling accuracy and regional boundary continuity.

[0029] Based on this, embodiments of this application provide a global modeling method based on a hybrid expert neural network, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the global modeling method based on hybrid expert neural networks in this application.

[0030] In this embodiment, the global modeling method based on hybrid expert neural networks includes steps S10 to S30: Step S10: Obtain sample data of the semiconductor device to be modeled; It should be noted that the scalable device to be modeled refers to a semiconductor device whose electrical characteristics need to be described and predicted through modeling methods, including but not limited to all-around gate field-effect transistors, fin field-effect transistors, or other transistor devices with scalable geometric parameters. The sample data refers to the dataset used to train the hybrid expert neural network model, and the sample data includes at least physical structure parameters, operating state parameters, and electrical characteristic data corresponding to the physical structure parameters and operating state parameters. The physical structure parameters are parameters characterizing the geometric dimensions and physical composition of the scalable device, including but not limited to one or more of gate length, channel width, fin height, and oxide thickness. The operating state parameters are parameters characterizing the electrical bias conditions that the scalable device withstands under actual operating conditions, including but not limited to one or more of drain-source voltage, gate-source voltage, and body bias voltage. The electrical characteristic data are physical quantities characterizing the electrical behavior of the scalable device obtained through process computer-aided design simulation or actual measurement under conditions corresponding to the physical structure parameters and operating state parameters, including but not limited to one or more of drain current, channel conductance, and transconductance.

[0031] Understandably, this step, by acquiring electrical characteristic data covering different combinations of physical structure parameters and operating state parameters, provides the necessary data foundation for the subsequent training of the hybrid expert neural network model. The diversity and coverage of the sample data directly affect the model's generalization ability within the complete geometric parameter space. Therefore, the sample data should, as far as possible, cover the range of variations in physical structure parameters and operating state parameters involved in the practical application of the scalable device to be modeled.

[0032] Step S20: Based on the sample data, the preset hybrid expert neural network model is trained in stages. The hybrid expert neural network model includes a gating network and multiple expert networks. In the first training stage, only the expert networks are trained. In the second training stage, the expert networks are fixed and the gating networks are trained. It should be noted that the hybrid expert neural network model refers to a neural network architecture that includes a gating network and multiple expert networks, and dynamically combines the outputs of each expert network through the gating network. The gating network is a sub-network in the hybrid expert neural network model used to dynamically generate weight coefficients corresponding to each expert network based on input information. The gating network takes the physical structure parameters as input and outputs a weight allocation vector equal to the number of expert networks. Each expert network in the hybrid expert neural network model refers to multiple parallel sub-networks used for feature extraction and prediction of input information. Each expert network takes the physical structure parameters and the operating state parameters as input and outputs an intermediate prediction value. The phased training refers to dividing the training process of the hybrid expert neural network model into at least two training phases executed sequentially, with different network parameter objects updated in different training phases. The first training phase refers to the phase in the phased training where only the expert networks are trained. In the first training phase, the network parameters of the gating network remain in their initial state or are in an inactive state, and the network parameters of each expert network are optimized only through training data. The second training phase refers to the training phase executed after the first training phase in the phased training. In the second training phase, the network parameters of each expert network that has been trained in the first training phase are fixed, that is, set to a non-updateable state, and the network parameters of the gated network are updated and optimized using the training data.

[0033] Understandably, this step divides the training process into a first training phase and a second training phase. In the first training phase, each expert network learns and acquires the electrical characteristics within different physical structure parameter regions. In the second training phase, the gating network learns how to optimally combine weights for each expert network based on the input physical structure parameters. Since the training processes of each expert network in the first training phase do not interfere with each other, parallel training can be used to improve training efficiency. Furthermore, since the parameters of the expert networks are fixed in the second training phase, the gating network only needs to learn the weight allocation strategy, resulting in a more stable training process and faster convergence. Through this phased training, the hybrid expert neural network model can ultimately assign smoothly transitioning weight coefficients to each expert network based on the input physical structure parameters, thereby achieving continuous changes in prediction results at the boundaries of different physical structure parameter regions.

[0034] Step S30: Input the target physical structure parameters and target operating state parameters corresponding to the semiconductor device to be modeled into the trained hybrid expert neural network model. The gating network outputs a weight allocation vector, the expert network outputs an intermediate prediction value, and the weight allocation vector and the intermediate prediction value are weighted and fused to obtain the global modeling result of the semiconductor device to be modeled.

[0035] It should be noted that the target physical structure parameters refer to the specific physical structure parameter values ​​possessed by the scalable device to be modeled in actual applications or simulations. The target operating state parameters refer to the specific operating state parameter values ​​experienced by the scalable device to be modeled in actual applications or simulations. The weight allocation vector is a multi-dimensional vector generated by the gating network based on the input target physical structure parameters. The dimension of the weight allocation vector is equal to the number of expert networks, and each element in the weight allocation vector corresponds to a weight coefficient of an expert network in the current prediction. The intermediate prediction value refers to the electrical characteristic prediction result independently calculated and output by each expert network based on the input target physical structure parameters and the target operating state parameters. The weighted fusion refers to the weighted summation of the intermediate prediction values ​​output by each expert network with the corresponding weight coefficients in the weight allocation vector to obtain the final prediction output value. The global modeling result refers to the predicted electrical characteristics of the scalable device to be modeled under given physical structure parameters and operating state parameters, ultimately output by the hybrid expert neural network model.

[0036] Understandably, in this step, during the model inference stage, a gating network dynamically generates weight assignments based on the physical structure parameters of the device. Each expert network then calculates its own predicted value for the current input in parallel, and finally, a weighted fusion is performed to obtain the global prediction result. Since the weight coefficients output by the gating network smoothly change with the continuous variation of the physical structure parameters, when the target physical structure parameters are in the transition zone between regions favored by different expert networks, the gating network assigns similar weight coefficients to expert networks in adjacent regions. This allows the weighted fusion prediction result to smoothly transition between different regions, avoiding the prediction value jump problem that occurs at region boundaries in traditional binning methods. Furthermore, because each expert network focuses on the electrical characteristics within a specific physical structure parameter region, the weighted fusion prediction result maintains high fitting accuracy across the entire physical structure parameter space.

[0037] This embodiment provides a global modeling method based on a hybrid expert neural network. When global modeling of scalable devices with large geometric parameter spans is required, the first training phase trains only the expert networks, enabling each expert network to focus on learning the electrical characteristics of the device within a specific geometric parameter subspace, achieving high-precision fitting of the device behavior within that subspace. The second training phase fixes the expert networks and trains the gating network, allowing the gating network to learn to assign smooth transition weight coefficients to each expert network based on the input physical structure parameters, thereby achieving continuous transition of prediction results at the boundaries of different geometric parameter regions. Therefore, based on the architecture design of the hybrid expert neural network model and the phased training strategy, the global modeling task can be decomposed into multiple expert networks focused on local regions, dynamically and softly combined by the gating network. This allows for the determination of a device electrical characteristic description that maintains high accuracy and smooth boundaries throughout the complete geometric parameter space, ultimately achieving global modeling of scalable devices that balances modeling accuracy and regional boundary continuity.

[0038] In one feasible implementation, the step of performing phased training on the preset hybrid expert neural network model based on the sample data includes: The sample data is normalized. Based on the distribution of physical structure parameters in the sample data, the normalized sample data is divided into multiple data subsets; The expert networks corresponding to each of the aforementioned data subsets are pre-trained in parallel using the data in the hybrid expert neural network model. The gating network is configured to receive the physical structure parameters, and the expert network is configured to receive the physical structure parameters and the working state parameters in the sample data. The network parameters of each of the expert networks after the parallel pre-training are fixed; The hybrid expert neural network model is trained as a whole using the complete sample data. During the overall training process, only the network parameters of the gating network are updated, so that the gating network learns to assign smooth transition weight coefficients to each expert network according to the input physical structure parameters, thereby achieving continuous modeling at the boundaries of different physical structure parameter regions.

[0039] It should be noted that the normalization process refers to a data preprocessing operation that maps the original values ​​of the physical structure parameters, operating state parameters, and electrical characteristic data in the sample data to a unified numerical range through a preset mathematical transformation. Specific normalization methods include, but are not limited to, any one of the equidistant scaling method and the standard deviation standardization method. The distribution of the physical structure parameters refers to the statistical distribution characteristics of the physical structure parameters within their preset range of variation, including but not limited to the value span, value density, and degree of difference in physical characteristics between different value ranges. The data subset refers to multiple non-overlapping data sets formed after dividing the normalized sample data according to the distribution characteristics of the physical structure parameters. Each data subset corresponds to a specific sub-region in the parameter space constituted by the physical structure parameters. The parallel pre-training refers to the operation of simultaneously and independently training the corresponding expert network on multiple computing units using their respective allocated data subsets. During the parallel pre-training process, each expert network only uses the sample data from its corresponding data subset for iterative parameter updates. The network parameter fixing refers to setting the weight parameters and bias parameters of each expert network after parallel pre-training to a non-updateable state. After the network parameters are fixed, the expert network only performs forward propagation calculations and no longer performs backpropagation updates during subsequent training. The overall training refers to training the hybrid expert neural network model as a whole using complete sample data under the condition that the network parameters of the expert networks are fixed. "Complete" refers to all sample data covering the entire preset range of variation of the physical structure parameters without undergoing the partitioning operation. The smooth transition weight coefficients refer to the element values ​​in the weight allocation vector generated by the gating network based on the input physical structure parameters, which continuously change with the continuous change of the physical structure parameters. The smooth transition characteristic of the weight coefficients ensures that when the input physical structure parameters gradually change from one region to an adjacent region, the weight coefficients of the two expert networks corresponding to the adjacent regions show a gradual trend of one increasing and the other decreasing, rather than a sudden switch. The continuous modeling refers to the modeling effect where the predicted electrical characteristic values ​​change continuously with the changes of the physical structure parameters throughout the entire preset range of variation of the physical structure parameters, without any abrupt changes or discontinuities.

[0040] It is understandable that by normalizing the sample data, the adverse effects on the network training convergence speed caused by differences in dimensions or large differences in numerical ranges among the physical structure parameters, operating state parameters, and electrical characteristic data are eliminated, ensuring that each dimension of data contributes equally in the training process. By dividing the sample data into multiple data subsets according to the distribution of the physical structure parameters, the sample data in each subset exhibits relatively consistent patterns in physical characteristics, providing a data foundation for the subsequent specialized training of each expert network. By using each data subset to perform parallel pre-training on its corresponding expert network, each expert network can focus on learning the local variation patterns of the device's electrical characteristics within its corresponding sub-region, avoiding underfitting or overfitting problems caused by large differences in data characteristics across the entire network. Simultaneously, the parallel training method significantly shortens the overall training time. By fixing the network parameters of each expert network after parallel pre-training, the local patterns learned by each expert network are fully preserved in the subsequent overall training stage, avoiding interference or overwriting of the knowledge already acquired by the expert network during global training. By training the hybrid expert neural network model using the complete sample data and updating only the network parameters of the gating network, the gating network can learn, based on complete parameter space information, how to optimally combine the weights of each expert network for different input physical structure parameters. Furthermore, since only the gating network needs to be trained, the convergence speed of the overall training phase is significantly improved. Through this overall training, the gating network ultimately learns the ability to assign smoothly transitioning weight coefficients to each expert network. Therefore, during the model inference phase, when the physical structure parameters of the target device are at the boundaries of different regions, the prediction result can be contributed by the expert networks of adjacent regions with a smooth transition. Ultimately, this achieves a continuous modeling effect that maintains both high accuracy and boundary continuity across the entire range of physical structure parameter variations.

[0041] In one feasible implementation, the step of dividing the normalized sample data into multiple data subsets based on the distribution of physical structure parameters in the sample data includes: Based on the critical size threshold for the occurrence of short-channel effect in the semiconductor device, the parameter space composed of the physical structure parameters is divided into multiple non-overlapping regions with continuous boundaries. The sample data belonging to the same region are grouped into the same data subset, such that each data subset uniquely corresponds to a specific sub-region in the parameter space; The partitioning operation enables each expert network to focus on learning the changes in the device's electrical characteristics within its corresponding sub-region during the parallel pre-training phase. After the division operation, adjacent regions are seamlessly connected in parameter space and completely cover the entire preset range of variation of the physical structure parameters.

[0042] It should be noted that the short-channel effect refers to a series of physical phenomena that deviate from the ideal characteristics of long-channel devices, such as a decrease in threshold voltage, an increase in subthreshold swing, and a decrease in drain-induced barrier, when the channel length of a field-effect transistor is shortened to be comparable to the width of the source-drain depletion layer. The critical size threshold refers to the boundary value of the physical structure parameter used to distinguish between long-channel and short-channel characteristics. For a fully all-around gate field-effect transistor, the critical size threshold is usually determined based on the range of gate length values ​​corresponding to the occurrence of a significant short-channel effect, including but not limited to a specific value within the range of 10 to 30 nanometers. The parameter space refers to a multi-dimensional space spanned by one or more dimensions of the physical structure parameters, where each point corresponds to a specific combination of the physical structure parameter values. The non-overlapping and continuous boundary regions refer to the subspaces obtained after dividing the parameter space. Any two adjacent subspaces are only adjacent at the boundary and do not contain the same parameter points internally, and there are no gaps or discontinuities at the boundaries of adjacent subspaces. The specific sub-region refers to a defined sub-space within the parameter space, where the combinations of physical structure parameter values ​​exhibit similar device physical characteristics, such as belonging to the long-channel region, short-channel region, or deep-short-channel region. Seamless connection means that adjacent regions are closely adjacent at their boundaries, with no uncovered blank intervals in the parameter space transition between them. Complete coverage means that the union of all regions after partitioning is exactly equal to the entire preset range of variation of the physical structure parameters, with no omitted parameter value intervals.

[0043] It is understandable that by dividing the parameter space according to the critical size threshold for the occurrence of the short-channel effect, the divided regions have inherent consistency in physical characteristics. The electrical characteristic variation patterns of devices within the same region are similar, while the differences in physical mechanisms between different regions are naturally separated. By dividing the parameter space into multiple non-overlapping and continuously bounded regions, a clear spatial division basis is provided for the subsequent division of data subsets, ensuring that each data subset corresponds to a parameter sub-interval with clear physical meaning. By grouping sample data belonging to the same region into the same data subset, the samples in each data subset are within adjacent and continuous value ranges in terms of physical structural parameters, which is beneficial for the corresponding expert network to learn the continuous function mapping relationship within this value range. By enabling each expert network to focus on learning the device electrical characteristic variation patterns within its corresponding sub-region during the parallel pre-training stage, each expert network only needs to master the input-output relationship within a local parameter range, thereby reducing the learning difficulty of a single expert network and improving the fitting accuracy within that sub-region. By ensuring seamless connection between adjacent regions in the parameter space and fully covering the entire preset range of variation of the physical structure parameters, it is ensured that in the subsequent inference stage, for any target physical structure parameter falling within the preset range of variation, there is a corresponding expert network or a combination of adjacent expert networks that can provide effective prediction output, and there are no blind spots in the coverage of the parameter space.

[0044] In one feasible implementation, the step of performing parallel pre-training of the respective expert networks in the hybrid expert neural network model using each of the data subsets further includes: The gated network is constructed by sequentially connecting an input layer, at least two hidden layers, and an output layer. The hidden layers use the hyperbolic tangent function as the activation function. The number of neurons in the output layer is equal to the number of neurons in the expert network, and the output layer uses the Softmax function as the activation function. Construct the plurality of expert networks, each of which consists of an input layer, at least two hidden layers and an output layer connected in sequence. The hidden layers of the expert network use the SigmaRhodes function as the activation function, and the output layer of the expert network uses the hyperbolic tangent function as the activation function. The number of neurons in the hidden layer of the expert network is less than the number of neurons in the hidden layer of a single global neural network model constructed for the same modeling task. The number of expert networks, the number of hidden layers and the number of neurons in each expert network are set according to the variation range of the physical structure parameters of the semiconductor device and the preset modeling accuracy requirements.

[0045] It should be noted that constructing the gated network refers to determining the specific network structure and parameter initialization method of the gated network before training the hybrid expert neural network model. The input layer refers to the first layer in the neural network used to receive external input data, and the number of neurons in the input layer matches the dimensionality of the input data. The hidden layer refers to the intermediate layer in the neural network located between the input and output layers, used for nonlinear feature transformation and abstract representation extraction of the input data. The number of hidden layers and the number of neurons in each hidden layer jointly determine the expressive power and model complexity of the network. The output layer refers to the last layer in the neural network, used to generate the final output result of the network. The hyperbolic tangent function is a nonlinear activation function centered at zero with an output value range between -1 and +1. Using the hyperbolic tangent function helps alleviate the gradient vanishing problem during neural network training. The Softmax function is a normalized exponential function that maps any real vector to a probability distribution space. Using the Softmax function as the activation function of the gated network's output layer ensures that the sum of all elements in the output weight allocation vector equals one, thus providing probabilistic interpretation of the weight allocation among expert networks. The Sigmaroid function is an S-shaped nonlinear activation function with an output value range between zero and one, suitable for capturing the smooth saturation characteristics of device electrical properties. The single global neural network model refers to a modeling method in related technologies that uses a deep neural network to uniformly fit all sample data within the entire parameter space. The modeling accuracy requirement refers to a pre-set threshold for measuring the deviation between the model's predicted results and the actual electrical property values, including but not limited to at least one of the following: mean squared error threshold, maximum relative error threshold, or coefficient of determination threshold.

[0046] It is understandable that by pre-constructing the gating network and the multiple expert networks before parallel pre-training, the overall architecture of the hybrid expert neural network model and the specific structural parameters of each component sub-network are clarified, providing a definite network structure foundation for subsequent staged training. By setting the hidden layers of the gating network to use the hyperbolic tangent function as the activation function, the gating network can learn a richer nonlinear mapping relationship between input parameters and output weights during training. By setting the output layers of the gating network to use the Softmax function as the activation function, it is ensured that the weight allocation vector output by the gating network is a normalized non-negative weight, thus giving the contribution ratio of each expert network to the final prediction result a clear physical meaning. By setting the hidden layers of each expert network to use the Sigmar function as the activation function, the expert network can effectively fit the saturation region characteristics commonly found in the electrical characteristics of semiconductor devices. By setting the output layers of each expert network to use the hyperbolic tangent function as the activation function, the intermediate predicted values ​​output by the expert networks can be consistent with the target value range of the normalized electrical characteristic data. By setting the number of hidden layer neurons in the expert network to be less than that in the single global neural network model, the computational complexity of a single expert network is effectively controlled, while the total computational cost of combining multiple expert networks remains within a reasonable range. By setting the number of expert networks and the specific structure of each expert network according to the variation range of the physical structure parameters and the preset modeling accuracy requirements, the overall size of the hybrid expert neural network model can be flexibly adjusted according to the complexity of the actual modeling task, achieving a balance between modeling accuracy and computational efficiency.

[0047] In one feasible implementation, the step of training the hybrid expert neural network model using the complete sample data, wherein only the network parameters of the gating network are updated during the overall training process, includes: The physical structure parameters in the complete sample data are input into the gating network to obtain the weight allocation vectors corresponding to each expert network output by the gating network. The physical structure parameters and working state parameters from the complete sample data are input into each of the expert networks whose parameters have been fixed, to obtain the intermediate prediction values ​​output by each of the expert networks. The intermediate predicted values ​​are weighted and fused according to the weight allocation vector to obtain the global electrical characteristic predicted value. Using the actual electrical characteristic values ​​in the complete sample data as the supervision signal, the error between the global electrical characteristic prediction value and the actual electrical characteristic value is calculated by minimizing a preset composite loss function, and the network parameters of the gated network are updated only based on the backpropagation of the error. The composite loss function is composed of a weighted sum of a first error term, a second error term, a third error term, and a fourth error term. The first error term is the mean square error between the predicted and actual values ​​of the electrical characteristic data. The second error term is the mean square error between the predicted and actual values ​​of the logarithm of the electrical characteristic data (base 10). The third error term is the mean square error between the predicted and actual values ​​of the partial derivative of the electrical characteristic data with respect to the first bias voltage in the operating state parameters. The fourth error term is the mean square error between the predicted and actual values ​​of the partial derivative of the electrical characteristic data with respect to the second bias voltage in the operating state parameters.

[0048] It should be noted that "complete" refers to the sample data set containing all sample points within the entire preset range of variation of the physical structure parameters, without undergoing data subset partitioning. The weight allocation vector is a multi-dimensional vector generated by the gating network based on the input physical structure parameters. The dimension of the weight allocation vector is equal to the number of expert networks, and each element in the weight allocation vector corresponds to the contribution weight of an expert network under the current input conditions. The intermediate prediction value refers to the electrical characteristic prediction result independently calculated and output by each expert network based on the input physical structure parameters and the operating state parameters. Due to the different data subsets learned during their pre-training phase, the accuracy of the intermediate prediction values ​​output by different expert networks varies across different physical structure parameter regions. The weighted fusion calculation refers to the process of weighted summation of the intermediate prediction values ​​output by each expert network with the corresponding weight coefficients in the weight allocation vector. The global electrical characteristic prediction value is the final prediction output value obtained after the weighted fusion calculation. This prediction value integrates the prediction results of all expert networks and is dynamically weighted by the gating network. The true electrical characteristic values ​​refer to the electrical characteristic values ​​actually recorded in the sample data, obtained through process computer-aided design simulation or actual measurement, including but not limited to the actual measured values ​​of drain current. The supervisory signal refers to the real data used as the reference value for the prediction target during supervised learning. The composite loss function refers to a comprehensive error evaluation function formed by weighting and summing multiple error terms with different physical meanings using preset weight coefficients. The error backpropagation refers to the process of passing the calculated loss function value layer by layer from the network output layer to the input layer and calculating the gradient of the network parameters at each layer. Updating only the network parameters of the gated network means that during the error backpropagation process, only the weight parameters and bias parameters of the gated network are updated and adjusted based on gradients, while no update operation is performed on the parameters of each expert network. The first error term refers to the mean square error of the deviation between the global electrical characteristic prediction value and the true electrical characteristic value, used to constrain the model's fitting accuracy to the main trend of electrical characteristic changes. The second error term refers to the mean square error of the deviation between the base-10 logarithmic value of the predicted global electrical characteristic value and the base-10 logarithmic value of the actual electrical characteristic value. This term enhances the model's ability to distinguish minute changes when the electrical characteristic value is small. The third error term refers to the mean square error of the deviation between the predicted and actual values ​​of the partial derivative of the electrical characteristic data with respect to the first bias voltage in the operating state parameters. The first bias voltage includes, but is not limited to, the gate-source voltage. The partial derivative, i.e., the transconductance, is used to constrain the model's accuracy in fitting the gate control capability.The fourth error term refers to the mean square error of the deviation between the predicted value and the true value of the partial derivative of the second bias voltage in the operating state parameters of the electrical characteristic data. The second bias voltage includes, but is not limited to, the drain-source voltage. The partial derivative is the channel conductance, which is used to constrain the accuracy of the model's fitting of the shape of the output characteristic curve.

[0049] It is understood that by inputting the physical structure parameters from the complete sample data into the gating network, the gating network can generate an optimal weight allocation strategy for the current input based on information from the global parameter space. By inputting the physical structure parameters and operating state parameters from the complete sample data into each of the expert networks whose parameters have been fixed, each expert network can output its own predictions based on local parameter region knowledge in parallel. Since the parameters are fixed, the local patterns acquired by the expert networks during the pre-training phase are fully preserved. By performing weighted fusion calculations on each intermediate prediction value according to the weight allocation vector, the final output global electrical characteristic prediction value can dynamically favor the expert network with higher prediction accuracy at the location of the input physical structure parameters, achieving a soft combination of knowledge from different regions. By using the real electrical characteristic values ​​as supervision signals and minimizing the composite loss function, a clear and multi-dimensional optimization objective is provided for the training of the gating network. By employing a composite loss function consisting of a weighted sum of the first, second, third, and fourth error terms, the gated network, during training, not only focuses on the fitting accuracy of the electrical characteristic values ​​themselves, but also on the logarithmic accuracy of the electrical characteristics in the low-current region and the consistency of physical trends at the level of first-order partial derivatives. This allows the trained gated network to better coordinate the outputs of the various expert networks, ensuring overall fitting accuracy while making the prediction results more physically plausible. By updating only the network parameters of the gated network through error backpropagation, the optimization problem in the second training phase involves only this small sub-network, significantly accelerating the convergence speed of the training process and avoiding the potential disruption of the learned knowledge of the expert networks caused by global joint training.

[0050] In one feasible implementation, the step of inputting the target physical structure parameters and target operating state parameters corresponding to the semiconductor device to be modeled into the trained hybrid expert neural network model, and having the gating network output a weight allocation vector, includes: In the gated network, an initial weight allocation vector is generated based on the target physical structure parameters; A predetermined number of weight coefficients with the largest values ​​are retained in the initial weight allocation vector, and the remaining weight coefficients in the initial weight allocation vector are set to zero to form the weight allocation vector; When performing the weighted fusion calculation on the weight allocation vector and the intermediate predicted value, only the expert network corresponding to the retained weight coefficients participates in the calculation, while the expert network corresponding to the weight coefficients set to zero does not participate in the calculation during the current prediction process.

[0051] It should be noted that the initial weight allocation vector refers to the original weight allocation vector generated by the gating network based on the input target physical structure parameters, without sparsification. Each element in the initial weight allocation vector is a non-negative value, and the sum of all elements is equal to one. The preset number of largest weight coefficients refers to the preset number of element values ​​selected from the elements of the initial weight allocation vector after sorting them in descending order of value. The preset number is a pre-set positive integer, and the preset number is less than the total number of expert networks. The remaining weight coefficients refer to all weight coefficients in the initial weight allocation vector other than the preset number of retained weight coefficients. Setting them to zero means forcibly modifying the values ​​of the remaining weight coefficients to zero. The weight allocation vector formed after the zeroing operation contains only the preset number of non-zero weight coefficients, and the remaining elements are all zero. The current prediction process refers to a complete model inference process performed on the currently input target physical structure parameters and target working state parameters. The phrase "not participating in the calculation" means that the corresponding expert network does not perform forward propagation operations during the current prediction process, that is, it does not calculate and output the intermediate prediction value based on the input target physical structure parameters and the target working state parameters, thereby not generating corresponding computational resource consumption.

[0052] Understandably, by first generating an initial weight allocation vector based on the target physical structure parameters in the gating network, the gating network can perform a preliminary evaluation of the applicability of each expert network under the current input conditions. By retaining a preset number of weight coefficients with the largest values ​​in the initial weight allocation vector and setting the remaining weight coefficients to zero, sparse activation of the expert networks is achieved, ensuring that only a portion of the expert networks whose predictive capabilities best match the current input are actually invoked during each inference process. Since only the expert networks corresponding to the retained weight coefficients participate in the calculation during the weighted fusion calculation, and the expert networks corresponding to the weight coefficients set to zero do not participate in the calculation during the current prediction process, the computational complexity and memory access overhead of the model inference stage are effectively reduced while ensuring prediction accuracy. For modeling tasks deployed on edge devices with limited computing resources, the above-mentioned sparse activation mechanism can significantly improve the inference speed of the model and reduce power consumption, thereby improving the deployment feasibility of the hybrid expert neural network model in practical engineering applications.

[0053] For example, to help understand the implementation process of the global modeling method based on a hybrid expert neural network obtained by combining this embodiment with the above embodiment one, please refer to... Figure 2 , Figure 2 A simplified flowchart of a global modeling method based on a hybrid expert neural network is provided, specifically: like Figure 2 As shown, it illustrates the backpropagation and loss calculation process during the overall training of the hybrid expert neural network model in the second training phase of this application embodiment.

[0054] In this overall training phase, complete sample data is input into the hybrid expert neural network model sequentially or in batches. Specifically, the physical structure parameters in the sample data first enter the gating network, which outputs weight allocation vectors corresponding to each expert network after forward propagation. Simultaneously, the physical structure parameters and operating state parameters from the sample data are input in parallel to each expert network whose parameters are fixed. Each expert network outputs its own intermediate prediction value after its own forward propagation. Subsequently, the weight allocation vectors and the intermediate prediction values ​​are weighted and fused to obtain the global electrical characteristic prediction value.

[0055] After obtaining the predicted global electrical properties, they are compared with the actual electrical property values ​​recorded in the sample data, and the error between the two is calculated using a preset composite loss function. This composite loss function integrates the equivalence error of the electrical property values ​​themselves, the logarithmic domain error, and the error at the first-order partial derivative level, thereby evaluating the accuracy of the prediction results from multiple physical dimensions.

[0056] After calculating the loss function value, the backpropagation phase begins. For example... Figure 2 As shown in the backpropagation path, the prediction error signal, represented by the loss function value, is propagated back layer by layer from the model's output to its input. It is important to note that since the network parameters of each expert network are pre-fixed at this stage, the parameter gradients generated during backpropagation only affect the gating network. Specifically, during backpropagation, the error signal only calculates gradients for the connection weights and bias parameters between neurons in each layer of the gating network, and updates and adjusts the network parameters of the gating network based on these gradients. The network parameters of each expert network remain unchanged throughout the entire backpropagation process.

[0057] Through the iterative cycle of loss calculation and backpropagation described above, the network parameters of the gated network are gradually optimized until the composite loss function converges to a preset threshold range. After the overall training phase described above, the gated network has the ability to accurately allocate the weights of each expert network under different input physical structure parameters, thereby generating the optimal weight allocation strategy based on the specific physical structure parameters of the semiconductor device to be modeled in the subsequent model inference phase.

[0058] like Figure 3 As shown, the global modeling method for scalable devices based on hybrid expert neural networks proposed in this application first acquires sample data of the scalable device to be modeled under different combinations of physical structure parameters and operating state parameters, and normalizes the sample data to eliminate the influence of differences in the dimensions of each parameter on the training process. Then, based on the distribution characteristics of the physical structure parameters, the normalized sample data is divided into multiple data subsets, each corresponding to a specific sub-region in the physical structure parameter space. In the first training stage, each data subset is used to perform parallel pre-training on the corresponding expert network in the hybrid expert neural network model, enabling each expert network to focus on learning the changes in the electrical characteristics of the device within its corresponding sub-region. After completing the parallel pre-training, the network parameters of each expert network are fixed, and the second training stage begins. In this stage, the complete sample data is used to train the hybrid expert neural network model as a whole, and only the network parameters of the gating network are updated during the training process, allowing the gating network to learn to assign smoothly transitioning weight coefficients to each expert network according to the input physical structure parameters. After the model training is completed, the target physical structure parameters and target operating state parameters of the semiconductor device to be modeled are input into the model. The gating network outputs the weight allocation vector, and the expert networks output the intermediate prediction values. By performing weighted fusion calculation on the weight allocation vector and the intermediate prediction values, the global modeling result of the semiconductor device to be modeled is finally obtained.

[0059] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the global modeling method based on hybrid expert neural networks in this application. Any simple transformations based on this technical concept are within the protection scope of this application.

[0060] This application also provides a global modeling device based on a hybrid expert neural network, please refer to... Figure 4 The global modeling device based on a hybrid expert neural network includes: The acquisition module 10 is used to acquire sample data of the semiconductor device to be modeled; Training module 20 is used to train a preset hybrid expert neural network model in stages based on the sample data. The hybrid expert neural network model includes a gating network and multiple expert networks. In the first training stage, only the expert networks are trained. In the second training stage, the expert networks are fixed and the gating networks are trained. The modeling module 30 is used to input the target physical structure parameters and target operating state parameters corresponding to the semiconductor device to be modeled into the trained hybrid expert neural network model, the gating network outputs a weight allocation vector, the expert network outputs an intermediate prediction value, and the weight allocation vector and the intermediate prediction value are weighted and fused to obtain the global modeling result of the semiconductor device to be modeled.

[0061] And / or, the training module includes: The normalization module is used to normalize the sample data; The subset division module is used to divide the normalized sample data into multiple data subsets based on the distribution of physical structure parameters in the sample data. The first training module is used to perform parallel pre-training of the expert networks corresponding to each of the respective data subsets in the hybrid expert neural network model, wherein the gated network is configured to receive the physical structure parameters, and the expert network is configured to receive the physical structure parameters and the working state parameters in the sample data; A parameter fixing module is used to fix the network parameters of each of the expert networks after parallel pre-training. The second training module is used to train the hybrid expert neural network model as a whole using the complete sample data. During the overall training process, only the network parameters of the gating network are updated, so that the gating network learns to assign smooth transition weight coefficients to each expert network according to the input physical structure parameters, so as to achieve continuous modeling at the boundaries of different physical structure parameter regions.

[0062] And / or, the subset partitioning module includes: The first partitioning submodule is used to divide the parameter space composed of the physical structure parameters into multiple non-overlapping and continuous regions based on the critical size threshold for the occurrence of short-channel effect in the semiconductor device. The second partitioning submodule is used to group the sample data belonging to the same region into the same data subset, such that each data subset uniquely corresponds to a specific subregion in the parameter space. The first focused learning submodule is used to enable each expert network to focus on learning the changes in the electrical characteristics of the device within its corresponding sub-region through a partitioning operation during the parallel pre-training phase. The first coverage submodule is used to make adjacent regions seamlessly connected in parameter space after the division operation, and to completely cover the entire preset range of variation of the physical structure parameters.

[0063] And / or, the global modeling apparatus based on a hybrid expert neural network further includes: The first construction module is used to construct the gated network, which consists of an input layer, at least two hidden layers and an output layer connected in sequence. The hidden layers use the hyperbolic tangent function as the activation function. The number of neurons in the output layer is equal to the number of neurons in the expert network, and the output layer uses the Softmax function as the activation function. The second construction module is used to construct the plurality of expert networks. Each expert network consists of an input layer, at least two hidden layers and an output layer connected in sequence. The hidden layers of the expert network use the SigmaRhythm function as the activation function, and the output layer of the expert network uses the hyperbolic tangent function as the activation function. The first setting module is used to set the number of neurons in the hidden layer of the expert network to be less than the number of neurons in the hidden layer of a single global neural network model constructed for the same modeling task. The second setting module is used to set the number of expert networks and the number of hidden layers and neurons of each expert network according to the variation range of the physical structure parameters of the semiconductor device and the preset modeling accuracy requirements.

[0064] And / or, the second training module includes: The first input submodule is used to input the physical structure parameters in the complete sample data into the gating network to obtain the weight allocation vectors output by the gating network corresponding to each of the expert networks; The second input submodule is used to input the physical structure parameters and working state parameters from the complete sample data into each of the expert networks whose parameters have been fixed, so as to obtain the intermediate prediction values ​​output by each of the expert networks. The first weighted fusion submodule is used to perform weighted fusion calculation on each of the intermediate prediction values ​​according to the weight allocation vector to obtain the global electrical characteristic prediction value. The first loss calculation submodule is used to calculate the error between the global electrical characteristic prediction value and the actual electrical characteristic value by minimizing a preset composite loss function, using the real electrical characteristic value in the complete sample data as a supervision signal. The first backpropagation submodule is used to backpropagate the error to update only the network parameters of the gated network. The composite loss function is composed of a weighted sum of a first error term, a second error term, a third error term, and a fourth error term. The first error term is the mean square error between the predicted and actual values ​​of the electrical characteristic data. The second error term is the mean square error between the predicted and actual values ​​of the logarithm of the electrical characteristic data (base 10). The third error term is the mean square error between the predicted and actual values ​​of the partial derivative of the electrical characteristic data with respect to the first bias voltage in the operating state parameters. The fourth error term is the mean square error between the predicted and actual values ​​of the partial derivative of the electrical characteristic data with respect to the second bias voltage in the operating state parameters.

[0065] And / or, the modeling module includes: The first generation submodule is used to generate an initial weight allocation vector in the gating network based on the target physical structure parameters. The first sparse submodule is used to retain a preset number of weight coefficients with the largest values ​​in the initial weight allocation vector, and set the remaining weight coefficients in the initial weight allocation vector to zero to form the weight allocation vector; The first participating calculation submodule is used to ensure that, when performing weighted fusion calculation on the weight allocation vector and the intermediate predicted value, only the expert network corresponding to the retained weight coefficients participates in the calculation, and the expert network corresponding to the weight coefficients set to zero does not participate in the calculation during the current prediction process.

[0066] The global modeling device based on hybrid expert neural networks provided in this application, employing the global modeling method based on hybrid expert neural networks in the above embodiments, can solve the technical problem in related technologies that global modeling of scalable devices cannot simultaneously ensure modeling accuracy and region boundary continuity. Compared with related technologies, the beneficial effects of the global modeling device based on hybrid expert neural networks provided in this application are the same as those of the global modeling method based on hybrid expert neural networks provided in the above embodiments, and other technical features in the global modeling device based on hybrid expert neural networks are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0067] This application provides a global modeling device based on a hybrid expert neural network. The global modeling device based on a hybrid expert neural network includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the global modeling method based on a hybrid expert neural network in the first embodiment described above.

[0068] The following is for reference. Figure 5 This document illustrates a structural schematic diagram of a global modeling device based on a hybrid expert neural network suitable for implementing embodiments of this application. The global modeling device based on a hybrid expert neural network in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, tablets, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital televisions and desktop computers. Figure 5 The global modeling device based on a hybrid expert neural network shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0069] like Figure 5As shown, the global modeling device based on a hybrid expert neural network may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the global modeling device based on the hybrid expert neural network. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the global modeling device based on hybrid expert neural networks to communicate wirelessly or wiredly with other devices to exchange data. Although a global modeling device based on hybrid expert neural networks with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0070] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0071] The global modeling device based on hybrid expert neural networks provided in this application, employing the global modeling method based on hybrid expert neural networks in the above embodiments, can solve the technical problem in related technologies that global modeling of scalable devices cannot simultaneously ensure modeling accuracy and region boundary continuity. Compared with related technologies, the beneficial effects of the global modeling device based on hybrid expert neural networks provided in this application are the same as those of the global modeling method based on hybrid expert neural networks provided in the above embodiments, and other technical features in this global modeling device based on hybrid expert neural networks are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0072] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0073] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0074] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the global modeling method based on a hybrid expert neural network in the above embodiments.

[0075] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0076] The aforementioned computer-readable storage medium may be included in a global modeling device based on a hybrid expert neural network; or it may exist independently and not assembled into a global modeling device based on a hybrid expert neural network.

[0077] The aforementioned computer-readable storage medium carries one or more programs that, when executed by a global modeling device based on a hybrid expert neural network, cause the global modeling device based on a hybrid expert neural network to: acquire sample data of the semiconductor device to be modeled; Based on the sample data, a preset hybrid expert neural network model is trained in stages. The hybrid expert neural network model includes a gating network and multiple expert networks. In the first training stage, only the expert networks are trained. In the second training stage, the expert networks are fixed and the gating network is trained. The target physical structure parameters and target operating state parameters corresponding to the semiconductor device to be modeled are input into the trained hybrid expert neural network model. The gating network outputs a weight allocation vector, the expert network outputs an intermediate prediction value, and the weight allocation vector and the intermediate prediction value are weighted and fused to obtain the global modeling result of the semiconductor device to be modeled.

[0078] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0079] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0080] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0081] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned global modeling method based on a hybrid expert neural network. This solves the technical problem in related technologies where global modeling of scalable devices cannot simultaneously achieve both modeling accuracy and region boundary continuity. Compared with related technologies, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the global modeling method based on a hybrid expert neural network provided in the above embodiments, and will not be elaborated upon here.

[0082] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the global modeling method based on a hybrid expert neural network as described above.

[0083] The computer program product provided in this application can solve the technical problem in related technologies that global modeling of scalable devices cannot simultaneously ensure modeling accuracy and the continuity of region boundaries. Compared with related technologies, the beneficial effects of the computer program product provided in this application are the same as those of the global modeling method based on hybrid expert neural networks provided in the above embodiments, and will not be repeated here.

[0084] All acquisition of signals, information, or actions in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization of the relevant device owner.

[0085] The above description is only a part of the embodiments of this application and does not limit the scope of protection of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the scope of protection of this application.

Claims

1. A global modeling method based on a hybrid expert neural network, characterized in that, Includes the following steps: Obtain sample data of the semiconductor device to be modeled; Based on the sample data, a preset hybrid expert neural network model is trained in stages. The hybrid expert neural network model includes a gating network and multiple expert networks. In the first training stage, only the expert networks are trained. In the second training stage, the expert networks are fixed and the gating network is trained. The target physical structure parameters and target operating state parameters corresponding to the semiconductor device to be modeled are input into the trained hybrid expert neural network model. The gating network outputs a weight allocation vector, the expert network outputs an intermediate prediction value, and the weight allocation vector and the intermediate prediction value are weighted and fused to obtain the global modeling result of the semiconductor device to be modeled. The step of training the preset hybrid expert neural network model in stages based on the sample data includes: The sample data is normalized. Based on the distribution of physical structure parameters in the sample data, the normalized sample data is divided into multiple data subsets; The expert networks corresponding to each of the aforementioned data subsets are pre-trained in parallel using the data in the hybrid expert neural network model. The gating network is configured to receive the physical structure parameters, and the expert network is configured to receive the physical structure parameters and the working state parameters in the sample data. The network parameters of each of the expert networks after the parallel pre-training are fixed; The hybrid expert neural network model is trained as a whole using the complete sample data. During the overall training process, only the network parameters of the gating network are updated, so that the gating network learns to assign smooth transition weight coefficients to each expert network according to the input physical structure parameters, thereby achieving continuous modeling at the boundaries of different physical structure parameter regions.

2. The method as described in claim 1, characterized in that, The step of dividing the normalized sample data into multiple data subsets based on the distribution of physical structure parameters in the sample data includes: Based on the critical size threshold for the occurrence of short-channel effect in the semiconductor device, the parameter space composed of the physical structure parameters is divided into multiple non-overlapping regions with continuous boundaries. The sample data belonging to the same region are grouped into the same data subset, such that each data subset uniquely corresponds to a specific sub-region in the parameter space; The partitioning operation enables each expert network to focus on learning the changes in the electrical characteristics of the device within its corresponding sub-region during the parallel pre-training phase. After the division operation, adjacent regions are seamlessly connected in parameter space and completely cover the entire preset range of variation of the physical structure parameters.

3. The method as described in claim 1, characterized in that, Before the step of performing parallel pre-training of the respective expert networks in the hybrid expert neural network model using each of the data subsets, the method further includes: The gated network is constructed by sequentially connecting an input layer, at least two hidden layers, and an output layer. The hidden layers use the hyperbolic tangent function as the activation function. The number of neurons in the output layer is equal to the number of neurons in the expert network, and the output layer uses the Softmax function as the activation function. Construct the plurality of expert networks, each of which consists of an input layer, at least two hidden layers and an output layer connected in sequence. The hidden layers of the expert network use the SigmaRhodes function as the activation function, and the output layer of the expert network uses the hyperbolic tangent function as the activation function. The number of neurons in the hidden layer of the expert network is less than the number of neurons in the hidden layer of a single global neural network model constructed for the same modeling task. The number of expert networks, the number of hidden layers and the number of neurons in each expert network are set according to the variation range of the physical structure parameters of the semiconductor device and the preset modeling accuracy requirements.

4. The method as described in claim 1, characterized in that, The step of training the hybrid expert neural network model using the complete sample data, and updating only the network parameters of the gating network during the overall training process, includes: The physical structure parameters in the complete sample data are input into the gating network to obtain the weight allocation vectors corresponding to each expert network output by the gating network. The physical structure parameters and working state parameters from the complete sample data are input into each of the expert networks whose parameters have been fixed, to obtain the intermediate prediction values ​​output by each of the expert networks. The intermediate predicted values ​​are weighted and fused according to the weight allocation vector to obtain the global electrical characteristic predicted value. Using the actual electrical characteristic values ​​in the complete sample data as the supervision signal, the error between the global electrical characteristic prediction value and the actual electrical characteristic value is calculated by minimizing a preset composite loss function, and the network parameters of the gated network are updated only based on the backpropagation of the error. The composite loss function is composed of a weighted sum of a first error term, a second error term, a third error term, and a fourth error term. The first error term is the mean square error between the predicted and actual values ​​of the electrical characteristic data. The second error term is the mean square error between the predicted and actual values ​​of the logarithm of the electrical characteristic data (base 10). The third error term is the mean square error between the predicted and actual values ​​of the partial derivative of the electrical characteristic data with respect to the first bias voltage in the operating state parameters. The fourth error term is the mean square error between the predicted and actual values ​​of the partial derivative of the electrical characteristic data with respect to the second bias voltage in the operating state parameters.

5. The method as described in claim 1, characterized in that, The step of inputting the target physical structure parameters and target operating state parameters corresponding to the semiconductor device to be modeled into the trained hybrid expert neural network model, and having the gating network output a weight allocation vector, includes: In the gated network, an initial weight allocation vector is generated based on the target physical structure parameters; A predetermined number of weight coefficients with the largest values ​​are retained in the initial weight allocation vector, and the remaining weight coefficients in the initial weight allocation vector are set to zero to form the weight allocation vector; When performing the weighted fusion calculation on the weight allocation vector and the intermediate predicted value, only the expert network corresponding to the retained weight coefficients participates in the calculation, while the expert network corresponding to the weight coefficients set to zero does not participate in the calculation during the current prediction process.

6. A global modeling device based on a hybrid expert neural network, characterized in that, The device includes: The acquisition module is used to acquire sample data of the semiconductor device to be modeled; The training module is used to train a preset hybrid expert neural network model in stages based on the sample data. The hybrid expert neural network model includes a gating network and multiple expert networks. In the first training stage, only the expert networks are trained. In the second training stage, the expert networks are fixed and the gating networks are trained. The modeling module is used to input the target physical structure parameters and target operating state parameters corresponding to the semiconductor device to be modeled into the trained hybrid expert neural network model. The gating network outputs a weight allocation vector, the expert network outputs an intermediate prediction value, and the weight allocation vector and the intermediate prediction value are weighted and fused to obtain the global modeling result of the semiconductor device to be modeled. The training module is also used to achieve: The sample data is normalized. Based on the distribution of physical structure parameters in the sample data, the normalized sample data is divided into multiple data subsets; The expert networks corresponding to each of the aforementioned data subsets are pre-trained in parallel using the data in the hybrid expert neural network model. The gating network is configured to receive the physical structure parameters, and the expert network is configured to receive the physical structure parameters and the working state parameters in the sample data. The network parameters of each of the expert networks after the parallel pre-training are fixed; The hybrid expert neural network model is trained as a whole using the complete sample data. During the overall training process, only the network parameters of the gating network are updated, so that the gating network learns to assign smooth transition weight coefficients to each expert network according to the input physical structure parameters, thereby achieving continuous modeling at the boundaries of different physical structure parameter regions.

7. A global modeling device based on a hybrid expert neural network, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the global modeling method based on a hybrid expert neural network as described in any one of claims 1 to 5.

8. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the global modeling method based on a hybrid expert neural network as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the global modeling method based on a hybrid expert neural network as described in any one of claims 1 to 5.

Images