Physical and neural network integrated fusion model construction method, equipment and medium

Abstract

The invention discloses a method and equipment for constructing a fusion model integrating physics and a neural network, and a medium, and relates to the technical field of circuit simulation, and the method comprises the steps: constructing a neural network according to an instance parameter of an electronic device and a first correction coefficient; constructing a fusion module based on the physical information of the device and a neural network; and constructing a fusion model according to the at least one fusion module and a physical model determined based on the device characteristics. According to the fusion model, physical rule constraint is carried out on the output of the neural network by utilizing the physical information of the device and the physical model, so that the physical interpretability is kept, and the strong nonlinear capture capability of the neural network on the complex physical parameter change is achieved, thereby improving the simulation accuracy of the fusion model.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and particularly to a method, device, and medium for constructing a fusion model integrating physics and neural networks. Background Art

[0002] With the development of semiconductor technology, the feature size of transistors has been reduced to the nanometer level, driving the development of electronic devices towards more miniaturized and efficient directions, and promoting the improvement of circuit performance. However, the changes in the electrical characteristics of these electronic devices have increased the complexity of the device compact models used for SPICE (Simulation Program with Integrated Circuit Emphasis) simulation.

[0003] The construction of compact models often relies on physical models and complex mathematical derivations. In the process of constructing compact models, if a unified model parameter set (i.e., a global model) is used for modeling, it is difficult to provide accurate simulation results within the size range of all electronic devices in the simulated circuit. This is because the global model parameters need to cover a wide range of operating conditions such as device size, voltage bias, and temperature, but their mathematical expressions are often overly simplified and unable to capture the complex coupling relationships of non-linear phenomena such as short-channel effects and quantum tunneling in nanoscale devices, resulting in a significant increase in the prediction deviation of the model under small sizes or extreme biases, that is, the accuracy of the compact model for predicting device characteristics is relatively low.

[0004] The above content is only used to assist in understanding the technical solution of the present application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of the present application is to provide a method, device, and medium for constructing a fusion model integrating physics and neural networks, aiming to improve the simulation accuracy of the model.

[0006] To achieve the above purpose, the present application proposes a method for constructing a fusion model integrating physics and neural networks, and the method includes:

[0007] Construct a neural network according to the instance parameters of the electronic device and the first correction coefficient;

[0008] Construct a fusion module based on the device physical information and the neural network;

[0009] Construct a fusion model according to at least one of the fusion modules and a physical model determined based on device characteristics.

[0010] In one embodiment, after the step of constructing the fusion model, it further includes:

[0011] Construct a target loss function for the fusion model according to the electrical characteristics of the electronic device;

[0012] Based on the backpropagation algorithm, adjust the weights and biases of the fusion model until the target loss function is minimized.

[0013] In one embodiment, the fusion model includes a current model, and the electrical characteristics include: linear current, logarithmically processed current, logarithmically processed output conductance, and linear transconductance. The step of constructing the target loss function of the fusion model according to the electrical characteristics of the electronic device includes:

[0014] Construct a first loss function for the current model according to the linear current, the logarithmically processed current, the logarithmically processed output conductance, and the linear transconductance.

[0015] In one embodiment, after the step of constructing the fusion model, the method further includes:

[0016] Generate offset data of the electronic device through the fusion module and the instance parameters;

[0017] Determine the device characteristics of the electronic device through the physical model, the offset data, the bias voltage configuration information of the electronic device, and the instance parameters.

[0018] In one embodiment, the step of generating offset data of the electronic device through the fusion module and the instance parameters includes:

[0019] Determine a second correction coefficient of the electronic device through the neural network and the instance parameters;

[0020] Generate the offset data according to the device physical information and the second correction coefficient.

[0021] In one embodiment, the device physical information includes a short-channel effect equation, and the offset data includes a threshold voltage offset. The step of generating the offset data according to the device physical information and the second correction coefficient includes:

[0022] Determine the threshold voltage offset according to the short-channel effect equation and the second correction coefficient.

[0023] In one embodiment, the device physical information includes a mobility degradation equation, and the offset data includes a mobility degradation factor. The step of generating the offset data according to the device physical information and the second correction coefficient includes:

[0024] Determine the mobility degradation factor according to the mobility degradation equation and the second correction coefficient.

[0025] In one embodiment, after the step of constructing the fusion model, the method further includes:

[0026] Optimizing hyperparameters of the fusion model based on the Bayesian optimization algorithm.

[0027] In addition, to achieve the above object, the present application further provides a fusion model construction device integrating physics and neural networks, where the fusion model construction device integrating physics and neural networks includes:

[0028] A neural network construction module, configured to construct a neural network according to instance parameters of an electronic device and a first correction coefficient;

[0029] A fusion module construction module, configured to construct a fusion module based on device physical information and the neural network;

[0030] A fusion model construction module, configured to construct a fusion model according to at least one of the fusion modules and a physical model determined based on device characteristics.

[0031] In addition, to achieve the above object, the present application further provides an electronic device, where the device includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the computer program is configured to implement the steps of the above-mentioned fusion model construction method integrating physics and neural networks.

[0032] In addition, to achieve the above object, the present application further provides a storage medium, where the storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium, and when the computer program is executed by a processor, the steps of the above-mentioned fusion model construction method integrating physics and neural networks are implemented.

[0033] In addition, to achieve the above object, the present application further provides a computer program product, where the computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned fusion model construction method integrating physics and neural networks are implemented.

[0034] One or more technical solutions proposed in this application have at least the following technical effects: First, a neural network is constructed based on the instance parameters of the electronic device and the first correction coefficient, so that the powerful non-linear fitting ability of the neural network can be fully utilized to ensure the accuracy of the correction coefficient solved during the application of the model. Furthermore, a fusion module is constructed based on the device physical information and the neural network. This fusion module combines the physical information with the neural network, and by introducing the physical principle of the device as prior knowledge, it constrains and optimizes the output of the neural network, thereby enhancing the interpretability and generalization ability of the model. Furthermore, a fusion model is constructed according to at least one fusion module and the physical model determined based on the device characteristics, enabling the fusion model to more accurately reflect the device characteristics of the electronic device, thereby improving the simulation accuracy of the fusion model. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings herein are incorporated into and constitute a part of this specification, showing embodiments consistent with this application, and are used together with the specification to explain the principles of this application.

[0036] In order to more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0037] Figure 1 It is a schematic flowchart provided for Embodiment 1 of the method for constructing a fusion model integrating physics and neural network in this application;

[0038] Figure 2 It is the overall framework diagram of the fusion model provided for Embodiment 1 of this application;

[0039] Figure 3 It is the voltage-current characteristic curve graph provided for Embodiment 3 of this application;

[0040] Figure 4 It is the capacitance characteristic curve graph provided for Embodiment 3 of this application;

[0041] Figure 5 It is the module structure schematic diagram of the device for constructing a fusion model integrating physics and neural network in the embodiment of this application;

[0042] Figure 6 It is the device structure schematic diagram of the hardware operating environment involved in the method for constructing a fusion model integrating physics and neural network in the embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of this application and are not used to limit this application.

[0044] To better understand the technical solution of this application, the following will be described in detail in conjunction with the accompanying drawings of the specification and specific implementation manners.

[0045] It should be noted that in this application, the "first correction coefficient" regarding the correction coefficient represents the quantity of the correction coefficient, and the "second correction coefficient" represents the specific value determined by the neural network during subsequent application processes. Thus, "first" and "second" are only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features.

[0046] In this embodiment, for the sake of convenience in description, the compiler integrated with the fusion model will be used as the execution subject for elaboration below.

[0047] The conventional intensive model modeling method also includes neural network modeling. Although the neural network model has high accuracy for circuit simulation results, the instance parameters of electronic devices in the simulated circuit are diverse. When training the neural network model, a large amount of data sets also need to be provided for it, and a large amount of resource investment is required for neural network modeling; the black-box characteristic of the neural network also makes it difficult to explain the specific derivation of the circuit simulation results. At the same time, due to the possible overfitting of the training data during the neural network training process, the model has poor generalization ability for unseen data and is difficult to provide accurate simulation results for the electronic devices not trained in the simulated circuit; since physical constraints are not considered, the data results (circuit simulation results) of the neural network model may not conform to physical laws and cannot meet the characteristics that natural physical models such as zero input-zero output satisfy.

[0048] The present application provides a fusion model integrating physics and neural networks. The model processes instance parameters through a preset fusion module to generate offset data of electronic devices. Since the instance parameters can reflect the process characteristics of electronic devices, by analyzing the instance parameters, the offset amounts of some parameters in the current electronic devices can be determined, realizing the simulation of the offset of data related to the electrical characteristics of electronic devices in a real circuit environment. Among them, the fusion module includes device physical information and neural networks. This setting, while retaining the high interpretability of physical equations, utilizes the powerful non-linear fitting ability of neural networks to capture other detailed characteristics of electronic devices in the simulation circuit, which is beneficial to improving the accuracy of circuit simulation. Furthermore, through the physical model, offset data, bias voltage configuration information of the electronic device, and instance parameters, the device characteristics of the electronic device are determined. The bias voltage configuration information simulates the working environment of the electronic device. Through the physical model, it can be ensured that the device characteristics of the finally output electronic device are obtained based on physical derivation, guaranteeing the accuracy of the captured device characteristics. In summary, the present application determines the device characteristics of electronic devices through a fusion model integrating physics and neural networks. On the one hand, it retains the accuracy and reliability of device physical information and physical models. On the other hand, it accelerates the training process of the fusion model through neural networks, thereby accurately and efficiently realizing circuit simulation and accurately capturing the device characteristics of electronic devices in a real circuit.

[0049] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication, and program running functions, such as a tablet computer, a personal computer, a mobile phone, a compiler, a circuit emulator, etc., or an electronic device capable of implementing the above functions.

[0050] Taking the terminal as the execution subject as an example, this embodiment and the following embodiments will be described. Based on this, the embodiments of the present application provide a method for constructing a fusion model integrating physics and neural networks, referring to Figure 1 , Figure 1 is a schematic flowchart of the first embodiment of the method for constructing a fusion model integrating physics and neural networks of the present application.

[0051] In this embodiment, the method for constructing a fusion model integrating physics and neural networks includes steps S10 to S30:

[0052] Step S10, construct a neural network according to the instance parameters of the electronic device and the first correction coefficient;

[0053] It should be noted that an electronic device is a basic component in a circuit and has specific electrical characteristics, such as capacitors, resistors, transistors, etc. The instance parameters of an electronic device refer to the specific numerical values used to describe the physical characteristics of the electronic device. Changes in the size, material, etc. of the electronic device will also have a direct impact on the device characteristics. Using electronic devices with different instance parameters will result in differences in the device characteristics of the electronic devices. Common instance parameters of electronic devices include: channel length, channel width, etc.

[0054] In addition, it should be noted that the correction factor refers to a set of numerical values obtained based on a neural network and used to adjust the electrical parameters of an electronic device. Due to the influence of the instance parameters (such as channel length, channel width, etc.) of the electronic device, the preset fixed parameters of the electronic device (such as threshold voltage, mobility, etc.) may shift. In order to calculate the degree of its shift, it can be determined through a trained neural network and physical model. The correction factor can be determined manually by the user or automatically by the system according to the model task.

[0055] Optionally, during the construction of the neural network, the number of input layer neurons is determined based on the dimension of the instance parameters of the electronic device, and each input neuron corresponds to a specific instance parameter feature (such as channel length, channel width, etc.); the number of output layer neurons is determined based on the number of correction factors to ensure that each output node can independently output the prediction value of the neural network for a specific correction factor, thereby completing the basic construction of the neural network.

[0056] Optionally, after determining the above basic architecture, the hidden layer structure of the neural network can be further designed. For example, according to the non-linear correlation characteristics between the input parameters and output parameters, the hierarchical depth of the hidden layer and the number of neurons in each layer are determined through cross-validation.

[0057] Optionally, after completing the construction of the neural network above, the weight matrix and bias vector of the neural network are randomly initialized, enabling the neural network to have the ability to perform forward propagation calculations, so that in the subsequent training process of the overall fusion model, the parameters of the neural network can be optimized through backpropagation.

[0058] It can be understood that through the non-linear modeling ability of the neural network, the complex relationship between the instance parameters and the correction factor is captured, thereby improving the accuracy and generalization ability of the subsequent constructed fusion model.

[0059] Step S20, construct a fusion module based on device physical information and a neural network;

[0060] It should be noted that device physical information refers to recognized theories, formulas, or empirical rules, etc., which can be used to predict or explain the electrical behavior of electronic devices.

[0061] Optionally, the prediction result of the neural network is combined with the device physical information to construct a fusion module.

[0062] Optionally, according to the fixed parameters by which the electronic device may deviate, the corresponding device physical information is determined, such as the mobility degradation equation, the short-channel effect equation, etc.; furthermore, according to the mapping relationship between the output parameters of the above neural network and the input parameters of the device physical information, a fusion module is constructed, and the fusion module can determine the offset data (fixed parameters of the offset) of the electronic device according to the correction coefficient and the corresponding physical equation.

[0063] It can be understood that by fusing physical information and the neural network, the fusion module simultaneously has the non-linear modeling ability of the neural network and the accuracy of physical information, thereby improving the subsequent ability to capture device characteristics.

[0064] Step S30, a fusion model is constructed according to at least one of the fusion modules and the physical model determined based on device characteristics.

[0065] It should be noted that the physical model refers to a mathematical model based on physical laws and theories, which is used to describe the behavior or device characteristics of an electronic device, such as describing the variation of characteristics such as gate capacitance and drain-source current with voltage. The device characteristics of an electronic device refer to the physical quantities exhibited by the electronic device under specific bias conditions, such as drain-source current and gate capacitance, which are used to evaluate the performance of the electronic device.

[0066] Exemplarily, in order to determine the variation of the drain-source current with the bias voltage, a mathematical model based on the long-channel equation is determined as the corresponding physical model, and a complete fusion model is constructed according to the mapping relationship between the output parameters of each fusion module and the input parameters of the physical model; after the construction is completed, a training data set including instance parameters, bias voltage, and corresponding device characteristics can be used to train the fusion model. Among them, the instance parameters and bias voltage in the training data set are marked as input data, while the device characteristics of the electronic device are marked as output data, that is, the target for the fusion model to predict and learn. Through such an input-output correspondence, the fusion model can learn the complex mapping relationship from instance parameters and bias voltage to the device characteristics of the electronic device, so as to accurately predict the device characteristics of unknown electronic devices (electronic devices outside the training data set), and has strong extrapolation ability.

[0067] Exemplarily, please refer to Figure 2 , Figure 2 A general framework diagram of a fusion model is provided. Specifically: W represents the channel width, L represents the channel length, W...L represents the instance parameters of the electronic device, and the bias voltage includes the gate voltage V G and the drain voltage V D , representing the gate voltage and the drain voltage respectively;Figure 2 The circles in it represent neurons in the neural network. The blue, orange, and green circles represent neurons in the input layer, hidden layer, and output layer of the neural network respectively. In addition, the neural network can also be composed of other models besides the fully connected neural network, and this embodiment does not make specific limitations in this regard. Figure 2 The fusion model in it includes a short-channel effect fusion module, a mobility degradation fusion module, and a long-channel core module (physical model), etc. For the specific number of fusion modules and the device physical information used therein, this embodiment does not make specific limitations; among them, the short-channel effect fusion module includes a neural network and a short-channel effect physical equation, which is mainly used to determine the threshold voltage shift ΔV of the electronic device th ; the mobility degradation fusion module includes a neural network and a mobility degradation physical equation, which is used to determine the mobility degradation factor D of the electronic device mob ; after determining the above offset data such as the threshold voltage shift and the mobility degradation factor, based on the long-channel core module, according to the determined offset data and the preset bias voltage configuration information, the device characteristics of the electronic device can be determined, such as the drain-source current I DS 、the drain-source resistance G DS 、the gate charge Q G 、the gate capacitance C GG etc. Figure 2 This kind of fusion model used to reflect the device characteristics in it not only retains the high efficiency and accuracy of the neural network, but also retains the strictness of the physical model, avoids the introduction of irrelevant parameters, and only requires a small amount of data sets or instance parameters to accurately capture the device characteristics of the electronic device.

[0068] This embodiment provides a method for constructing a fusion model integrating physics and neural network. By combining the physical model and the neural network through the fusion model, the collaborative modeling of physical law constraints and data-driven optimization is realized; through the powerful generalization ability of the neural network, the offset error of some data is reduced, and then through the device physical information and the physical model, it is ensured that the device characteristics predicted by the model strictly follow the basic equations of semiconductor devices, thus ensuring the simulation accuracy of the fusion model.

[0069] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to the above-mentioned embodiment one can be referred to the above introduction and will not be elaborated later. On this basis, after step S30, it further includes:

[0070] Step S01, construct the target loss function of the fusion model according to the electrical characteristics of the electronic device;

[0071] It should be noted that the electrical characteristics of electronic devices are used to describe the operating behavior of electronic devices under different working conditions, including output characteristics and transfer characteristics. Among them, the output characteristics mainly focus on the working state of electronic devices under different working conditions, such as the change of operating data of electronic devices under different drain-source voltages; the transfer characteristics mainly focus on the input-output relationship of electronic devices under different working conditions, such as the relationship between operating data and gate-source voltage change of electronic devices under a fixed drain-source voltage.

[0072] It should be noted that the electrical characteristics of electronic devices are the same as the device characteristics. Different expressions are used here to distinguish the training process of the fusion model and the actual application process of the fusion model. During the training process of the fusion model, based on the actual values of the electrical characteristics obtained, the corresponding model can be optimized based on backpropagation to determine the relevant parameters of the corresponding electrical characteristics; while in the actual application process of the fusion model, the finally determined device characteristics are the predicted values of the true electrical characteristics.

[0073] In addition, it should be noted that the target loss function is used to evaluate the difference between the model prediction value and the actual value, and usually appears in the training process of the model. The construction forms of the loss function include mean-square error (MSE), Mean Absolute Error (MAE), etc., which are not specifically limited in this embodiment.

[0074] Exemplarily, during the training process of the model, the fusion model is used to simulate the operating state of the electronic device in the simulation circuit and output relevant prediction values (device characteristics), such as drain-source current, linear transconductance, output conductance, etc.; then, the actual value measured by the electronic device in the real circuit is compared with the prediction value of the model, and a specific weight is assigned to the difference between the corresponding actual value and the prediction value to obtain the target loss function; furthermore, the target loss function can be used to train and optimize the fusion model until the target loss function is minimized.

[0075] It can be understood that by constructing a suitable target loss function, the performance of the fusion model can be evaluated and optimized more accurately, thereby improving the simulation accuracy of the model for electronic devices.

[0076] Step S02: Based on the backpropagation algorithm, adjust the weights and biases of the fusion model until the target loss function is minimized.

[0077] It should be noted that the backpropagation algorithm is a supervised learning algorithm. By calculating the gradients of the loss function with respect to the network weights and biases, and using optimization algorithms such as gradient descent to update the weights and biases, the loss function is minimized. The weights refer to the parameters connecting two neurons in a neural network, representing the strength of the influence of the previous layer of neurons on the next layer of neurons; the bias is a constant term used to adjust the activation threshold of neurons, and appropriate bias settings can enable the neural network to better fit the data.

[0078] Exemplarily, obtain the instance parameters and bias voltage configuration information of the electronic device, and determine the prediction result of the fusion model accordingly; then, based on the prediction result of the model and the actual operation data, calculate the value of the target loss function and determine the gradient of the target loss function; then, based on the gradient of the target loss function, backtrack and adjust the weights and biases layer by layer, and repeat the above steps until the target loss function converges.

[0079] It can be understood that by introducing the backpropagation algorithm, the training process of the model is the process of adjusting the parameters of the fusion model, significantly shortening the time and resources required for parameter adjustment of traditional physical models.

[0080] In this embodiment, by adjusting the weights and biases of the neural network to minimize the target loss function, the output characteristics or transfer characteristics predicted by the model can be made as close as possible to the actual measured values, thereby optimizing the circuit design and accurately capturing the device characteristics of the electronic device. At the same time, based on the automatic optimization method of backpropagation, the optimization of the model parameters is completed while training the fusion model, thereby improving the training efficiency and extrapolation ability of the fusion model, and at the same time improving the accuracy of the device characteristics determined by the fusion model.

[0081] In a feasible implementation manner, the fusion model includes a current model, and the electrical characteristics include: linear current, logarithmically processed current, logarithmically processed output conductance, and linear transconductance. Step S01 includes:

[0082] Step A01, construct the first loss function of the current model according to the linear current, logarithmically processed current, logarithmically processed output conductance, and linear transconductance.

[0083] It should be noted that the current model refers to a mathematical model used to predict the current behavior in an electronic device. Based on the current model, parameters such as the linear current, output conductance, and linear transconductance of the electronic device can be determined, and these parameters reflect the electrical characteristics of the electronic device under different operating conditions.

[0084] Additionally, it should be noted that the linear current is used to study the behavior of electronic devices in the on-state; the logarithmically processed current refers to the value obtained by taking the logarithm of the current. This is mainly because in the subthreshold region, the current increases exponentially with the increase of the gate voltage. Using the logarithmically processed current in the loss function can more accurately learn the electrical characteristics in the subthreshold region; the output conductance is a physical quantity that describes the degree of influence of the change in the output port voltage of the simulation circuit on the output current, and can be expressed as the ratio of the change in the output current to the change in the output voltage. Similarly, by taking the logarithm, the exponential change characteristic of the output conductance with the increase of voltage at low drain voltages can be learned; the linear transconductance is a physical quantity that describes the degree of influence of the change in the input port voltage of the simulation circuit on the output current, and can be expressed as the ratio of the change in the output current to the change in the input gate voltage, and is used to learn the derivative characteristic of the current with respect to the gate voltage.

[0085] Exemplarily, a loss function is constructed in the form of the mean absolute error. By summing the losses between the predicted values and the actual values of all samples according to the weight setting and then taking the average, this is used as the objective that needs to be minimized during the training process of the model. The definition formula of the first loss function is as follows:

[0086]

[0087] Among them, the parameter with a hat represents the simulated value of the device characteristics generated by the fusion model, and the parameter without a hat represents the measured value of the device characteristics of the electronic device in the real circuit; The symbol represents taking the Euclidean norm (L2 norm), that is, the square root of the sum of the squares of the vector elements, which is used to measure the difference between the simulated value and the measured value; I represents the linear current, and both the normal version and the logarithmic version are set. The logarithmic version is mainly to amplify the current in the subthreshold region; G DS represents the output conductance. Taking the logarithm is to amplify the output conductance in the saturation region. In the saturation region, the drain current no longer changes significantly with the drain-source voltage, so the output conductance is used to represent the current characteristics in the saturation region; G M represents the linear transconductance, which is used to enhance the learning effect of the transmission characteristics; α, β, γ, δ represent the weights of each parameter, which can be manually configured by the user according to actual needs or automatically determined through model training. This embodiment does not make specific limitations on this. N represents the number of samples, and Loss I represents the loss function (the first loss function) of the current model. This loss function sums the losses of all samples according to the weight setting and then takes the average, which is used as the objective that needs to be minimized during the training process of the fusion model, so that the current value (simulated value) and other operating data predicted by the model are as close as possible to the actual measured value, thereby accurately capturing the device characteristics of the electronic device.

[0088] In this embodiment, incorporating the linear current into the loss function of the current model enables the current model to learn the current variations in the linear region and saturation region of the electronic device under different instance parameters and / or bias voltage configuration information. Meanwhile, adding the logarithm-processed current allows the current model to learn the current variations in the subthreshold region under different instance parameters and / or bias voltage configuration information. Adding the logarithm-processed output conductance to the loss function is to enhance the learning ability of the derivative of the drain current with respect to the drain voltage change, and taking the linear transconductance is to enhance the learning effect of the transmission characteristics of the electronic device. The addition of these two conductances helps to improve the learning effect of the current derivative. By introducing the above loss function, the fusion model can more accurately capture the device characteristics in different regions of the electronic device under different operating conditions.

[0089] In a feasible embodiment, the fusion model includes a capacitance model, and the electrical characteristics include: gate capacitance, drain capacitance, and gate-source capacitance. Step S01 includes:

[0090] Step B01, constructing a second loss function of the capacitance model according to the gate capacitance, drain capacitance, and gate-source capacitance.

[0091] Exemplarily, the second loss function is also constructed in the form of the mean absolute error, and the specific formula is defined as follows:

[0092]

[0093] where the meanings of the parameters with ^, the meanings of the parameters without ^, and the meaning of the symbol are the same as those in the above first loss function formula, so they will not be elaborated here; C GG is the gate capacitance, that is, the total gate capacitance value, which is used to describe the storage and release ability of gate charge when the gate voltage changes; C DD is the drain capacitance, that is, the total drain capacitance value, which is used to describe the storage and release ability of drain charge when the drain voltage changes; C GS is the gate-source capacitance, which is used to describe the influence of the source voltage on the gate charge.

[0094] Since the gate capacitance can affect the input impedance and high-frequency response, the drain capacitance affects the output impedance and high-frequency response, and the gate-source capacitance will affect the input impedance and signal transmission, these parameters are very important for evaluating the electrical performance of the electronic device.

[0095] Exemplarily, in the process of determining the capacitance through the capacitance model, there is no need to rely on a large amount of data to describe the output characteristics and transmission characteristics of each terminal capacitance. Instead, through a physical model, only a small amount of key data is required, such as the gate capacitance C GG and the drain capacitance C DDThe training of the model can be completed, accelerating the training process of the model. Moreover, the terminal charge value determined based on the physical formula completely eliminates the offset phenomenon, making the simulation value as close as possible to the true value, thereby improving the accuracy of circuit simulation.

[0096] In this embodiment, by minimizing the second loss function of the capacitance model and optimizing the weights and biases of the fusion model, the capacitance parameter changes of the electronic devices in the simulation circuit can be made to fit as closely as possible to their capacitance changes in the real circuit, thereby accurately capturing the device characteristics of the electronic devices.

[0097] Based on the first embodiment and / or the second embodiment of the present application, in the third embodiment of the present application, the same or similar content as that in the above-mentioned first embodiment and second embodiment can be referred to the above introduction and will not be repeated hereinafter. On this basis, after step S30, the following further includes:

[0098] Step E10, generating offset data of the electronic device through the fusion module and instance parameters;

[0099] It should be noted that the offset data of the electronic device refers to the physical model characteristic parameters under different working conditions, such as threshold voltage offset, mobility degradation factor, etc. These data can be predicted and analyzed through physical models and / or neural networks.

[0100] Exemplarily, the relevant configuration of the instance parameters of the electronic device can be determined by responding to the user's electronic device configuration operation, supporting various electronic device configuration operations such as text box entry and parameter file import; or the key parameters such as bias voltage and aspect ratio defined in the SPICE netlist or simulation project file can be identified by integrating an EDA (Electronic Design Automation) interface to realize the automatic extraction of the device configuration information in the simulation circuit. In addition, intelligent fusion processing of multi-source input data can be performed by establishing a parameter verification matrix. When it is detected that there is a conflict between the instance parameters input by the user and the simulation circuit configuration, a preset parameter synchronization protocol will be automatically triggered and optimization suggestions will be generated.

[0101] In a feasible embodiment, based on the neural network and the obtained instance parameters, the corresponding correction coefficient is determined; the relevant coefficient in the device physical information is updated to the determined correction coefficient, and the corresponding offset data is generated based on the relevant coefficient of the electronic device. Specifically, which device physical information is included in the fusion module can be determined according to actual needs, and no specific limitation is made in this embodiment. For example, different offset data can be calculated according to the velocity saturation effect, quantum confinement effect, channel length modulation effect, etc.

[0102] Exemplarily, in the case where the offset data is the threshold voltage of an electronic device, through a neural network model, based on the instance parameters of the electronic device, such as the channel length L and the channel width W, the correlation coefficients of the subsequent physical model are determined, such as the change in threshold voltage, mobility-related parameters, substrate-related parameters, etc.; and then through device physical information (such as the short-channel effect equation), based on the above-mentioned corrected correlation coefficients, the threshold voltage offset of the electronic device is determined.

[0103] In another feasible embodiment, the physical model in the fusion model calculates based on the input data to generate preliminary offset data; and then the neural network in the fusion model fits and optimizes the calculation results in the physical model to generate the final offset data and output it for further analysis. For example, through the physical model, the theoretical threshold voltage offset under the current bias voltage condition is calculated; but in fact, the threshold voltage may have different degrees of offset due to the size of the electronic device, the thickness of the oxide layer, etc., and then the neural network learns the actual threshold voltage offsets of different historical electronic devices, and based on the instance parameters of the current electronic device, corrects the above-mentioned theoretical threshold voltage offset.

[0104] It can be understood that by combining the reliability of the physical model and the fitting ability of the neural network, the fusion module can more accurately predict and analyze the performance changes of electronic devices under different conditions, improving the accuracy and reliability of circuit simulation.

[0105] Step E20, determine the device characteristics of the electronic device through the physical model, offset data, bias voltage configuration information of the electronic device, and instance parameters.

[0106] It should be noted that the bias voltage refers to the external DC voltage applied to the electronic device in the simulation circuit, which is used to adjust the performance of the electronic device in the circuit, such as the gate voltage, source voltage, etc.

[0107] Exemplarily, after determining the physical model characteristic parameters (offset data) of the electronic device under the current instance parameters, the physical model in the fusion model can finally determine the device characteristics of the electronic device according to the determined physical model characteristic parameters and the obtained bias voltage configuration information and instance parameters.

[0108] Exemplarily, according to the bias voltage range, a suitable physical theory in the physical model can be automatically selected for calculation. For example, the current density in the electronic device under different bias voltages is calculated through physical formulas. When the bias voltage is low, the Drift-Diffusion model is used to calculate the electron current density, and when the bias voltage is high, the Impact Ionization model is used to calculate the current density.

[0109] Exemplarily, for the circuit simulation of a long-channel MOSFET (Metal-Oxide-Semiconductor Field-Effect Transistor), the threshold voltage offset and the mobility degradation factor (offset data) are obtained respectively through the fusion model, and the initial threshold voltage and mobility are corrected to obtain the corrected threshold voltage and mobility. Then, according to the long-channel equation, using the corrected threshold voltage and mobility, as well as the bias voltage configuration information and relevant parameters in the instance parameters, the magnitude of the drain-source current of the long-channel MOSFET (device characteristics) is determined. By observing the change of the drain-source current of the long-channel MOSFET under the working conditions of different bias voltages and / or instance parameters, the electrical characteristics such as the current driving ability of the long-channel MOSFET can be evaluated.

[0110] In this embodiment, the device characteristics of the electronic device are determined through the fusion model. While retaining the basic device physical knowledge, the non-ideal effects in the actual device are corrected through the neural network and the physical model, so as to meet the simulation requirements of nano-scale electronic devices.

[0111] In a feasible implementation manner, step E10 includes:

[0112] Step E11, determining a second correction coefficient of the electronic device through the neural network and the instance parameters;

[0113] Exemplarily, since some fixed parameters of the electronic device may change under different instance parameters of the electronic device, the neural network can be used to capture the change rules therein. Then, when facing unknown or known electronic devices, the correction coefficient of fixed parameters such as the threshold voltage or mobility can be determined according to the instance parameters of the current electronic device.

[0114] Step E12, generating offset data according to the device physical information and the second correction coefficient.

[0115] It should be noted that the offset data refers to a set of physical model characteristic parameters of the electronic device that are more in line with the real circuit environment after correction.

[0116] In a feasible implementation manner, the device physical information includes the short-channel effect equation, and the offset data includes the threshold voltage offset. Step E12 includes:

[0117] Step A20, determining the threshold voltage offset according to the short-channel effect equation and the second correction coefficient.

[0118] It should be noted that the short-channel effect equation refers to the mathematical expressions of various physical phenomena (such as threshold voltage shift, subthreshold slope change, etc.) caused by the reduction of channel length, DIBL (Drain Induced Barrier Lowering) effect, or substrate effect in short-channel devices.

[0119] Exemplarily, an appropriate short-channel effect equation can be selected according to the device structure and process conditions; then, through a neural network model, a correction coefficient is determined based on historical experimental data; the equation and the correction coefficient are input into a simulation software for calculation to obtain the threshold voltage shift. The threshold voltage shift method based on the short-channel effect equation can refer to the following formula:

[0120]

[0121] where DVT0 NN is the threshold voltage correction coefficient of the short-channel effect determined based on the neural network; ETA0 NN is the adjustment parameter of the DIBL effect obtained based on the neural network, DVTSUB NN is the related parameter of the substrate effect determined based on the neural network. Both the DIBL effect and the substrate effect can cause changes in the threshold voltage of electronic devices; ΔV th,all_NN is the threshold voltage shift determined based on the correction coefficient, where the correction coefficient is obtained based on the neural network; V bi , ψ st , V dsx respectively represent the built-in voltage, surface potential, and drain-source voltage of the electronic device under specific working conditions. In the short-channel effect, changes in these parameters will affect parameters such as the threshold voltage and carrier mobility; cosh is the hyperbolic cosine function, L eff represents the effective channel length, and λ is the characteristic length describing the attenuation characteristics of the potential distribution.

[0122] In this embodiment, by introducing the short-channel effect equation and the correction coefficient, the offset of the threshold voltage of the electronic device in actual operation can be more accurately simulated and predicted, thereby reducing the influence of data offset on the results of circuit simulation.

[0123] In a feasible embodiment, the device physical information includes the mobility degradation equation, the offset data includes the mobility degradation factor, and step E12 includes:

[0124] Step B20, determining the mobility degradation factor according to the mobility degradation equation and the second correction coefficient.

[0125] It should be noted that the mobility degradation equation refers to a mathematical model that describes the variation of carrier mobility with time and stress conditions in electronic devices. The Mobility Degradation Factor is a parameter used to describe the degree to which carrier mobility decreases in semiconductor devices due to various factors (such as electric field, temperature, lattice scattering, etc.). It is usually used to characterize the degradation of carrier mobility under high electric field or high temperature conditions compared to low electric field or low temperature conditions, and helps to predict and explain the behavior of electronic devices under different operating conditions.

[0126] Exemplarily, for a MOSFET, the calculation method of its mobility degradation factor based on the mobility degradation equation can refer to the following formula:

[0127]

[0128] where UA NN is the vertical electric field correction term for mobility degradation determined based on a neural network; EU NN is the comprehensive correction term for mobility degradation determined based on a neural network, including temperature or electric field scaling coefficients, slope adjustment parameters of exponential functions, threshold parameters, etc.; UD NN is the correction value for mobility degradation under thin film conditions determined based on a neural network; D mob represents the mobility degradation factor determined based on a correction coefficient, where the correction coefficient is obtained based on a neural network; E effa is the effective electric field in the vertical channel direction of the electronic device; q ia2 is the average inversion charge density in the channel, and these charges will affect the carrier mobility; Cox represents the capacitance of the silicon dioxide layer.

[0129] In this embodiment, by combining the use of the mobility degradation equation and the correction coefficient determined based on a neural network, the degradation law of mobility can be more accurately described, thereby improving the prediction accuracy of device performance.

[0130] Exemplarily, for the device physical information used in the physical model, in addition to the short-channel effect equation and the mobility degradation equation mentioned above, it can also include velocity saturation effect, quantum confinement effect, channel length modulation effect, etc. The specific device physical information adopted can be determined according to actual needs, and this embodiment does not make specific limitations on this.

[0131] In this embodiment, by introducing a correction coefficient, the prediction result of the physical model can be adjusted to more accurately reflect the actual situation or experimental data, thereby improving the accuracy of the device characteristics determined by the fusion model.

[0132] In a feasible embodiment, after step S30, it further includes:

[0133] Step S31: Optimize the hyperparameters of the fusion model based on the Bayesian optimization algorithm.

[0134] It should be noted that the Bayesian optimization algorithm is a global optimization algorithm based on a probability model. It gradually approaches the optimal value of the objective function by constructing a surrogate model and using an acquisition function to select the next evaluation point.

[0135] Exemplarily, during the optimization of the hyperparameter of the learning rate, first, it is necessary to determine the objective function and the value range of the hyperparameter. For example, the mean square error between the predicted drain-source current of the model and the actually measured value is used, and the learning rate is set to [0.001, 0.1]. The objective function can also adopt the above-mentioned constructed objective loss function. Then, a surrogate model and an acquisition function are selected to evaluate the hyperparameter. For example, the Gaussian equation is determined as the surrogate model, and the Expected Improvement (EI) is determined as the acquisition function, and the surrogate function value is calculated to update the surrogate model. Finally, the hyperparameter combination that minimizes the surrogate function value is found.

[0136] In this embodiment, through the Bayesian optimization algorithm, the hyperparameter combination that minimizes the objective function value can be found more precisely, thereby improving the performance of the model and the accuracy of the device characteristics determined by the fusion model.

[0137] After the model training is completed, the fusion model is tested using the test set, and the test results can be referred to Figure 3 and Figure 4 , Figure 3 and Figure 4 where the curves of different colors in

[0138] Figure 3 correspond to electronic devices of different sizes respectively. Among them, the green curve represents the variation curve of the electronic device with a channel length L of 16 nm, the blue curve represents the variation curve of the electronic device with a channel length L of 36 nm, and the red curve represents the variation curve of the electronic device with a channel length L of 96 nm. The unit of the vertical axis, a.u., represents arbitrary units, which are usually used to represent the relative value of a physical quantity rather than the absolute value, that is, used to represent the relative magnitude of current, voltage, etc.; the unit of the vertical axis is volts (V).

[0138] Figure 3 A voltage-current characteristic curve measured based on a current model is provided. Figure 3 In (a) of GS under the cases where the gate-source voltages V GS are 0.3 V, 0.5 V, and 0.8 V respectively, the simulation data and real circuit data of the drain-source current I DS vary with the drain-source voltage V DS , and the vertical axis is the drain-source current I DS, the horizontal axis is the drain-source voltage V DS ; Figure 3 (b) reflects the different drain-source voltage V DS When the voltages are 0.05V, 0.4V, and 0.8V respectively, the drain-source current I DS The running simulation data and real circuit data vary with the gate-source voltage V GS The left vertical axis is the logarithmic drain-source current I DS The vertical axis on the right is the drain-source current I DS , the horizontal axis is the gate-source voltage V GS ; Figure 3 (c) reflects the different gate-source voltage V GS When the voltages are 0.3V, 0.5V and 0.8V respectively, the output conductance G DS , the running simulation data and the real circuit data change with the drain-source voltage V DS The vertical axis is the logarithmic processed output conductance G DS , the horizontal axis is the drain-source voltage V DS ; Figure 3 (d) reflects the different drain-source voltage V DS When the voltages are 0.05V, 0.4V, and 0.8V respectively, the linear transconductance G M The running simulation data and real circuit data vary with the gate-source voltage V GS The vertical axis is the linear transconductance G M , the horizontal axis is the gate-source voltage V GS ; Figure 3 The curves of the simulation data and the real circuit data in (a), (b), (c), and (d) are basically consistent, which shows that the current model in the fusion model has a high accuracy.

[0139] Figure 4 Provides a capacitance characteristic curve graph based on the capacitance model. Figure 4 (a) reflects the different gate-source voltage V GS When the voltages are 0.3V, 0.5V, and 0.8V respectively, the gate capacitance C GG The running simulation data and real circuit data vary with the drain-source voltage V DS The vertical axis of the curve is the gate capacitance C GG , the horizontal axis is the drain-source voltage V DS ; Figure 4 (b) reflects the different drain-source voltage V DS When the voltages are 0.05V, 0.4V, and 0.8V respectively, the gate capacitance C GG The running simulation data and real circuit data vary with the gate-source voltage V GS The vertical axis of the curve is the gate capacitance CGG , the horizontal axis is the gate-source voltage V GS ; Figure 4 In (c), it reflects the drain-source capacitance C GS under the conditions where the gate-source voltages V are 0.3V, 0.5V, and 0.8V respectively, and the curves of the operating simulation data and the real circuit data of the drain-source capacitance C DD changing with the drain-source voltage V DS , the vertical axis is the drain capacitance C DD , and the horizontal axis is the drain-source voltage V DS ; Figure 4 In (d), it reflects the drain capacitance C DS under the conditions where the drain-source voltages V are 0.05V, 0.4V, and 0.8V respectively, and the curves of the operating simulation data and the real circuit data of the drain capacitance C DD changing with the gate-source voltage V GS , the vertical axis is the drain capacitance C DD , and the horizontal axis is the gate-source voltage V GS ; Figure 4 In (a), (b), (c), and (d), the curves of the operating simulation data and the real circuit data basically coincide, indicating that the accuracy of the capacitance model in the fusion model is relatively high.

[0140] In addition, this application also conducts error statistics for the drain-source current I DS , gate capacitance C GG and drain capacitance C DD of the nanoscale electronic device FinFET (Fin Field-Effect Transistor) under different channel lengths L. The error calculation methods include the NAE (Normalized Absolute Error) calculation method and the MAPE (Mean Absolute Percentage Error) calculation method. For the specific error results, please refer to the following table:

[0141]

[0142] Among them, the ones without parentheses are the NAE errors, and the ones in parentheses are the MAPE errors. Based on the above error data, it is not difficult to see that the method for constructing the fusion model integrating physics and neural networks in this application has relatively high accuracy for circuit simulations of nanoscale electronic devices of different sizes.

[0143] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the method for constructing the fusion model integrating physics and neural networks in this application. Based on this technical concept, more forms of simple transformations are within the protection scope of this application.

[0144] The embodiment of the present application further provides a fusion model construction device integrating physics and neural networks. Please refer to Figure 5 , the fusion model construction device integrating physics and neural networks includes:

[0145] A neural network construction module 10, configured to construct a neural network according to the instance parameters of the electronic device and the first correction coefficient;

[0146] A fusion module construction module 20, configured to construct a fusion module based on device physical information and the neural network;

[0147] A fusion model construction module 30, configured to construct a fusion model by combining at least one of the fusion modules and a physical model determined based on device characteristics.

[0148] The fusion model construction device integrating physics and neural networks provided by the embodiment of the present application adopts the fusion model construction method integrating physics and neural networks in the above embodiment, and can solve the technical problem of how to accurately capture the device characteristics of electronic devices. Compared with the prior art, the beneficial effects of the fusion model construction device integrating physics and neural networks provided by the present application are the same as those of the fusion model construction method integrating physics and neural networks provided by the above embodiment, and other technical features in the fusion model construction device integrating physics and neural networks are the same as those disclosed in the above embodiment method, and will not be elaborated herein.

[0149] The embodiment of the present application provides an electronic device, which 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 so that the at least one processor can execute the fusion model construction method integrating physics and neural networks in the first embodiment above.

[0150] Next, refer to Figure 6 , which shows a schematic structural diagram of an electronic device suitable for implementing the embodiment of the present application. The electronic device in the embodiment of the present application may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (such as in-vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 6 The electronic device shown is only an example and should not impose any limitation on the functions and usage scope of the embodiment of the present application.

[0151] As shown Figure 6 in the figure, the electronic device may include a processing device 1001 (such as a central processing unit, a graphics processing unit, etc.), which may perform various appropriate actions and processes according to the program stored in the read-only memory 1002 or the program loaded from the storage device 1003 into the random access memory 1004. In the random access memory 1004, various programs and data required for the operation of the electronic device are also stored. The processing device 1001, the read-only memory 1002, and the random access memory 1004 are connected to each other through a bus 1005. The input / output interface 1006 is also connected to the bus. Generally, the following systems may be connected to the input / output interface 1006: an input device 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 may allow the electronic device to communicate with other devices wirelessly or wiredly to exchange data. Although the figure shows an electronic device having various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems may be implemented or had alternatively.

[0152] Specifically, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart may be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes program codes for executing the method shown in the flowchart. In such an embodiment, the computer program may be downloaded and installed from the network through the communication device, or installed from the storage device 1003, or installed from the read-only memory 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiments disclosed in the present application are executed.

[0153] The electronic device provided by the embodiments of the present application adopts the method for constructing an integrated physical and neural network fusion model in the above-mentioned embodiments, and can solve the technical problem of how to accurately capture the device characteristics of electronic devices. Compared with the prior art, the beneficial effects of the electronic device provided by the present application are the same as those of the method for constructing an integrated physical and neural network fusion model provided by the above-mentioned embodiments, and other technical features in the electronic device are the same as those disclosed in the method of the previous embodiment, and will not be elaborated here.

[0154] It should be understood that each part disclosed in this application can be implemented by hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in a suitable manner in any one or more embodiments or examples.

[0155] As described above, the above are only specific embodiments of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed in this application can easily think of changes or substitutions, which should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

[0156] An embodiment of this application provides a computer-readable storage medium having computer-readable program instructions (i.e., computer programs) stored thereon, and the computer-readable program instructions are used to execute the method for constructing a fusion model integrating physics and neural networks in the above embodiments.

[0157] The computer-readable storage medium provided by the embodiment of this application can be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to: electrical connections with 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 fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above. In this embodiment, the computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or combined with an instruction execution system or device. The program code contained on the computer-readable storage medium can be transmitted by any appropriate medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0158] The above computer-readable storage medium can be included in an electronic device; or it can exist alone without being assembled into the electronic device.

[0159] The above computer-readable storage medium stores one or more programs, which, when executed by an electronic device, cause the electronic device to: construct a neural network according to instance parameters of electronic devices and a first correction coefficient; construct a fusion module based on device physical information and the neural network; and construct a fusion model by combining at least one fusion module and a physical model constructed based on device characteristics.

[0160] Computer program code for performing the operations of this application can be written in one or more programming languages or combinations thereof. The programming languages include object-oriented programming languages such as Java, Smalltalk, and C++, as well as 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, executed as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (for example, by using an Internet service provider to connect through the Internet).

[0161] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in the flowchart or block diagram can represent a module, a program segment, or a part of code that contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks can occur in a different order than that marked in the accompanying drawings. For example, two consecutive blocks shown can actually be executed substantially in parallel, and they can sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, as well as combinations of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system for performing the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0162] The modules described in the embodiments of this application can be implemented in software or in hardware. In some cases, the name of the module does not constitute a limitation on the unit itself.

[0163] The readable storage medium provided by the embodiment of the present application is a computer-readable storage medium. The computer-readable storage medium stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned method for constructing a fusion model integrating physics and neural networks, and can solve the technical problem of how to accurately capture the device characteristics of electronic devices. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present application are the same as those of the method for constructing a fusion model integrating physics and neural networks provided by the above embodiment, and will not be elaborated herein.

[0164] The embodiment of the present application further provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned method for constructing a fusion model integrating physics and neural networks are implemented.

[0165] The computer program product provided by the embodiment of the present application can solve the technical problem of how to accurately capture the device characteristics of electronic devices. Compared with the prior art, the beneficial effects of the computer program product provided by the present application are the same as those of the method for constructing a fusion model integrating physics and neural networks provided by the above embodiment, and will not be elaborated herein.

[0166] The above are only partial embodiments of the present application, and thus do not limit the patent scope of the present application. Any equivalent structural transformation made under the technical concept of the present application by using the content of the specification and drawings of the present application, or any direct / indirect application in other related technical fields, is included in the patent protection scope of the present application.

Claims

1. A method for constructing a fusion model integrating physics and neural networks, characterized in that The method for integrating physics and neural network modeling includes: Construct a neural network according to the instance parameters of the electronic device and the first correction coefficient; Construct a fusion module based on the device physical information and the neural network; Construct a fusion model according to at least one of the fusion modules and a physical model determined based on device characteristics.

2. The method for constructing a fusion model integrating physics and neural networks according to claim 1, characterized in that, After the step of constructing the fusion model, it further includes: Construct an objective loss function of the fusion model according to the electrical characteristics of the electronic device; Based on the backpropagation algorithm, adjust the weights and biases of the fusion model until the objective loss function is minimized.

3. The method for constructing a fusion model integrating physics and neural networks according to claim 2, characterized in that, The fusion model includes a current model, and the electrical characteristics include: linear current, logarithmically processed current, logarithmically processed output conductance, and linear transconductance. The step of constructing the objective loss function of the fusion model according to the electrical characteristics of the electronic device includes: Construct a first loss function of the current model according to the linear current, the logarithmically processed current, the logarithmically processed output conductance, and the linear transconductance.

4. The method for constructing a fusion model integrating physics and neural networks according to claim 1, wherein, After the step of constructing the fusion model, the method further includes: Generate offset data of the electronic device through the fusion module and the instance parameters; Determine the device characteristics of the electronic device through the physical model, the offset data, the bias voltage configuration information of the electronic device, and the instance parameters.

5. The method for constructing an integrated physical and neural network fusion model according to claim 4, wherein The step of generating offset data of the electronic device through the fusion module and the instance parameters includes: Determine a second correction coefficient of the electronic device through the neural network and the instance parameters; Generate the offset data according to the device physical information and the second correction coefficient.

6. The method for constructing an integrated physical and neural network fusion model according to claim 5, wherein The device physical information includes a short-channel effect equation, and the offset data includes a threshold voltage offset. The step of generating the offset data according to the device physical information and the second correction coefficient includes: Determine the threshold voltage offset according to the short-channel effect equation and the second correction coefficient.

7. The method for constructing a fusion model integrating physics and neural networks according to claim 5, characterized in that, The device physical information includes a mobility degradation equation, and the offset data includes a mobility degradation factor. The step of generating the offset data according to the device physical information and the second correction coefficient includes: Determine the mobility degradation factor according to the mobility degradation equation and the second correction coefficient.

8. The method for constructing a fusion model integrating physics and neural networks according to claim 1, wherein After the step of constructing the fusion model, it further includes: Optimize the hyperparameters of the fusion model based on the Bayesian optimization algorithm.

9. An electronic device, characterized in that, The device includes: a memory, a processor, and a computer program stored on the memory and executable on the processor. The computer program is configured to implement the steps of the method for constructing a fusion model by integrating physics and neural network as described in any one of claims 1 to 8.

10. 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 method for constructing a fusion model by integrating physics and neural network as described in any one of claims 1 to 8.

Images