Transistor high-power electromagnetic thermal model optimization system based on deep learning

Abstract

The invention discloses a transistor high-power electromagnetic heat model optimization system based on deep learning, and relates to the technical field of electromagnetic heat. Comprising an input module, a deep learning module, a global optimization module, an online adaptive module and a real-time simulation module. The input module obtains various key parameters and constructs a high-dimensional input space; the deep learning module constructs a deep learning model by embedding a physical constraint and uncertainty quantization mechanism; the global optimization module is used for carrying out global search on design parameters, rapidly evaluating a target function by utilizing an agent model, and searching a Pareto optimal solution in a design parameter space; the online adaptive module is used for drift detection and updating; the real-time simulation module simulates an electromagnetic field and a thermal field in real time, and data exchange ensures timely synchronization of boundary data through dynamic scheduling and load balancing; all the modules cooperate with one another through data and feedback, and a complete, closed-loop and continuously self-adaptive updating optimization system is achieved.

Description

Technical Field

[0001] The present invention relates to the field of electromagnetic thermal technology, and particularly to an optimization system for a high-power electromagnetic thermal model of a transistor based on deep learning. Background Art

[0002] High-power transistors are key components in modern wireless communication, radar, satellite communication and other fields. With the continuous increase of working power and frequency, the heat generated inside the device and the resulting electromagnetic and thermal effects become increasingly significant; the high-power electromagnetic thermal model of a transistor is a multi-physical field model that couples phenomena such as electromagnetic fields, heat conduction and possibly mechanical stress, etc., and comprehensively numerically simulates and predicts the electromagnetic response and thermal effects generated when a high-power transistor (such as a high electron mobility transistor HEMT, LDMOS, etc.) is working; therefore, the high-power electromagnetic thermal model of a transistor is very important;

[0003] Traditional models usually rely on explicit physical parameters and material properties, but in actual production, these parameters may be uncertain or change with the working state. Models based solely on prior physical formulas are sensitive to initial parameters, boundary conditions and local non-linear effects, resulting in an increase in prediction errors under certain extreme or complex working conditions; at the same time, in the process of global design optimization, conventional methods are prone to falling into local optimal solutions and it is difficult to comprehensively search the entire parameter space to obtain the optimal design. Summary of the Invention

[0004] Based on this, it is necessary to provide an optimization system for a high-power electromagnetic thermal model of a transistor based on deep learning to solve the problems mentioned in the above background art.

[0005] The object of the present invention can be achieved by the following technical solutions: An optimization system for a high-power electromagnetic thermal model of a transistor based on deep learning, comprising: an input module, a deep learning module, a global optimization module, an online adaptive module and a real-time simulation module; the input module obtains various key parameters from transistor process documents and online sensors, and constructs a high-dimensional input space; the deep learning module constructs a deep learning model by embedding physical constraints and an uncertainty quantification mechanism; the global optimization module globally searches for design parameters, quickly evaluates the objective function using a surrogate model, and searches for Pareto optimal solutions in the design parameter space; the online adaptive module is used for drift detection and updating;

[0006] The real-time simulation module performs real-time simulation on the electromagnetic field and the thermal field to quickly verify the actual effects of design parameters; divides the three-dimensional space into several sub-regions, there are several computing units, and the computing units are responsible for updating the field values. There are overlapping regions or boundary regions at the grid boundaries of each sub-region; for the update of the electromagnetic field values in adjacent regions, boundary data exchange is used to ensure data consistency. The boundary data exchange mechanism is specifically as follows:

[0007] Step 1, real-time monitoring and priority calculation: In each computing unit, for boundary data exchange, real-time monitor and obtain the FIFO occupancy rate, data waiting time, and link penalty value; normalize the FIFO occupancy rate, data waiting time, and penalty link value and take their numerical values for weighted calculation to obtain the priority Pry of data transmission; the calculation process of the link penalty value is as follows: Calculate through the round-trip time (RTT), send a probe packet every 1ms, and calculate the delay: Based on the number of CRC check failures, calculate the error rate: Penalty link value:

[0008]

[0009] Step 2, FIFO data exchange scheduling: When Pry≥0.8 exists, preempt immediately; when all Pry<0.8, the scheduler periodically polls all computing units and allocates transmission time slices according to the calculated dynamic priority; if congestion occurs during data transmission, forward the data through the intermediate computing unit. If the forwarding fails continuously for 3 times, gradually increase the number of intermediate computing units, and distribute the data in this link direction to several intermediate computing units for forwarding;

[0010] Step 3, data exchange execution: Each computing unit packs the boundary data to be exchanged into data blocks of a fixed size and writes them into the FIFO buffer; adjacent computing units realize data exchange through the FIFO buffer. The sending end writes the data into the FIFO, and the receiving end reads the data from the FIFO; adopt asynchronous communication and double-buffer technology to make data transmission and local computing proceed in parallel.

[0011] In some embodiments, the real-time simulation module is also used to detect anomalies and trigger safety control in a timely manner. Its execution process is as follows:

[0012] Introduce the Isolation Forest algorithm as an anomaly detection method. For each key parameter collected in real time, use the Isolation Forest algorithm to calculate its path length δ. The path length represents the number of nodes that need to be passed from the root node to the key parameter in the Isolation Forest; set a path threshold. If the path length of the key parameter is less than the path threshold, then regard this key parameter as an abnormal parameter and trigger the safety mode; the specific safety mode is: reduce the device operating power to 80, and start the emergency heat dissipation device; regard the abnormal parameter and the triggered safety mode as an abnormal event.

[0013] In some embodiments, the execution process of the input module for constructing a high-dimensional input space is as follows:

[0014] The key parameters include electrical parameters, thermal parameters, material parameters, and dynamic sensing parameters; among them, the electrical parameters include operating voltage, current density, switching frequency, and carrier mobility; the thermal parameters include thermal conductivity, specific heat capacity, and boundary convection coefficient; the material parameters include dielectric constant, magnetic permeability, and thermal expansion coefficient; the dynamic sensing parameters include the real-time temperature field distribution T(x, y, z, t), the electromagnetic field distributions E(x, y, z, t) and B(x, y, z, t), the current waveform I(t), the voltage waveform V(t), and the stress field distribution Y(x, y, z, t);

[0015] Preprocessing: For the collected electromagnetic signals, use wavelet threshold denoising. Use the Daubechies 9 / 7 wavelet basis. After wavelet transform decomposition, apply soft threshold processing: τ is a preset threshold, and W(E) represents the discrete wavelet coefficients of the electromagnetic signal; establish a Kalman filter, and the state space model is: where k represents the time, and x k represents the state vector. In temperature drift correction, x k is set as a combination of temperature and temperature change rate: T(k) represents the temperature value at time k, and represents the change rate of temperature; A is the state transition matrix, Δt is the sampling time step, and w k-1 is the process noise, assuming it follows a Gaussian distribution with a mean of 0 and a covariance of Q. The matrix Q quantitatively describes the statistical attributes of the process noise w k-1 , that is, the variance of the noise and the covariance between different state variables; z k is the observation parameter, H is the observation matrix, assuming it follows a Gaussian distribution with a mean of 0 and a covariance of R; R is the observation noise covariance matrix; establish a Kalman filter and perform optimal estimation to provide accurate state correction information for the subsequent construction of the deep learning model; perform Box-Cox transformation on all key parameters to construct dimensionless numbers;

[0016] A graph neural network is constructed, where each node represents a physical parameter and the weight is determined based on the mutual information entropy between different parameters. Spatial features are extracted using a 3D convolutional network, and temporal features are extracted using an LSTM network. Ultimately, the two features are fused into a high-dimensional input.

[0017] In some embodiments, the deep learning module performs the following steps:

[0018] Use the encoder to extract multi-scale features, and extract multi-scale features of the input data through residual blocks and dense connections. Use cross-layer connections to fuse the encoder features with the decoder upsampled features. Finally, an activation function needs to be applied to the output to ensure that some predicted physical quantities meet physical constraints.

[0019] The model output is regularized using a physical information network. According to Maxwell's equations, the constraints of the electromagnetic part should ideally satisfy: is the curl operation, which is used to measure the degree of rotation of a vector field near a certain point. G represents the electric field displacement. It represents the partial derivative of the electric displacement field with respect to time, reflecting the change of the electric field with time. According to Fourier's law and thermal equilibrium, the heat conduction constraints satisfy: in is the temperature gradient, Q Joule is the Joule heat loss; the physical constraint loss is L phys :

[0020] Where λ1 and λ2 are physical regularization weights;

[0021] The Monte Carlo method and Baye network neural method are used to quantify the uncertainty of the output. The Monte Carlo method is to add Monte Carlo before the key layer, and perform M = 50 forward propagations during inference to obtain the predicted mean. and variance σ 2 , the expression is: Output confidence interval Quantify model uncertainty; where represents the prediction result of the neural network using the current parameters for the input parameter X after the mth forward propagation. θ represents all learnable parameters in the deep learning model, including the convolution kernel weights of the convolution layer, the weights and biases of the fully connected layer, etc. M is the total number of forward propagation samples. is the mean of the predicted results, which is the point estimate output of the final model, and the variance σ 2 is the variance of the prediction results; variational inference is used to approximate the posterior of some parameters to a Gaussian distribution, and training is performed by maximizing ELBO; data loss, physical loss and uncertainty loss are combined and analyzed to construct the total loss.

[0022] In some embodiments, the global optimization module uses the non-dominated sorting genetic algorithm and Bayesian optimization to perform a global search for design parameters, utilizes a surrogate model to quickly evaluate the objective function, and searches for Pareto optimal solutions in the design parameter space. Specifically:

[0023] Construct the objective function, and the additional restrictions imposed on the design parameters are the constraint conditions, which are: the breakdown voltage of the device ≥ 1.2 times the rated operating voltage, and the switching efficiency reaches at least 95%; starting from the generation of the initial population, use LHS to generate design parameters for continuous variables, and then randomly select discrete variables; perform electromagnetic-thermal simulations on each individual, calculate the objective function value, stratify the individuals through non-dominated sorting, and then use the Das-Dennis method to generate reference points to ensure population diversity; subsequently, generate offspring through SBX crossover and polynomial mutation, and finally retain excellent individuals according to the reference point association to form the next generation; use the Gaussian process surrogate model to model the objective function, then perform local search on continuous variables through the expected improvement acquisition function, combine the enumeration method of discrete variables, and at the same time use the Ray framework to achieve parallel acceleration to quickly evaluate a large number of design parameters, thereby accelerating the global search; through the collaborative work of the non-dominated sorting genetic algorithm and Bayesian optimization, perform local fine search and update the population every certain number of generations until the preset termination condition is reached.

[0024] In some embodiments, the execution process of the online adaptive module is as follows:

[0025] Extract key parameters in real time, and use the KS test to compare the real-time data distribution P real (X) with the training data P train (X) to obtain the drift value KS: Preset the drift threshold. If the drift value KS > the drift threshold, then trigger model update; the specific model update is: use variational inference to update the model parameters, assuming the parameter posterior The optimization objective is to maximize the ELBO, and the expression is: Set the learning rate, update and iterate several times, and adjust the model to adapt to the new data.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] 1. By utilizing multi-source data and preprocessing methods (noise reduction, filtering, dimensionless conversion), the data deviation problems caused by parameter uncertainty, sensor noise, and drift are effectively solved; multi-modal fusion ensures that not only the characteristics of a single physical quantity are captured, but also electricity, heat, materials, and dynamic working conditions are comprehensively considered, providing reliable input for subsequent model training; realizing provides high-quality and robust input data for the entire system, reducing the limitations brought by the dependence of traditional models on explicit physical parameters and ensuring that subsequent models work on a more accurate data basis;

[0028] 2. Through multi-scale feature extraction and cross-layer connection, fine-grained and global information in the input data are effectively captured, enhancing the model's fitting ability for complex physical phenomena; embedding Maxwell's equations and heat conduction equations as physical constraints and using physical information to regularize the training process enable the model's prediction results to be not only accurate in data fitting but also satisfy basic physical laws, improving the extrapolation ability and robustness; the Monte Carlo Dropout method is adopted to quantify the uncertainty of model predictions and give confidence intervals; the Bayesian neural network method is used to obtain the posterior distribution of parameters through variational inference, improving the prediction reliability of the model under complex and extreme working conditions; the network structure utilizes multi-scale features and physical constraints, making the model perform more robustly in scenarios with local non-linearity and boundary condition changes; uncertainty quantification provides prediction confidence, helping engineers comprehensively consider in risk assessment and alleviating the uncertainty problems brought by the sensitivity of traditional models to prior parameters; realizing through data-driven and physical information embedding enables the model to accurately predict the electromagnetic and thermal field states under complex working conditions, effectively overcoming the problem of increased prediction errors of traditional prior formula-based methods under extreme conditions;

[0029] 3. By constructing the objective function and setting constraints, initializing the population and non-dominated sorting, it is ensured that the optimization scheme has reliable safety while achieving high performance; according to the collaborative work of NSGA-Ⅲ and Bayesian optimization, local fine search and population update are carried out every certain number of generations until the preset termination conditions (such as the number of iterations or convergence rate) are reached, effectively searching the entire design parameter space and obtaining balanced solutions from multiple objective perspectives; combining the local fine search of Bayesian optimization ensures that it will not fall into local optima and obtains the Pareto optimal solution set; solves the problem that traditional methods are prone to falling into local optima and difficult to comprehensively search the parameter space when performing global design optimization. Using surrogate model acceleration and global co-evolution methods, multi-objective optimization is realized, ensuring that the design parameters reach the best balance under various constraint conditions;

[0030] 4. By using the Kolmogorov–Smirnov (KS) test to compare the distributions of real-time collected data and training data, data drift is determined. After detecting data drift, the Bayes-by-Backprop method is adopted to update the model parameters according to the ELBO optimization formula. During the update process, the underlying feature extraction network can be frozen and only the top layer is fine-tuned to improve the update speed and avoid catastrophic forgetting. It can monitor and detect the drift of data distribution in the production environment in real time, and adjust the model parameters through online update, enabling the model to continuously adapt to actual production changes, effectively alleviating the prediction error problem caused by parameter uncertainty in traditional models.

[0031] 5. Through FPGA hardware acceleration and parallel computing, real-time simulation of complex electromagnetic-thermal fields is achieved. Data exchange ensures the timely synchronization of boundary data through dynamic scheduling and load balancing, providing reliable real-time feedback for the entire closed-loop system, thereby greatly reducing the overall calculation time and improving the system security and accuracy. Each module collaborates with each other through data and feedback to achieve a complete, closed-loop and continuously self-adaptive updated optimization system, providing high-precision, robust and highly automated technical support for the design of high-power transistor electromagnetic-thermal models. This solution can not only significantly reduce the prediction error and calculation time-consuming, but also improve the passing rate and long-term reliability of the system under extreme working conditions.

[0032] In summary, through the physics-informed neural network, multi-objective global optimization and online adaptive mechanism, the present invention constructs a full-chain closed-loop system covering "data - model - optimization - control", effectively solving the problem of multi-physical field coupling of high-power transistors and providing a complete technical path for the design of highly reliable power electronic devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application or in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0034] Figure 1 It is a schematic diagram of the connection of the system modules of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following describes in detail the specific embodiments of the present invention with reference to the accompanying drawings. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the spirit of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0036] As Figure 1 shown, the optimization system for the high-power electromagnetic thermal model of a transistor based on deep learning includes: an input module, a deep learning module, a global optimization module, an online adaptive module, and a real-time simulation module;

[0037] The input module obtains various key parameters from transistor process documents, online sensors, etc., preprocesses and performs multimodal fusion on them, and constructs a high-dimensional input space; specifically:

[0038] The key parameters include electrical parameters, thermal parameters, material parameters, and dynamic sensing parameters; among them, the electrical parameters include the operating voltage V, current density J, switching frequency f sw , carrier mobility μ; the thermal parameters include thermal conductivity κ0, specific heat capacity C p and boundary convection coefficient h0; the material parameters include dielectric constant ε r , magnetic permeability μ m , coefficient of thermal expansion α T ; the dynamic sensing parameters include the real-time temperature field distribution T(x, y, z, t), electromagnetic field distributions E(x, y, z, t) and B(x, y, z, t), current waveform I(t), voltage waveform V(t), and stress field distribution Y(x, y, z, t). For the current-voltage waveform, the time-domain and frequency-domain features are extracted simultaneously, and the spectrogram is obtained using the fast Fourier transform. For the stress field distribution, it is obtained in real time through a microstructure sensor (such as a MEMS sensor);

[0039] Preprocessing: For the collected electromagnetic signals, wavelet threshold denoising is used. The Daubechies 9 / 7 wavelet basis is used. After wavelet transform decomposition, soft threshold processing is applied: τ is a preset threshold, and W(E) represents the discrete wavelet coefficients of the electromagnetic signal; a Kalman filter is established, and the state space model is: where k represents the time, and x k represents the state vector. In temperature drift correction, x k is set as a combination of temperature and temperature change rate: T(k) represents the temperature value at time k, and represents the rate of change of temperature; A is the state transition matrix, which represents the deterministic evolution rule of the system state from time k-1 to time k. When there is no external interference, x k can be obtained from A and x k-1 ; In temperature drift correction, A can include the heat conduction, inertia of the physical system, and the dynamic relationship of temperature change; for example, if a simple first-order system is adopted, then where Δt is the sampling time step, representing the update of temperature due to the change rate at the previous moment; w k-1 is the process noise, which represents the process noise caused by incomplete modeling, internal random disturbances or parameter fluctuations of the system. It is assumed to follow a Gaussian distribution with a mean of 0 and a covariance of Q; it reflects the influence of external or internal uncertainty factors (such as environmental temperature fluctuations, randomness of material properties, etc.) on temperature and the rate of temperature change during the state transition process; the matrix Q quantitatively describes the statistical properties of the process noise w k-1 , that is, the variance of the noise and the covariance between different state variables; z k is the observation parameter, representing the observation parameter obtained by the sensor at time k. In this scheme, it can be the actually measured temperature, temperature field parameter or other related physical quantities; H is the observation matrix, which represents mapping the system state x k to the observation parameter z k ; v k is the observation noise, which represents the noise introduced during the measurement process. It is assumed to follow a Gaussian distribution with a mean of 0 and a covariance of R; R is the observation noise covariance matrix, which represents the reliability and measurement accuracy of the sensor. A smaller value indicates a lower noise level of the observed data; through this state space model, we can use the Kalman filtering technology to optimally estimate the system state (temperature and its change rate) in the presence of uncertainties (process noise and observation noise), and provide accurate state correction information for the subsequent construction of the deep learning model;

[0040] Apply the Box-Cox transformation to all key parameters to construct dimensionless numbers;

[0041] Construct a graph neural network (GNN), where each node represents a certain physical parameter (such as V, J, f sw etc.), and the weights are determined according to the mutual information entropy between different parameters; the spatial features are extracted using a 3D convolutional network: F spatial = Conv3D(X in , W 3D );

[0042] The time features are extracted using an LSTM network: h t = LSTM(F spatial (t), h t-1 ); Finally, the two parts of the features are fused into a high-dimensional input: Conv() represents the convolution operation; represents feature concatenation to enhance feature reuse;

[0043] By utilizing multi-source data and preprocessing methods (noise reduction, filtering, dimensionless transformation), the data deviation problems caused by parameter uncertainty, sensor noise, and drift are effectively solved; multi-modal fusion ensures that not only the characteristics of a single physical quantity are captured, but also electricity, heat, materials, and dynamic operating conditions are comprehensively considered, providing reliable input for subsequent model training; it realizes providing high-quality and robust input data for the entire system, reducing the limitations brought by the dependence of traditional models on explicit physical parameters, and ensuring that subsequent models work on a more accurate data basis.

[0044] The deep learning module uses an improved U-Net++ and a parallel PINN network to embed physical constraints and uncertainty quantification mechanisms, efficiently learning the coupling relationships between multiple physical fields, constructing a deep learning model that can capture the complex coupling relationships of multiple physical fields and satisfy physical constraints, and at the same time quantifying prediction uncertainties; specifically:

[0045] The encoder uses ResNet-50 to extract multi-scale features, and extracts multi-scale features of the input data through multiple layers of residual blocks and dense connections; uses cross-layer connections to fuse encoder features with decoder upsampled features to form a structure similar to U-Net++, and the expression is: Finally, an activation function (such as Softplus) needs to be applied to the output to ensure that some predicted physical quantities (such as thermal conductivity κ0) satisfy physical constraints (i.e., κ0 > 0); where l represents any layer; represents the feature map extracted by the encoder at the l-th level. Feature maps of different layers capture information at different levels, and this multi-scale characteristic is utilized here; represents the feature map generated by the decoder at the l-th level. This feature map is the output after the convolution operation on the input features of the current layer, and is the result of gradually restoring and fusing feature information during the decoding process; this structure is often used in tasks such as image segmentation, detection, and other vision tasks, and its goal is to fully retain and utilize the features extracted by each layer to improve the overall model performance;

[0046] Use a Physics-Informed Neural Network (PINN) to regularize the model output. According to Maxwell's equations, the constraints of the electromagnetic part should ideally satisfy: is the curl operation, which is used to measure the degree of rotation of a vector field near a certain point, and G represents the electric displacement field, Denotes the partial derivative of the electric displacement field with respect to time, reflecting the variation of the electric field over time; according to Fourier's law and thermal equilibrium, the thermal conduction part satisfies the constraint: where is the temperature gradient, and Q Joule is the Joule heat loss; the physical constraint loss is introduced as L phys : where λ1 and λ2 are physical regularization weights used to adjust the relative importance of each constraint term, usually dimensionless numbers;

[0047] Two methods, Monte Carlo and Bayesian neural network, are used to quantify the uncertainty of the output. In the Monte Carlo method, Monte Carlo is added before the key layers (such as the Transformer or U-Net++ part) (the technical staff will set the dropout rate to 0.2 during use, and this dropout rate can be fine-tuned according to the actual situation). During inference, M = 50 forward propagations are performed to obtain the predicted mean and variance σ 2 , and the expression is: Output confidence interval to quantify the model uncertainty; where represents the prediction result generated by the neural network using the current parameters (the subset of parameters after random inactivation) for the input parameter X after the m-th forward propagation. θ represents all the learnable parameters in the deep learning model, including the convolutional kernel weights of the convolutional layer, the weights and biases of the fully connected layer, etc.; M is the total number of sampling times for forward propagation, which is set to 50 in this scheme; is the mean of the prediction results and serves as the point estimate output of the final model. The variance σ 2 is the variance of the prediction results, reflecting the uncertainty level of the model output. The Bayesian neural network method uses variational inference to approximate the posterior of some parameters as a Gaussian distribution q φ (θ) = N(θ|μ θ ,σ θ 2 ), and is trained by maximizing the ELBO. The training expression is: where q φ (θ) is the posterior approximation distribution of the model parameter θ, controlled by the parameter φ, and its form is assumed to be a Gaussian distribution; μ θ and σ θ 2 are the mean and variance of the Gaussian distribution, obtained through training, represents the expected value of the log-likelihood of calculating the data D under the approximate posterior distribution. KL(q φ (θ)||p(θ)) represents the approximate posterior distribution q φThe Kullback-Leibler divergence between \(q(\theta)\) and the prior distribution \(p(\theta)\) is used to constrain the update. This method enables the network to not only output point estimation results but also obtain the distribution information of the parameters, which is helpful for the uncertainty analysis of the model.

[0048] The data loss, physical loss, and uncertainty loss are combined and analyzed to construct the total loss. The calculation expression of the total loss is: where and are the predicted value and the true value respectively, \(N\) is the total number of samples in the training dataset, which is used for normalizing the error calculation, and \(n\) is any one of the samples. NLL is the negative log-likelihood, which is an index used to measure the probability of the true data under the predicted probability distribution of the model. By minimizing the NLL loss, the model can not only learn the correct output value but also reasonably estimate the uncertainty of the prediction, which is very helpful for improving the reliability of the model when facing parameter uncertainty and complex working conditions.

[0049] Through multi-scale feature extraction and cross-layer connection, fine-grained and global information in the input data can be effectively captured, enhancing the model's fitting ability for complex physical phenomena. Embedding Maxwell's equations and heat conduction equations as physical constraints and using physical information to regularize the training process enable the model's prediction results to be not only accurate in data fitting but also satisfy the basic physical laws, improving the extrapolation ability and robustness. The Monte Carlo Dropout method is adopted to quantify the uncertainty of the model prediction and give the confidence interval. Using the Bayesian neural network method, the posterior distribution of the parameters is obtained through variational inference, improving the prediction reliability of the model under complex and extreme working conditions. The network structure utilizes multi-scale features and physical constraints, making the model perform more robustly in scenarios with local nonlinearity and boundary condition changes. Uncertainty quantification provides the prediction confidence, helping engineers comprehensively consider in risk assessment and alleviating the uncertainty problem caused by the sensitivity of traditional models to prior parameters. By realizing data-driven and physical information embedding, the model can accurately predict the electromagnetic and thermal field states under complex working conditions, effectively overcoming the problem of increased prediction error of traditional prior formula-based methods under extreme conditions.

[0050] Under the premise of ensuring safety and high-performance requirements, the global optimization module uses NSGA-Ⅲ (Non-dominated Sorting Genetic Algorithm) and Bayesian optimization to globally search for the design parameters, uses the surrogate model to quickly evaluate the objective function, and finds the Pareto optimal solution in the design parameter space to achieve the best balance of indicators such as temperature, thermal stress, and energy consumption. Specifically:

[0051] Construct the objective function

[0052]

[0053] , where \(T\)max (θ design , λ design ) represents the maximum temperature of the device under extreme operating conditions, and σ vonMises (θ desgin , λ desgin ) represents the thermal stress calculated according to the VonMises criterion and is used to evaluate the risk of mechanical failure; P loss (θ desgin , λ desgin ) represents the power loss; θ desgin represents the parameters that need to be optimized in the actual device design; the additional restrictions imposed on the design parameters are the constraint conditions, and the constraint conditions are: the breakdown voltage of the device ≥ 1.2 times the rated operating voltage to ensure that the device will not break down easily during actual operation; and the switching efficiency (i.e., the ratio of the output power to the input power) is at least 95% to ensure the energy conversion efficiency;

[0054] Population initialization, define the design parameter θ desgin , where the design parameter θ desgin includes continuous variables and discrete variables; for continuous variables, Latin Hypercube Sampling (LHS) is used to generate 100 uniformly distributed sampling points in their respective domains to ensure that the samples are uniformly distributed in each dimension and avoid data clustering. For discrete variables, a random uniform assignment method is used to determine the discrete options for each individual; after combination, each initial individual is represented as θ desgin = {gate thickness, doping concentration, heat dissipation structure type, packaging material}; for each individual, an electromagnetic-thermal coupling simulation model is called to calculate three key target indicators (the maximum temperature T max (θ design , λ design ), the von Mises stress σ vonMises (θ desgin , λ desgin ) and the power loss P loss (θ desgin , λ desgin )); the dominance relationship is determined as: individual dominates individual if and only if the performance on each individual index is at least the same as , and on at least one index is significantly better than That is, if a design is at least as good as another design in all key metrics and better in at least one metric, then this design can "dominate" the other design, meaning it is overall superior; where i and j are any two individuals in the population, that is, the indices used to distinguish two different design solutions in the population; it should be noted that the continuous variables specifically include: gate thickness, which is a continuous variable with a value range of [0.1μm, 5μm]; doping concentration, which is a continuous variable with a value range of [1×10^16 cm -3 , 1×10^20 cm -3 ; the discrete variables specifically include: heat dissipation structure type, options {fin, microchannel}, and encapsulation material, options {AlN, SiC, epoxy resin}; all non-dominated individuals (i.e., individuals not dominated by other individuals) are selected and marked as Front 1, and the dominated individuals of Front 1 are removed from the remaining individuals, and then the non-dominated set is found again from the remaining individuals and marked as Front 2; this process is repeated until all individuals are stratified;

[0055] The value of each objective function is normalized and mapped to [0, 1], and the Das-Dennis method is used to generate reference points and obtain the coordinates of each reference point. The reference points are used to maintain population diversity in the environmental selection step, and offspring are generated through SBX crossover and polynomial mutation; for each individual Calculate the distance between its normalized objective function vector and the reference point (calculated using the Euclidean distance formula), assign the individual to the reference point with the closest distance, and perform environmental selection according to the density of the region where the reference point is located; according to the non-dominated sorting rank, first retain the individuals in Front 1; if the number of individuals in the front exceeds the population capacity, then those individuals with weaker association with the reference point and greater diversity are preferentially selected; use the Gaussian process surrogate model (Matérn 5 / 2 kernel function) to model the objective function, and then perform local search on the continuous variables through the EI (expected improvement) acquisition function. Combining with the enumeration method of discrete variables, and at the same time using the Ray framework to achieve parallel acceleration, quickly evaluate a large number of design parameters, so as to accelerate the global search; After NSGA-Ⅲ generates 10 generations of populations, Bayesian optimization is started to perform local fine search on the current Pareto front; the improved points obtained by Bayesian optimization are fed back to the NSGA-Ⅲ population to replace the individuals with lower performance, forming a closed loop of co-evolution until the preset termination conditions are reached; the specific termination conditions are: reaching the maximum number of iterations (which the technical personnel set to 100 generations) or reaching the convergence condition (the technical personnel set the convergence condition to that when the Pareto front improvement rate for 20 consecutive generations is less than 1%, it is considered convergent and no longer significantly improves);

[0056] Through objective function construction, constraint setting, population initialization, and non-dominated sorting, it is ensured that the optimization scheme has reliable security while achieving high performance; based on the collaborative work of NSGA-Ⅲ and Bayesian optimization, local fine search is performed and the population is updated every certain number of generations until the preset termination conditions (such as the number of iterations or the convergence rate) are reached, effectively searching the entire design parameter space and obtaining balanced solutions from multiple objective perspectives; combined with the local fine search of Bayesian optimization, it is ensured that it will not fall into local optima and obtain the Pareto optimal solution set; it solves the problems that traditional methods are prone to falling into local optima and difficult to comprehensively search the parameter space during global design optimization. Using surrogate model acceleration and global co-evolution methods, multi-objective optimization is achieved, ensuring that the design parameters reach the best balance under various constraint conditions.

[0057] The online adaptive module realizes closed-loop feedback through drift detection and update, ensuring that the model continuously adapts to production fluctuations and real-time working conditions; specifically:

[0058] Extract key parameters in real time, and use the Kolmogorov–Smirnov (KS) test to compare the real-time data distribution P real (X) with the training data P train (X) to obtain the drift value KS: Preset the drift threshold (which is set to 0.05 by the technicians in this field). If the drift value KS > the drift threshold, it means that the data distribution has drifted, and then the model update is triggered; the specific model update is: use variational inference to update the model parameters. Assume that the parameter posterior q φ (θ) = N(θ|μ θ , σ θ 2 ), and the optimization objective is to maximize the ELBO, and the expression is: Set a lower learning rate, such as 1×10 -5 , update iteratively 50 times, and adjust the model to adapt to the new data; when data drift occurs, only fine-tune the top layer of the deep model and freeze the bottom layer feature extraction part to speed up the training and avoid catastrophic forgetting;

[0059] By using the Kolmogorov–Smirnov (KS) test to compare the real-time collected data and the training data distribution, it is determined that data drift has occurred. After detecting data drift, the Bayesian posterior update (Bayes-by-Backprop) method is used to update the model parameters according to the ELBO optimization formula; during the update process, the bottom layer feature extraction network can be frozen and only the top layer is fine-tuned to improve the update speed and avoid catastrophic forgetting; it can monitor and detect the drift of the data distribution in the production environment in real time, and adjust the model parameters through online update, enabling the model to continuously adapt to actual production changes, and effectively alleviating the prediction error problem caused by parameter uncertainty of traditional models.

[0060] The real-time simulation module performs real-time simulation on the electromagnetic field and the thermal field to quickly verify the actual effects of the design parameters and trigger safety control in a timely manner when abnormalities occur. Specifically:

[0061] The three-dimensional space is divided into several sub-regions, and there are several computing units. The computing units are responsible for updating the field values of the Yee grid. There are overlapping regions or boundary regions at the grid boundaries of each sub-region, and data needs to be exchanged with adjacent computing units to ensure the continuity of numerical calculations. Adjacent computing units exchange boundary field values through a FIFO buffer to ensure data consistency, thereby completing the synchronous update of the electric field and the magnetic field. During the data exchange process, dynamic scheduling and load balancing are performed on the FIFOs between multiple computing units to ensure that the data transmission channel is always in the best state, avoid excessive data waiting time for some units, and thus improve the synchronization efficiency of the overall system. The boundary data exchange mechanism is specifically as follows:

[0062] Step 1, real-time monitoring and priority calculation: Two layers of overlapping grids (overlapping regions and boundary regions) are reserved at the boundaries of adjacent sub-regions. In each computing unit, for boundary data exchange, real-time monitoring is carried out to obtain the FIFO occupancy rate (the amount of data in the current FIFO divided by the total capacity of the FIFO to obtain the FIFO occupancy rate), the data waiting time (the time difference between the writing time and the reading time of each data packet is calculated to obtain a single waiting time, and the mean value of all single waiting times in the sliding window is calculated to obtain the data waiting time), and the link penalty value. The calculation process of the link penalty value is as follows: Through the round-trip time (RTT) calculation, a probe packet is sent every 1 ms to calculate the delay: Based on the number of CRC check failures, calculate the error rate: Penalty link value: The FIFO occupancy rate, the data waiting time, and the penalty link value are normalized and their numerical values are weighted and calculated to obtain the priority Pry (abbreviation of Priority) of data transmission;

[0063] Step 2, FIFO data exchange scheduling: When Pry≥0.8, preemption is immediately triggered (i.e., the current low-priority data transmission is paused, the high-priority FIFO data is transmitted, and the original data task transmission is resumed after completion) to quickly empty the queue and prevent overflow; when all Pry<0.8, the scheduler periodically polls all computing units and allocates transmission time slices according to the calculated dynamic priorities. The FIFO with a higher priority obtains a longer transmission time window, so that the boundary data can be exchanged to adjacent computing units faster; if congestion occurs during data transmission, the data is forwarded through intermediate computing units to relieve the pressure on a single link. If the forwarding fails three times in a row, the number of intermediate computing units is gradually increased, and the data in this link direction is distributed to multiple intermediate computing units for forwarding;

[0064] Step 3, data exchange execution: Each computing unit packs the boundary data to be exchanged into data blocks of a fixed size (which can be compressed as needed to reduce the transmission volume) and writes them into the FIFO buffer; adjacent computing units realize data exchange through the FIFO buffer. The sending end writes the data into the FIFO, and the receiving end reads the data from the FIFO; to reduce latency and waiting time, asynchronous communication and double-buffering technologies are adopted to enable parallel data transmission and local computing; after the data transmission is completed, the receiving unit uses a verification algorithm (such as CRC verification) to detect the data integrity. After ensuring that the received data is correct, the boundary data is updated to the local computing area to provide accurate boundary conditions for the next internal field value update;

[0065] The Isolation Forest algorithm is introduced as an anomaly detection method. Under normal operating conditions, the field value distributions calculated for each grid point and each region should be relatively reasonable; if abnormal data (such as extreme high temperature or abnormal stress) appears, it indicates that there may be problems with the design scheme or equipment failures; for each key parameter collected in real time, the Isolation Forest algorithm is used to calculate its path length δ. The path length represents the number of nodes that need to be passed from the root node to the key parameter in the Isolation Forest. Normal data usually requires more nodes to be "isolated", while abnormal data, due to being far from the normal distribution, can often be isolated quickly and has a shorter path; a path threshold is set. If the path length of a key parameter is less than the path threshold, the key parameter is regarded as an abnormal parameter and the safety mode is triggered; the specific safety mode is: reducing the device operating power to 80 to reduce the generated heat and stress; starting the emergency cooling equipment (such as the fan running at full speed or the liquid cooling system starting) to quickly reduce the temperature and prevent device damage; regarding the abnormal parameter and the triggered safety mode as an abnormal event, and feeding it back to the upper-layer control system through the online adaptive module, as well as updating the real-time data in the input module for subsequent optimization and model update; at the same time, recording the abnormal parameters for training and improving the proxy model;

[0066] Through FPGA hardware acceleration and parallel computing, real-time simulation of complex electromagnetic-thermal fields is achieved. Data exchange ensures the timely synchronization of boundary data through dynamic scheduling and load balancing, providing reliable real-time feedback for the entire closed-loop system, thereby greatly reducing the overall calculation time and improving system security and accuracy. Each module collaborates with each other through data and feedback to achieve a complete, closed-loop, and continuously self-adaptive updated optimization system, providing high-precision, robust, and highly automated technical support for the design of high-power transistor electromagnetic-thermal models; this solution can not only significantly reduce prediction errors and calculation time-consuming, but also improve the passing rate of the system under extreme working conditions and long-term reliability.

[0067] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0068] The above embodiments only represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent should be subject to the appended claims.

Claims

1. The optimization system of the high-power electromagnetic thermal model of a transistor based on deep learning, comprising an input module, a deep learning module, a global optimization module, and an online adaptive module; The input module obtains various key parameters from transistor process documents and online sensors, and constructs a high-dimensional input space; The deep learning module constructs a deep learning model by embedding physical constraints and uncertainty quantification mechanisms; The global optimization module performs a global search on the design parameters, uses a surrogate model to quickly evaluate the objective function, and finds the Pareto optimal solution in the design parameter space; the online adaptive module is used for drift detection and update; It is characterized in that it further includes a real-time simulation module; The real-time simulation module performs real-time simulation on the electromagnetic field and the thermal field to quickly verify the actual effects of the design parameters; the three-dimensional space is divided into several sub-regions, there are several computing units, and the computing units are responsible for updating the field values. There are overlapping regions or boundary regions at the grid boundaries of each sub-region; for the update of the electromagnetic field values in adjacent regions, boundary data exchange is adopted to ensure data consistency. The boundary data exchange mechanism is specifically as follows: Step 1, real-time monitoring and priority calculation: In each computing unit, for boundary data exchange, real-time monitor and obtain the FIFO occupancy rate, data waiting time, and link penalty value; normalize the FIFO occupancy rate, data waiting time, and penalty link value and take their numerical values for weighted calculation to obtain the priority Pry of data transmission; Step 2, FIFO data exchange scheduling: When Pry≥0.8 exists, preempt immediately; when all Pry<0.8, use the scheduler to periodically poll all computing units and allocate transmission time slices according to the calculated dynamic priority; if congestion occurs during data transmission, forward the data through the intermediate computing unit. If the forwarding fails continuously for 3 times, gradually increase the number of intermediate computing units and distribute the data in this link direction to several intermediate computing units for forwarding; Step 3, data exchange execution: Each computing unit packs the boundary data to be exchanged into data blocks of a fixed size and writes them into the FIFO buffer; adjacent computing units realize data exchange through the FIFO buffer. The sending end writes the data into the FIFO, and the receiving end reads the data from the FIFO; asynchronous communication and double-buffer technology are adopted to enable data transmission and local calculation to proceed in parallel.

2. The optimization system for the high-power electromagnetic thermal model of a transistor based on deep learning according to claim 1, wherein The real-time simulation module is also used to detect anomalies and trigger safety control in a timely manner. Its execution process is as follows: Introduce the isolation forest algorithm as an anomaly detection method. For each key parameter collected in real time, use the isolation forest algorithm to calculate its path length δ. The path length represents the number of nodes that need to be passed from the root node to the key parameter in the isolation forest; set a path threshold. If the path length of the key parameter is less than the path threshold, regard the key parameter as an abnormal parameter and trigger the safety mode; the specific safety mode is: reduce the device operating power to 80, start the emergency heat dissipation device; regard the abnormal parameter and the triggered safety mode as an abnormal event.

3. The optimized system for the high-power electromagnetic thermal model of a transistor based on deep learning according to claim 2, characterized in that The execution process of the input module constructing a high-dimensional input space is as follows: The key parameters include electrical parameters, thermal parameters, material parameters, and dynamic sensing parameters. Among them, the electrical parameters include operating voltage, current density, switching frequency, and carrier mobility. The thermal parameters include thermal conductivity, specific heat capacity, and boundary convection coefficient. The material parameters include dielectric constant, magnetic permeability, and thermal expansion coefficient. The dynamic sensing parameters include the real-time temperature field distribution T(x, y, z, t), electromagnetic field distributions E(x, y, z, t) and B(x, y, z, t), current waveform I(t), voltage waveform V(t), and stress field distribution Y(x, y, z, t). Preprocessing: For the collected electromagnetic signals, wavelet threshold denoising is adopted. Using the Daubechies 9 / 7 wavelet basis, after wavelet transform decomposition, soft threshold processing is applied. A Kalman filter is established for optimal estimation to provide accurate state correction information for the subsequent construction of the deep learning model. The Box-Cox transformation is applied to all key parameters to construct dimensionless numbers. Construct a graph neural network where each node represents a certain physical parameter, and the weights are determined according to the mutual information entropy between different parameters. Spatial features are extracted using a 3D convolutional network, and temporal features are extracted using an LSTM network. Finally, the two parts of the features are fused into a high-dimensional input.

4. The system for optimizing the high-power electromagnetic thermal model of a transistor based on deep learning according to claim 3, characterized in that, The execution process of the deep learning module is as follows: Use an encoder to extract multi-scale features, and extract the multi-scale features of the input data through residual blocks and dense connections in multiple layers. Use cross-layer connections to fuse the encoder features with the upsampled features of the decoder. Finally, an activation function needs to be applied to the output to ensure that some predicted physical quantities satisfy physical constraints. Regularize the model output using the cyber-physical network. According to Maxwell's equations, the constraints on the electromagnetic part should ideally satisfy: is the curl operation, which is used to measure the degree of rotation of a vector field near a certain point. G represents the electric displacement, represents the partial derivative of the electric displacement field with respect to time, reflecting the change of the electric field over time. According to Fourier's law and thermal equilibrium, the constraints on the heat conduction part are satisfied: where is the temperature gradient, and Q Joule is the Joule heat loss. Introduce the physical constraint loss as L phys : where λ1 and λ2 are physical regularization weights; Two methods, Monte Carlo and Bayesian neural network, are used to quantify the uncertainty of the output. In the Monte Carlo method, Monte Carlo is added before the critical layer, and during inference, M = 50 forward propagations are performed to obtain the predicted mean and variance σ 2 , and the expression is: Output confidence interval Quantify the uncertainty of the model; where represents the prediction result of the neural network using the current parameters for the input parameter X after the m-th forward propagation. θ represents all learnable parameters in the deep learning model, including the convolutional kernel weights of the convolutional layer, the weights and biases of the fully connected layer, etc.; M is the total number of sampling times for forward propagation; is the mean of the prediction results and is used as the point estimate output of the final model. The variance σ 2 is the variance of the prediction results; variational inference is used to approximate the posterior of some parameters as a Gaussian distribution and is trained by maximizing the ELBO; the data loss, physical loss, and uncertainty loss are combined and analyzed to construct the total loss.

5. The system for optimizing the high-power electromagnetic thermal model of a transistor based on deep learning according to claim 4, wherein The global optimization module uses the non-dominated sorting genetic algorithm and Bayesian optimization to perform a global search on the design parameters, and uses a surrogate model to quickly evaluate the objective function to find the Pareto optimal solution in the design parameter space. Specifically: Construct an objective function, and the additional restrictions imposed on the design parameters are used as constraint conditions. The constraint conditions are: the breakdown voltage of the device ≥ 1.2 times the rated operating voltage, and the switching efficiency reaches at least 95%. Starting from the generation of the initial population, Latin Hypercube Sampling (LHS) is used to generate design parameters for continuous variables, and then discrete variables are randomly selected. For each individual, electromagnetic-thermal simulations are performed to calculate the objective function value. The individuals are stratified through non-dominated sorting, and then the Das-Dennis method is used to generate reference points to ensure population diversity. Subsequently, offspring are generated through simulated binary crossover (SBX) and polynomial mutation. Finally, excellent individuals are retained according to the reference point association to form the next generation. Use a Gaussian process surrogate model to model the objective function, and then perform local search on the continuous variables through the expected improvement acquisition function. Combining with the enumeration method of discrete variables, the Ray framework is used to achieve parallel acceleration to quickly evaluate a large number of design parameters, thereby accelerating the global search. Through the collaborative work of the non-dominated sorting genetic algorithm and Bayesian optimization, local fine searches are performed and the population is updated every certain number of generations until the preset termination conditions are reached.

6. The system for optimizing the high-power electromagnetic thermal model of a transistor based on deep learning according to claim 5, characterized in that, The execution process of the online adaptive module is as follows: Extract key parameters in real time and use KS test to compare real-time data distribution P real (X) and training data P train (X) Get the drift value KS: The drift threshold is preset. If the drift value KS> the drift threshold, the model update is triggered. The specific model update is: the model parameters are updated using variational inference, assuming that the parameter posterior q φ (θ)=N(θ|μ θ ,σ θ 2 ), the optimization goal is to maximize ELBO, the expression is: Set the learning rate and perform a number of updates to adjust the model to the new data.

Images