Intelligent optimization design method for electric field distribution of semiconductor device

Abstract

This invention relates to the field of intelligent design technology for semiconductor devices, and discloses an intelligent optimization design method for the electric field distribution of semiconductor devices. The method includes: acquiring region division data and electric field distribution data of the semiconductor device, generating an electric field spectrum feature vector; collecting adjustment action vectors from each decision agent, detecting whether there are conflicts in the adjustment actions of adjacent regions, and if conflicts exist, inputting the action proposals of the conflicting regions into a central coordinator neural network, outputting a coordinated compromise action vector; performing doping parameter adjustments and calculating the updated electric field distribution, calculating the improvement amounts for low-frequency and high-frequency components respectively; each decision agent updating the parameters of the strategy neural network using an allocated spectrum-sensing reward signal, and outputting an optimized multi-region doping distribution scheme. This invention solves the technical problem of oscillations that are difficult to converge due to mutual interference between regions in the intelligent optimization design of multi-region electric field distribution of semiconductor devices.

Description

Technical Field

[0001] This invention relates to the field of intelligent design technology for semiconductor devices, and more specifically, to an intelligent optimization design method for the electric field distribution of semiconductor devices. Background Technology

[0002] The design of electric field distribution in semiconductor devices is a core technical challenge in the field of power electronics. Semiconductor devices, represented by IGBTs (Insulated Gate Bipolar Transistors), comprise multiple functional regions, including the emitter region, gate region, drift region, and collector region. The doping distribution in each region collectively determines the device's electric field distribution characteristics. From a frequency domain perspective, the electric field distribution exhibits different frequency characteristics in space: the overall trend of the main conductive field in the drift region corresponds to the low-frequency components of the electric field distribution; the gate and junction regions generate local electric field spikes, corresponding to the high-frequency components of the electric field distribution.

[0003] Intelligent optimization design of electric field distribution in semiconductor devices faces the challenge of multi-region coupling: the doping optimization objectives of different functional regions are coupled and conflicting. For example, parameter adjustments to improve the electric field uniformity in the drift region may worsen the high-frequency spikes in the gate region; a unified electric field optimization objective (such as minimizing the electric field peak value) cannot distinguish between two different types of electric field anomalies: high low-frequency background and excessive high-frequency spikes; when each region is optimized according to a unified objective, it is easy for them to interfere with each other. The lack of an intelligent task allocation mechanism based on the physical characteristics of each region leads to repeated oscillations and difficulty in convergence during the device design optimization process. Summary of the Invention

[0004] This invention provides an intelligent optimization design method for the electric field distribution of semiconductor devices, which solves the technical problem in the intelligent optimization design of multi-region electric field distribution of semiconductor devices where mutual interference between regions leads to oscillations that are difficult to converge.

[0005] This invention provides a multi-agent cooperative optimization method for the electric field distribution of semiconductor devices, comprising:

[0006] The system acquires region partitioning data and electric field distribution data of semiconductor devices, performs frequency domain transformation on the electric field distribution data, and divides the electric field spectrum into low-frequency and high-frequency components according to a preset frequency threshold to generate an electric field spectrum feature vector.

[0007] Independent policy neural networks are created for each functional region of the semiconductor device as decision-making agents. Dominant spectral component labels are preset according to the contribution characteristics of each region to the electric field spectrum, and a multi-agent set is generated.

[0008] The global device state is input to each decision agent, and each decision agent outputs the adjustment action vector of the doping parameter in the corresponding region.

[0009] Collect the adjustment action vectors of each decision-making agent, detect whether there is a conflict in the adjustment actions of adjacent regions, and if there is a conflict, input the action proposal of the conflicting region into the central coordinator neural network, and output the coordinated compromise action vector.

[0010] Perform doping parameter adjustment and calculate the updated electric field distribution. Perform frequency domain decomposition on the updated electric field distribution and calculate the improvement of low-frequency components and high-frequency components respectively.

[0011] Spectrum perception reward signals are assigned based on the dominant spectrum component labels of each decision-making agent. The reward signal of the low-frequency dominant agent is generated by weighting the low-frequency improvement amount and the global uniformity improvement amount, while the reward signal of the high-frequency dominant agent is generated by weighting the high-frequency improvement amount and the global uniformity improvement amount.

[0012] Each decision-making agent uses the allocated spectrum sensing reward signal to update the parameters of the policy neural network, iterates until the electric field distribution meets the design goal, and outputs the optimized multi-region doping distribution scheme.

[0013] Furthermore, the frequency domain transformation of the electric field distribution data includes: performing a two-dimensional discrete cosine transform on the electric field intensity distribution matrix to convert the spatial electric field distribution into a frequency domain representation; and dividing the spectral coefficients into low-frequency components and high-frequency components based on the comparison result between the magnitude of the spatial frequency coordinates and a preset frequency threshold.

[0014] Furthermore, the step of presetting the dominant spectral component label based on the contribution characteristics of each region to the electric field spectrum includes: marking the region with the overall trend of the dominant electric field as a low-frequency dominant agent, and marking the region that generates local electric field spikes as a high-frequency dominant agent.

[0015] Furthermore, the detection of whether there is a conflict between the adjustment actions of adjacent regions includes: estimating the change of electric field at the interface of adjacent regions after each adjustment action. If the change direction of the low-frequency component at the interface is opposite or the change direction of the high-frequency component is opposite, then it is determined that there is a conflict.

[0016] Furthermore, the central coordinator neural network receives the action proposal vector and electric field spectrum feature vector of the conflict region, calculates the influence weight of each action proposal on the global electric field spectrum, performs a weighted average of the conflict actions based on the influence weight, and outputs the coordinated compromise action vector.

[0017] Furthermore, the calculation of the improvement amount of the low-frequency component and the improvement amount of the high-frequency component respectively includes: calculating the difference between the energy norm of the low-frequency component before the update and the energy norm of the low-frequency component after the update as the improvement amount of the low-frequency component; and calculating the difference between the energy norm of the high-frequency component before the update and the energy norm of the high-frequency component after the update as the improvement amount of the high-frequency component.

[0018] Furthermore, the improvement in global uniformity is the difference between the standard deviation of the electric field distribution before the update and the standard deviation of the electric field distribution after the update.

[0019] Furthermore, the policy neural network includes an observation encoder and a policy output layer. The observation encoder extracts features from the global device state to generate a hidden feature representation, and the policy output layer outputs doping parameters to adjust the action vector based on the hidden feature representation.

[0020] Furthermore, the iterative execution until the electric field distribution meets the design target includes: determining whether the low-frequency component energy is lower than the low-frequency target threshold, whether the high-frequency component energy is lower than the high-frequency target threshold, and whether the electric field peak value of each region meets the design specifications. If all are met, the iteration is terminated.

[0021] This invention provides an intelligent optimization design system for the electric field distribution of semiconductor devices, comprising:

[0022] The frequency domain decomposition module is used to acquire the region division data and electric field distribution data of semiconductor devices, perform frequency domain transformation on the electric field distribution data and divide it into low-frequency components and high-frequency components, and generate electric field spectrum feature vectors.

[0023] The agent creation module is used to create decision-making agents for each functional area and preset the dominant spectral component labels;

[0024] The action generation module is used to input the global device state into each decision agent and output the doping parameter adjustment action vector for each region.

[0025] The conflict coordination module is used to detect action conflicts between adjacent areas and output a compromise action vector through a central coordinator neural network;

[0026] The spectrum evaluation module is used to perform doping parameter adjustments and calculate the improvement in low-frequency and high-frequency components of the updated electric field distribution.

[0027] The reward allocation module is used to allocate spectrum sensing reward signals to each decision-making agent based on the dominant spectrum component label.

[0028] The iterative optimization module is used to update the parameters of the strategy neural network based on the reward signal, determine the termination condition of the iteration, and output the optimized multi-region doping distribution scheme.

[0029] The beneficial effects of this invention are as follows:

[0030] This invention separates the electric field distribution into low-frequency background components and high-frequency detail components through frequency domain decomposition. Since the low-frequency components mainly reflect the overall trend of the electric field, while the high-frequency components mainly reflect local peak characteristics, agents in different regions can be assigned to the spectral components they are physically good at improving for focused optimization. Drift regions and collector regions mainly affect the overall distribution of the electric field; they are marked as low-frequency dominant agents and assigned reward signals primarily based on low-frequency improvement, enabling them to focus on improving the low-frequency background of the electric field. Gate regions and emitter regions mainly affect local electric field peaks; they are marked as high-frequency dominant agents and assigned reward signals primarily based on high-frequency improvement, enabling them to focus on suppressing high-frequency peaks in the electric field. This spectrum-aware reward allocation mechanism overcomes the deficiency of unclear task allocation caused by a unified optimization objective failing to distinguish between different types of electric field anomalies.

[0031] This invention uses a central coordinator neural network to arbitrate when conflicts are detected at the interface of adjacent regions. It calculates a compromise action based on the weights of each action's impact on the global electric field spectrum, ensuring that improvements in low-frequency and high-frequency components do not interfere with each other. Since conflict detection is based on the direction of spectral change, the coordination mechanism can identify and reconcile interference situations where "adjustments in one region worsen the optimization results in another region." Therefore, this invention solves the technical problem of oscillations that are difficult to converge due to mutual interference between regions in the intelligent optimization design of multi-region electric field distribution in semiconductor devices. Attached Figure Description

[0032] Figure 1 This is a flowchart of an intelligent optimization design method for the electric field distribution of a semiconductor device according to the present invention;

[0033] Figure 2 This is a trend chart of multiple indicators changing during the optimization iteration process of this invention;

[0034] Figure 3 This is a comparison diagram of the doping parameters of each region of the present invention before and after optimization;

[0035] Figure 4 This is a stacked area diagram of the spectral feature improvement process of the present invention;

[0036] Figure 5 This is a comparison diagram of the instant reward signals of the intelligent agent according to the present invention;

[0037] Figure 6 This is the multi-agent coordination network diagram of the present invention. Detailed Implementation

[0038] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0039] At least one embodiment of the present invention discloses a multi-agent cooperative optimization method for the electric field distribution of semiconductor devices, such as... Figure 1 As shown, it includes the following steps:

[0040] Step 1: Obtain the region division data and initial electric field distribution data of the IGBT device, perform a two-dimensional discrete cosine transform on the electric field distribution data, and generate an electric field spectrum feature vector.

[0041] Specifically, the IGBT device's region partitioning data is retrieved from the semiconductor device design database. This data includes the spatial coordinate boundary information of the emitter, gate, drift, and collector regions, as well as the initial doping concentration parameters for each region. The electric field solver is then used to numerically calculate the initial doping distribution, outputting the initial electric field intensity distribution matrix. ,in and These are the spatial coordinates of the device's cross-section.

[0042] For the electric field intensity distribution matrix Perform a two-dimensional discrete cosine transform (2D-DCT) to convert the spatial domain electric field distribution into a frequency domain representation. ,in and For spatial frequency coordinates. Set a frequency threshold. The spectrum is divided into low-frequency components. and high frequency components :satisfy The spectral coefficients are classified into the low-frequency components. ,satisfy The spectral coefficients are classified into high-frequency components. Calculate the energy proportion of the low-frequency component. and the proportion of high-frequency component energy low-frequency components High-frequency components Energy percentage and The data is spliced ​​together to generate an electric field spectrum feature vector.

[0043] Furthermore, the proportion of low-frequency component energy The calculation formula is: High-frequency component energy ratio The calculation formula is: ,in Represents the L2 norm, For the energy represented in the complete frequency domain, satisfying .

[0044] It should be noted that the aforementioned two-dimensional discrete cosine transform is a conventional frequency domain transform algorithm, and its input is the electric field intensity distribution matrix. The output is a frequency domain representation matrix. The two-dimensional discrete cosine transform (DCT) performs orthogonal decomposition of spatial domain signals using cosine basis functions, achieving a mapping from the spatial domain to the frequency domain. The aforementioned frequency threshold... The value of is an empirical value predetermined based on the physical characteristics of the IGBT device, and is usually taken as a value between one-tenth and one-fifth of the size of the electric field distribution matrix.

[0045] Furthermore, frequency threshold The specific calculation formula is as follows ,in and These are the electric field intensity distribution matrices. exist direction and Dimensions of direction This is the frequency threshold coefficient, and its value range is... The frequency threshold coefficient is selected within the above range based on the device type and design requirements.

[0046] For a specific IGBT device with a rated voltage of 1200V, region partitioning data was retrieved from the device design database. The total thickness of this IGBT device is 120 micrometers, and the initial doping concentration parameters for each region are shown in Table 1.

[0047] Table 1 Initial doping parameters for different regions of the IGBT device:

[0048]

[0049] The electric field solver is invoked to perform numerical calculations on the initial doping distribution, and the size of the electric field distribution matrix is ​​set to [value missing]. , This corresponds to a spatial sampling resolution of 0.5 micrometers. The frequency threshold coefficient is then used. The frequency threshold was calculated. The spectral characteristic data obtained by performing a two-dimensional discrete cosine transform on the initial electric field intensity distribution matrix and segmenting it according to the frequency threshold are shown in Table 2.

[0050] Table 2. Spectral characteristics of the initial electric field distribution:

[0051]

[0052] Step 2: Instantiate independent policy neural networks as decision agents for each functional region of the IGBT device, and pre-define the dominant spectral component labels of each region according to its physical characteristics to generate a multi-agent set.

[0053] Specifically, independent policy neural network instances are created for the emitter region, gate region, drift region, and collector region, with each policy neural network serving as the decision-making agent for its corresponding region. Each policy neural network has the same network structure, including an observation encoder and a policy output layer, but the parameters of each network are initialized and updated independently.

[0054] Furthermore, the parameter initialization method for the policy neural network is as follows:

[0055] For the weight matrices of the observation encoder and the policy output layer The Xavier initialization method is used to initialize the data from a uniform distribution. Randomly sample initial values, where Input the number of nodes. Number of output nodes; bias vector Initialize to a zero vector.

[0056] Based on the physical characteristics of each region, a predefined dominant spectral component label is assigned: the drift region and collector region are labeled as low-frequency dominant agents, with the dominant spectral component label indicating that adjustments to the doping parameters in these two regions primarily affect the low-frequency components of the electric field distribution; the gate region and emitter region are labeled as high-frequency dominant agents, with the dominant spectral component label indicating that adjustments to the doping parameters in these two regions primarily affect the high-frequency components of the electric field distribution. These four decision-making agents and their corresponding dominant spectral component labels are then organized into a multi-agent set.

[0057] It should be noted that the aforementioned policy neural network comprises two units: an observation encoder and a policy output layer. The observation encoder consists of multiple fully connected layers, and its input layer receives the global device state vector. The global device state vector includes the doping parameters, electric field intensity distribution samples, and electric field spectrum characteristics of all regions; the data transfer relationship between the fully connected layers of the observation encoder is as follows: ,in For the first Hidden features of the layer and These are trainable parameters;

[0058] The output of the observation encoder is the hidden feature representation. The policy output layer is a fully connected layer that receives the hidden feature representation. As input, first calculate the linear transformation. Then, the output is restricted to a certain value using the Tanh activation function. The range is then multiplied by the preset maximum adjustment range. Obtain the adjustment action vector of the doping parameters in this region. Each component of the action vector represents the adjustment range of the corresponding doping parameter, and its value range is... .

[0059] Furthermore, the observation encoder contains two fully connected layers. The number of hidden nodes in the first layer is twice the dimension of the input state vector, and the number of hidden nodes in the second layer is once the dimension of the input state vector. The number of output nodes in the policy output layer is equal to the number of doping parameters in the corresponding region. For a region containing a single doping concentration parameter, the number of output nodes is 1. For a region containing multiple doping parameters, the number of output nodes is equal to the total number of doping parameters in the corresponding region.

[0060] Furthermore, the maximum adjustment range The calculation formula is determined based on the initial values ​​of the doping parameters in each region and the process tolerance. ,in This represents the initial concentration value of the doping parameters for the corresponding region. To adjust the amplitude coefficient, the value range is: The adjustment range coefficient is selected within the above range based on the safety margin of the device design.

[0061] The aforementioned policy neural network employs reinforcement learning training, specifically the policy gradient algorithm, and the Adam optimization algorithm is used for policy optimization. The loss function of the policy neural network is the policy gradient loss, defined as:

[0062]

[0063] in The policy distribution for parameterizing a policy neural network. and They are time points Actions and states, To accumulate rewards, For trajectory data, This represents the summation over all moments in the trajectory.

[0064] Furthermore, cumulative rewards The calculation formula is: ,in For a moment The instantaneous reward signal (calculated in step 6 based on the improvement in the electric field spectrum), at time... Execute steps 3 to 6 accordingly. iteration This refers to the number of iterations within a single training batch. This is the discount factor, and its value range is... The discount factor is used to balance the weights of near-term and long-term rewards.

[0065] Step 3: Input the global device state into the observation encoder of each agent, and each agent outputs the adjustment action vector of the doping parameters in its own region.

[0066] Specifically, the global device state vector is constructed as follows: it includes the current doping parameter values ​​for all regions, the sampled values ​​of the electric field intensity distribution matrix, and the electric field spectrum feature vector generated in step 1. Before constructing the state vector, different physical quantities are normalized to eliminate dimensional differences: the doping parameter values ​​and electric field intensity sampled values ​​are respectively processed using Z-score normalization, and the low-frequency components, high-frequency components, and energy proportions in the electric field spectrum feature vector are normalized to a zero-mean, unit-variance distribution. The normalized features are then concatenated to generate the global device state vector. The global device state vector is input into the observation encoders of the emitter region agent, gate region agent, drift region agent, and collector region agent, respectively.

[0067] Furthermore, the sampling method for the electric field intensity distribution matrix is ​​as follows: on the cross-section of the device... direction and Sampling grids are uniformly set in each direction, and the spacing between the sampling grids is set to one-tenth of the feature size of the device. Electric field intensity values ​​are extracted at each sampling grid node to form an electric field intensity sampling value sequence. The length of the electric field intensity sampling value sequence is equal to the total number of sampling grid nodes.

[0068] Each agent's observation encoder extracts features from the global device state vector to generate hidden feature representations. These hidden feature representations are then input into the policy output layer, which outputs adjustment action vectors for the doping parameters in its local area. Each component of the adjustment action vector represents the increment of the corresponding doping parameter; positive values ​​indicate an increase in concentration, while negative values ​​indicate a decrease in concentration.

[0069] In the first optimization iteration, a global device state vector was constructed based on the data in Tables 1 and 2. This vector was then input into the policy neural networks of four agents: the emitter region, gate region, drift region, and collector region. Each agent adjusted its state according to a preset adjustment coefficient. The maximum adjustment range is calculated, and the output adjustment action vector is shown in Table 3.

[0070] Table 3 shows the doping parameter adjustment actions of each agent in the first iteration:

[0071]

[0072] Step 4: Collect the adjustment action vectors of all agents, detect action conflicts between adjacent regions, and if conflicts exist, output compromise action vectors through the central coordinator neural network.

[0073] Specifically, the adjustment action vectors output by the transmitter agent, gate agent, drift agent, and collector agent are aggregated. For physically adjacent region pairs (transmitter-gate, gate-drift, drift-collector), conflicts in their adjustment actions are detected.

[0074] The conflict determination rule is as follows: after applying the adjustment action vector of each agent to the current doping parameter, the electric field change at the interface of adjacent regions is estimated; if the change direction of the low frequency component at the interface is opposite (one side is estimated to increase while the other side is estimated to decrease), or the change direction of the high frequency component is opposite, then a conflict is determined to exist.

[0075] It should be noted that the aforementioned prediction of electric field changes uses a local linear approximation method: for a pair of adjacent regions, the adjustment vectors of the two regions are... and The adjusted doping parameters are obtained by superimposing them onto the current doping parameters. and ;

[0076] Spatial sampling points are taken at the interface, and the Poisson equation is used. The change in local potential gradient at the calculation interface, where For electric potential, The space charge density is the result of the doping concentration. The dielectric constant is denoted by ; the change in electric field is the negative of the change in electric potential gradient. ; the change in electric field at the interface Perform a one-dimensional discrete cosine transform to separate it into low-frequency variation components. and high-frequency variation components Compare the signs of the low-frequency change components on both sides of the interface of adjacent regions. If the signs are opposite, there is a low-frequency conflict. Compare the signs of the high-frequency change components. If the signs are opposite, there is a high-frequency conflict.

[0077] Furthermore, the method for selecting spatial sampling points at the interface is as follows: Sampling line segments are set along the interface line of adjacent regions in a direction perpendicular to the interface. These sampling line segments cross both sides of the interface. Lateral extension distance is the area One-fifth of the feature size, in the region Lateral extension distance is the area One-fifth of the feature size; sampling points are evenly set along the sampling line segment, with the spacing between sampling points being one-twentieth of the interface feature size, and the change in electric field is calculated at each sampling point.

[0078] Furthermore, the mathematical condition for conflict determination is defined as follows: for adjacent regions and Low-frequency variation components on both sides of the interface and If satisfied If the frequency is low, it is determined to be a low-frequency conflict; for high-frequency variation components... and If satisfied If a conflict is found to be high-frequency, it is determined to be a high-frequency conflict; if either a low-frequency conflict or a high-frequency conflict is found to be true, it is determined that there is an action conflict between adjacent regions.

[0079] If a conflict is detected, the action proposal vector of the agent in the conflict region, the current doping parameters of the conflict region, and the electric field spectrum feature vector are input into the central coordinator neural network. The central coordinator neural network calculates the influence weight of each action proposal on the global electric field spectrum, performs a weighted average of the conflict actions based on the influence weights, and outputs a compromise action vector. The compromise action vector replaces the original action vector of the agent in the conflict region for subsequent execution.

[0080] It should be noted that the aforementioned central coordinator neural network consists of multiple fully connected layers. Its input layer receives the concatenated vector. ,in For the first Action proposal vectors for agents in conflict zones For the first The current doping parameters of each conflict region The eigenvectors of the electric field spectrum The number of conflict zones. Number the conflict regions; normalize each input component before splicing, and standardize the action proposal vector and doping parameters using Z-score, while normalizing the electric field spectrum feature vector using the same method as in step 3; the data transfer relationship between each fully connected layer is as follows: ,in For the first The hidden state of the layer and These are trainable parameters; the output layer generates a coordinated weight vector using the Softmax function. Each component of the coordination weight vector corresponds to the weight coefficient of the action proposal of each conflicting agent, satisfying the following conditions: ;

[0081] The compromise action vector is calculated by weighted average as follows: The compromise action vector replaces the original action of the agent in the conflict region for subsequent execution.

[0082] Furthermore, the central coordinator neural network consists of three fully connected layers. The number of hidden nodes in the first layer is 1.5 times the dimension of the input vector, the number of hidden nodes in the second layer is 1 times the dimension of the input vector, and the number of hidden nodes in the third layer is 0.5 times the dimension of the input vector. The number of output nodes in the output layer is equal to the number of conflict regions. This is used to output the coordination weights of each conflicting agent.

[0083] The aforementioned central coordinator neural network employs a supervised learning training mode, and the optimization strategy uses the gradient descent algorithm. During training, a pre-built conflict scenario dataset is used, which contains multiple sets of conflict action proposals and their corresponding optimal coordination weight labels.

[0084] Furthermore, the method for constructing the conflict scenario dataset is as follows: random doping parameter configurations are generated using the Monte Carlo sampling method. For each configuration, agents in each region generate random action proposals. The interface between adjacent regions is checked for conflict. If a conflict exists, the conflict scenario is retained. For the retained conflict scenarios, the electric field spectrum improvement effect after each conflict agent's action proposal is executed individually is tested. The weight corresponding to the action proposal with the best improvement effect is set to 1, and the weights of the others are set to 0, which serves as the optimal coordination weight label for the conflict scenario. Several sets of conflict scenarios and their labels are collected to form the conflict scenario dataset.

[0085] The loss function of the central coordinator neural network is the cross-entropy loss, defined as:

[0086]

[0087] in For the first The optimal weight labels for action proposals from conflicting agents. These are the prediction weights output by the central coordinator neural network.

[0088] If no conflict is detected, the central coordinator is skipped, and the original adjustment action vectors output by each agent are used directly.

[0089] The adjustment action vectors of the four agents in Table 3 were collected, and collision detection was performed on three adjacent region pairs (emitter region-gate region, gate region-drift region, and drift region-collector region). High-frequency collisions were detected at the gate region-drift region interface (located at 8 micrometers): gate region adjustment prediction increased the high-frequency components at the interface. V / m, drift region adjustment prediction reduces high-frequency components at the interface. V / m satisfies the conflict determination condition. The action proposals, doping parameters, and spectral characteristics of the gate region and drift region are input into the central coordinator neural network, and the conflict coordination results are shown in Table 4.

[0090] Table 4. Results of conflict resolution at the gate region-drift region interface:

[0091]

[0092] Figure 6 This demonstrates the network of interactions between four regional agents and the central coordinator.

[0093] Step 5: Perform the coordinated doping parameter adjustment, call the electric field solver to calculate the updated electric field distribution, perform frequency domain decomposition on the updated electric field distribution, and calculate the improvement of low-frequency components and high-frequency components.

[0094] Specifically, the adjustment vectors (coordinated or original) output from step 4 are superimposed onto the current doping parameters of the corresponding regions to generate updated doping parameter configurations. The electric field solver is then invoked to perform numerical calculations on the updated doping distribution, outputting the updated electric field intensity distribution matrix.

[0095] Perform a two-dimensional discrete cosine transform on the updated electric field intensity distribution matrix, using the same frequency threshold as in step 1. The component is divided into updated low-frequency and high-frequency components. The improvement in the low-frequency component is calculated. :

[0096]

[0097] in The low-frequency components before the update. For the updated low-frequency components, This represents the L2 norm. A positive improvement indicates a reduction in the energy of the low-frequency components, meaning that the low-frequency electric field background is improved.

[0098] Similarly, calculate the improvement amount of high-frequency components. :

[0099]

[0100] A positive improvement indicates a reduction in the energy of the high-frequency components, meaning that the high-frequency electric field spikes are suppressed.

[0101] The compromise actions in Table 4 are applied to the gate and drift regions, while the original actions in Table 3 are applied to the emitter and collector regions, generating updated doping parameter configurations. The updated electric field intensity distribution matrix is ​​calculated using the electric field solver, and a two-dimensional discrete cosine transform is performed on it, followed by segmentation at the same frequency threshold. The calculated spectral improvement is shown in Table 5.

[0102] Table 5. Improvement of electric field spectrum in the first iteration:

[0103]

[0104] Figure 4 It displays the changing trends of low-frequency and high-frequency component energy during the optimization process, and simultaneously shows the decrease in total spectral energy through stacking.

[0105] Step 6: Assign spectrum-aware reward signals based on the dominant spectrum component labels of each agent.

[0106] Specifically, the improvement in low-frequency components calculated in step 5. and high-frequency component improvement Based on the pre-set dominant spectral component labels of each agent in step 2, differentiated spectral perception reward signals are assigned to each agent.

[0107] global electric field uniformity improvement The value is calculated as the difference in the standard deviation of the electric field distribution before and after the update. A positive value indicates that the electric field distribution is more uniform.

[0108] Furthermore, the improvement in global electric field uniformity The specific calculation formula is as follows ,in To update the standard deviation of the electric field distribution before, The standard deviation of the updated electric field distribution. and These are the average values ​​of the electric field distribution matrix before and after the update. and denoted as the size of the electric field distribution matrix.

[0109] To eliminate the dimensional differences between different improvement indicators, the improvement of low-frequency components was analyzed. High-frequency component improvement and global uniformity improvement Normalization was performed separately, and the mean normalization method based on the range was used to normalize each improvement amount to the mean. The interval is used to obtain the normalized improvement amount. , and .

[0110] For low-frequency dominant agents (drift region agents and collector region agents), the instantaneous reward signal The calculation is as follows:

[0111]

[0112] in The reward weight for low-frequency improvement volume The index is used to assign global uniformity reward weights. This indicates the current optimization iteration time.

[0113] For high-frequency dominant agents (gate agent and transmitter agent), their instantaneous reward signal The calculation is as follows:

[0114]

[0115] in The reward weight is for high-frequency improvement.

[0116] Furthermore, reward weight parameters , and Satisfy normalization constraints and The reward weight for low-frequency improvements The range of values ​​is Reward weight for high-frequency improvement The range of values ​​is Global uniformity reward weight Correspondingly or .

[0117] In this embodiment of the application, in order to balance the objectives of local optimization and global optimization, the above-mentioned reward weights are... , and An adaptive adjustment strategy can be adopted: in the early stage of optimization, increase the weight of local improvement to quickly reduce the energy of the corresponding spectral components, and in the later stage of optimization, increase the weight of global uniformity to finely adjust the overall electric field distribution.

[0118] Based on the normalized improvement data in Table 5, and combined with the preset reward weight parameters... , , Differentiated instantaneous reward signals are calculated for the four agents. The reward signal allocation results are shown in Table 6.

[0119] Table 6. Instantaneous reward signals for each agent in the first iteration:

[0120]

[0121] Figure 5 The instant reward signals obtained by the four decision-making agents at different optimization stages are displayed.

[0122] Step 7: Each agent uses the allocated spectrum to sense the reward signal and updates its own policy neural network parameters through the policy gradient algorithm. Steps 3 to 7 are executed iteratively until the electric field distribution meets the design goal, and the optimized multi-region doping distribution scheme of the semiconductor device is output.

[0123] Specifically, each agent calculates its policy gradient using the spectrum-sensing reward signal allocated in step 6. For each agent, state-action-instant reward trajectory data is collected within a training batch. According to the cumulative reward calculation formula defined in step 2 Calculate the cumulative reward at each time point. Organize trajectory data into Calculate the policy gradient And update the policy neural network parameters using the Adam optimization algorithm:

[0124] Furthermore, the number of iterations within a training batch The value is determined based on the complexity of the optimization task, and its range is [value range missing]. The number of iterations within a training batch represents the number of consecutive optimization steps that need to be collected before each parameter update.

[0125]

[0126] in For learning rate, These are the parameters of the policy neural network. After updating the parameters, the probability of the policy neural network outputting a high-reward action increases under the same conditions.

[0127] Furthermore, the learning rate The range of values ​​is The learning rate is selected within the above range based on the convergence speed during the training process.

[0128] The aforementioned policy gradient algorithm is a conventional reinforcement learning optimization algorithm, which receives state-action-reward trajectory data as input. The output is the gradient of the policy neural network parameters. The policy gradient algorithm adjusts policy parameters by maximizing the expected value of the cumulative reward.

[0129] Determine the termination condition for the iteration: If the energy of the low-frequency component is lower than the low-frequency target threshold, the energy of the high-frequency component is lower than the high-frequency target threshold, and the peak electric field of each region meets the design specifications, then terminate the iteration; otherwise, return to step 3 to execute the next iteration.

[0130] Furthermore, the low-frequency target threshold is defined as follows: ,in The low-frequency component of the initial electric field distribution. The low-frequency target coefficient has a value range of [value range missing]. High-frequency target threshold is defined as ,in The high-frequency components of the initial electric field distribution. For high-frequency target coefficients, the value range is: ;

[0131] The design specifications for the peak electric field in each region are defined as follows: ,in The breakdown electric field strength of the device material. For the safety factor, the value range is: The target coefficients and safety factors mentioned above are selected within the corresponding ranges based on the device performance requirements and design objectives.

[0132] When the termination condition is met, the current doping parameter configurations for each region are output as the multi-region doping distribution scheme for the semiconductor device after intelligent optimization design. The multi-region doping distribution scheme for the semiconductor device includes the optimal doping concentration values ​​and doping distribution gradient parameters for the emitter region, gate region, drift region, and collector region.

[0133] In this embodiment of the application, in order to accelerate convergence and avoid getting trapped in local optima, an entropy regularization term can be introduced into the policy gradient algorithm to encourage the agent to maintain a certain degree of exploratory behavior in the early stages of optimization.

[0134] Each agent uses the instantaneous reward signal in Table 6 to calculate the policy gradient. The number of training batch iterations is set. Discount factor Learning rate After every 20 optimization iterations, each agent updates the policy neural network parameters. An optimization termination condition is set: low-frequency target coefficients. High-frequency target coefficient Safety factor After 8 training batches and a total of 160 optimization iterations, the electric field distribution met the design target. The optimization process and final results are shown in Table 7.

[0135] Table 7. Iterative process and final results of electric field distribution optimization for IGBT devices:

[0136]

[0137] The optimized final doping distribution scheme is as follows: emitter region doping concentration cm Gate doping concentration cm Doping concentration in the drift region cm Doping concentration in the collector region cm At this time, the low-frequency component energy V / m Below the low-frequency target threshold V / m High-frequency component energy V / m Below the high-frequency target threshold V / m Maximum electric field peak value V / m is lower than the design specifications. V / m (breakdown electric field strength of silicon material) V / m divided by the safety factor 2.0) satisfies all termination conditions.

[0138] Figure 2 The optimization effect of the electric field distribution of IGBT devices is demonstrated in 8 training batches and 160 iterations.

[0139] Figure 3 Compare the changes in doping concentration in the four functional regions (emitter region, gate region, drift region, and collector region) of the IGBT device before and after optimization.

[0140] This implementation separates the electric field distribution into low-frequency background components and high-frequency detail components through frequency domain decomposition. Since the low-frequency components mainly reflect the overall trend of the electric field, while the high-frequency components mainly reflect local peak characteristics, agents in different regions can be assigned to the spectral components they are physically good at improving for focused optimization. Drift regions and collector regions mainly affect the overall distribution of the electric field; they are marked as low-frequency dominant agents and assigned reward signals primarily based on low-frequency improvement, enabling them to focus on improving the low-frequency background of the electric field. Gate regions and emitter regions mainly affect local electric field peaks; they are marked as high-frequency dominant agents and assigned reward signals primarily based on high-frequency improvement, enabling them to focus on suppressing high-frequency peaks in the electric field. This spectrum-aware reward allocation mechanism overcomes the deficiency of unclear task allocation caused by a unified optimization objective failing to distinguish between electric field anomalies of different natures.

[0141] This implementation uses a central coordinator neural network to arbitrate when conflicts are detected at the interface of adjacent regions. It calculates a compromise action based on the weighted impact of each action on the global electric field spectrum, ensuring that improvements in low-frequency and high-frequency components do not interfere with each other. Since conflict detection is based on the direction of spectral change, the coordination mechanism can identify and reconcile interference situations where "adjustments in one region worsen the optimization results in another region." Therefore, this implementation solves the technical problem of oscillations that are difficult to converge due to mutual interference between regions in the intelligent optimization design of multi-region electric field distribution in semiconductor devices.

[0142] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A smart optimization design method for electric field distribution in semiconductor devices, characterized in that, Includes the following steps: The system acquires region partitioning data and electric field distribution data of semiconductor devices, performs frequency domain transformation on the electric field distribution data, and divides the electric field spectrum into low-frequency and high-frequency components according to a preset frequency threshold to generate an electric field spectrum feature vector. Independent policy neural networks are created for each functional region of the semiconductor device as decision-making agents. Dominant spectral component labels are preset according to the contribution characteristics of each region to the electric field spectrum, and a multi-agent set is generated. The global device state is input to each decision agent, and each decision agent outputs the adjustment action vector of the doping parameter in the corresponding region. Collect the adjustment action vectors of each decision-making agent, detect whether there is a conflict between the adjustment actions of adjacent regions, and if there is a conflict, input the action proposal of the conflicting region into the central coordinator neural network, and output the coordinated compromise action vector. Perform doping parameter adjustment and calculate the updated electric field distribution. Perform frequency domain decomposition on the updated electric field distribution and calculate the improvement of low-frequency components and high-frequency components respectively. Spectrum perception reward signals are assigned based on the dominant spectrum component labels of each decision-making agent. The reward signal of the low-frequency dominant agent is generated by weighting the low-frequency improvement amount and the global uniformity improvement amount, while the reward signal of the high-frequency dominant agent is generated by weighting the high-frequency improvement amount and the global uniformity improvement amount. Each decision-making agent uses the allocated spectrum sensing reward signal to update the parameters of the policy neural network, iterates until the electric field distribution meets the design goal, and outputs the optimized multi-region doping distribution scheme.

2. The intelligent optimization design method for electric field distribution of semiconductor devices according to claim 1, characterized in that, The frequency domain transformation of the electric field distribution data includes: performing a two-dimensional discrete cosine transform on the electric field intensity distribution matrix to convert the spatial electric field distribution into a frequency domain representation; and dividing the spectral coefficients into low-frequency components and high-frequency components based on the comparison between the magnitude of the spatial frequency coordinates and a preset frequency threshold.

3. The intelligent optimization design method for electric field distribution of semiconductor devices according to claim 1, characterized in that, The method of presetting dominant spectral component labels based on the contribution characteristics of each region to the electric field spectrum includes: marking regions with the overall trend of the dominant electric field as low-frequency dominant agents, and marking regions that generate local electric field spikes as high-frequency dominant agents.

4. The intelligent optimization design method for electric field distribution of semiconductor devices according to claim 1, characterized in that, The detection of whether there is a conflict between the adjustment actions of adjacent regions includes: estimating the change of electric field at the interface of adjacent regions after each adjustment action. If the change direction of the low-frequency component at the interface is opposite or the change direction of the high-frequency component is opposite, it is determined that there is a conflict.

5. The intelligent optimization design method for electric field distribution of semiconductor devices according to claim 1, characterized in that, The central coordinator neural network receives the action proposal vector and electric field spectrum feature vector of the conflict region, calculates the influence weight of each action proposal on the global electric field spectrum, performs a weighted average of the conflict actions based on the influence weight, and outputs the coordinated compromise action vector.

6. The intelligent optimization design method for electric field distribution of semiconductor devices according to claim 1, characterized in that, The calculation of the improvement amount of low-frequency components and high-frequency components respectively includes: calculating the difference between the energy norm of the low-frequency components before the update and the energy norm of the low-frequency components after the update as the improvement amount of low-frequency components; and calculating the difference between the energy norm of the high-frequency components before the update and the energy norm of the high-frequency components after the update as the improvement amount of high-frequency components.

7. The intelligent optimization design method for electric field distribution of semiconductor devices according to claim 1, characterized in that, The improvement in global uniformity is the difference between the standard deviation of the electric field distribution before the update and the standard deviation of the electric field distribution after the update.

8. The intelligent optimization design method for electric field distribution of semiconductor devices according to claim 1, characterized in that, The policy neural network includes an observation encoder and a policy output layer. The observation encoder extracts features from the global device state to generate hidden feature representations, and the policy output layer outputs doping parameters to adjust the action vector based on the hidden feature representations.

9. The intelligent optimization design method for electric field distribution of semiconductor devices according to claim 1, characterized in that, The iterative execution until the electric field distribution meets the design target includes: determining whether the low-frequency component energy is lower than the low-frequency target threshold, whether the high-frequency component energy is lower than the high-frequency target threshold, and whether the electric field peak value of each region meets the design specifications. If all of these are met, the iteration is terminated.

10. A smart optimization design system for the electric field distribution of a semiconductor device, used to execute the smart optimization design method for the electric field distribution of a semiconductor device according to any one of claims 1-9, characterized in that, include: The frequency domain decomposition module is used to acquire the region division data and electric field distribution data of semiconductor devices, perform frequency domain transformation on the electric field distribution data and divide it into low-frequency components and high-frequency components, and generate electric field spectrum feature vectors. The agent creation module is used to create decision-making agents for each functional area and preset the dominant spectral component labels; The action generation module is used to input the global device state into each decision agent and output the doping parameter adjustment action vector for each region. The conflict coordination module is used to detect action conflicts between adjacent areas and output a compromise action vector through a central coordinator neural network; The spectrum evaluation module is used to perform doping parameter adjustments and calculate the improvement in low-frequency and high-frequency components of the updated electric field distribution. The reward allocation module is used to allocate spectrum sensing reward signals to each decision-making agent based on the dominant spectrum component label. The iterative optimization module is used to update the parameters of the strategy neural network based on the reward signal, determine the termination condition of the iteration, and output the optimized multi-region doping distribution scheme.

Images