A Machine Learning-Based Optimal Terminal Structure Selection Method

Abstract

This invention discloses a machine learning-based method for selecting optimal junction termination structures, comprising: S1, training intelligent design models of different junction termination technology structure parameters based on machine learning; S2, constructing a model filter; S3, one-stage joint optimization; S4, two-stage specific optimization; and S5, selecting the optimal junction termination structure. This invention achieves multi-model collaborative design, which can quickly obtain the optimal junction termination structure that meets the designer's needs. Compared with traditional design relying on experience, it has advantages such as speed and accuracy, improving the design efficiency of designers, saving design time, and exhibiting small errors compared with TCAD simulation.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device design and multi-objective optimization, specifically to a method for selecting the optimal junction terminal structure based on machine learning. Background Technology

[0002] Lateral power devices, with their advantages of high breakdown voltage, high-frequency performance, and ease of integration, are widely used in high-voltage power switching and power conversion systems in industrial automation, lighting drivers, and consumer electronics. Junction termination technology significantly improves the breakdown voltage of lateral power devices, reduces leakage current, and optimizes switching speed and stability by rationally distributing the electric field at the device edge region and suppressing the accumulation of electric field peaks. Typical technologies include field limiting loops, field plates, and surface electric field reduction structures. Field limiting loops and field plate structures effectively improve the breakdown voltage by introducing induced charges at the edge of the PN junction, dispersing the edge electric field peaks, and preventing the electric field from focusing on the active junction. Surface electric field reduction technology reduces the surface electric field intensity by designing lateral gradient doping or multilayer doped regions in the drift region, achieving the effect of improving the breakdown voltage while reducing the on-resistance per unit area.

[0003] Currently, the selection of junction termination structures in device design mainly relies on traditional methods such as physical mechanism analysis and TCAD simulation. These methods require tedious calculations of various structural parameters and process variables, and physical model analysis often requires repeated verification based on experience and simplifying assumptions. Although TCAD simulation has high accuracy, it is computationally intensive, time-consuming, and prone to poor convergence when optimizing multi-dimensional parameters. For example, the design of field plate / field limiting ring termination structures is extremely sensitive to many parameters such as ring depth, spacing, and length, requiring extensive engineering experiments to obtain a better selection scheme. Machine learning has a natural advantage in handling high-dimensional parameter spaces and multi-objective optimization problems. With the increasing complexity of power device design, there is an urgent need to introduce intelligent design and selection technologies based on machine learning. By learning from large amounts of data, the optimal junction termination structure can be found automatically and quickly in the high-dimensional design space, significantly improving design efficiency and optimizing device performance. Therefore, developing a machine learning-based intelligent junction termination selection technology is of great significance for accelerating the power device design process and improving the overall performance of devices. Summary of the Invention

[0004] The purpose of this invention is to provide a machine learning-based optimal junction termination structure selection method. This method fully utilizes the powerful learning capabilities of machine learning, and through joint parameter space search and automatic optimization of junction termination parameters, it can simultaneously optimize the structural parameters of two different models in a short time to meet the breakdown voltage (V) requirement. B ) and on-resistance (R) onThe design goal is to screen out a junction termination structure with better device performance, thereby greatly accelerating the design and optimization process of semiconductor devices.

[0005] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0006] A machine learning-based optimal terminal structure selection method, the method comprising the following steps:

[0007] S1, Training intelligent design models for different terminal technology structure parameters based on machine learning:

[0008] Based on deep neural networks and differential evolution algorithms, an intelligent design model for structural parameters of different junction termination technologies is constructed. This intelligent design model for structural parameters of different junction termination technologies consists of two parts: a forward model and a reverse model. The forward model predicts the breakdown voltage and on-resistance based on the structural parameters of the device, while the reverse model designs the structural parameters of the device based on the breakdown voltage and on-resistance.

[0009] S2, Constructing the model filter:

[0010] Based on the different terminal technology structure parameters in step S1, an intelligent design model is constructed to build a model filter, which integrates the functions of the forward model and the reverse model, and the Baliga static merit coefficient is defined as the model screening condition.

[0011] S3, Phase One Joint Optimization:

[0012] Obtain the common structural parameters of devices corresponding to different junction termination technologies, design a joint optimization objective function, use the intelligent design model of structural parameters of different junction termination technologies in step S1 to obtain the extreme value of the joint optimization objective function, and perform joint optimization on the common structural parameters of devices corresponding to different junction termination technologies in the entire parameter space so that the breakdown voltage and on-resistance of the forward model and the reverse model tend to their respective preset target values, and the common structural parameters of the forward model and the reverse model are consistent.

[0013] S4, Phase Two Specific Optimization:

[0014] The design incorporates a first objective function focused on breakdown voltage optimization and a second objective function focused on on-resistance optimization. Based on fixed common structural parameters, the unique structural parameters of devices corresponding to different junction termination technologies are optimized separately. During optimization, if the on-resistance deviation in the optimization result is less than a preset deviation threshold, the first objective function focused on breakdown voltage optimization is used to improve the unique structural parameters of devices corresponding to different junction termination technologies to maximize the breakdown voltage. If the on-resistance deviation in the optimization result reaches the preset deviation threshold, the second objective function focused on on-resistance optimization is used to improve the unique structural parameters of devices corresponding to different junction termination technologies, making the on-resistance approach the target on-resistance. The optimized breakdown voltage and on-resistance are then calculated based on the optimized unique structural parameters.

[0015] S5, Selecting the optimal terminal structure:

[0016] Based on the optimized breakdown voltage and on-resistance obtained in step S4, compare the Baliga static figure of merit of devices corresponding to different junction termination technologies, and output the junction termination technology corresponding to the device with the larger Baliga static figure of merit.

[0017] Furthermore, the types of junction termination technologies include dual-thinned surface electric fields and field plates;

[0018] For the electric field of a double-thinned surface, the structural parameters include: drift region concentration N. d Drift region length L d Top silicon thickness T s Buried oxygen layer thickness T ox P-type buried oxygen layer length L p Distance D from P-type buried layer to oxide layer, thickness T2 of P-type buried layer, and doping concentration P of P-type buried layer. bd The distance from the P-type buried layer to the drain is L1, and the N+ doping concentration at the drain is N. nd Ohmic contact P+ doping concentration P d ;

[0019] For the field plate, its structural parameters include: drift region concentration N. d Drift region length L d Top silicon thickness T s Buried oxygen layer thickness T ox Drain field plate length L dfp Thickness T of oxide layer under drain field plate dfp Gate field plate length L gfp Thickness T of oxide layer under gate field plate gfp .

[0020] Step S1 further includes:

[0021] S11. For different junction termination technologies, breakdown voltage and on-resistance are selected as electrical performance parameters. Then, key structural parameters that affect breakdown voltage and on-resistance are selected respectively, the value space of each structural parameter is determined, and a dataset of the relationship between structural parameters and electrical performance parameters is established based on the TCAD simulation tool.

[0022] S12, Establish an intelligent design model for structural parameters of different junction terminal technologies based on deep neural networks and differential evolution algorithms. This intelligent design model for structural parameters of different junction terminal technologies consists of two parts: a forward model and a reverse model. The input parameters of the forward model are the structural parameters of the device, and the output parameters are the electrical performance parameters of the device. The input parameters of the reverse model are the electrical performance parameters of the device, and the output parameters are the structural parameters of the device.

[0023] S13 uses a relational dataset to train intelligent design models for different terminal technology structural parameters.

[0024] Furthermore, in step S2, the formula for calculating the Baliga static figure of merit is:

[0025] ;

[0026] In the formula, V B and R on These are the breakdown voltage and the on-resistance, respectively.

[0027] Step S3 further includes:

[0028] S31, Input performance specifications: Breakdown voltage V B and on-resistance R on ;

[0029] S32, based on the value space of each structural parameter, extract the maximum and minimum values ​​of each structural parameter in the forward and reverse models;

[0030] S33, determine the common structural parameters of the forward model and the backward model, and for each common structural parameter, extract the value space of the common structural parameter in the forward model and the backward model respectively, and take the intersection of the two as the effective range of the common structural parameter;

[0031] S34. Device models are constructed for two junction termination technologies: dual-thinning surface electric field and field plate. At the same time, a high-dimensional parameter space vector is constructed as a joint parameter vector, and the parameter space is divided into three parts: common structural parameters, dual-thinning surface electric field-specific structural parameters, and field plate-specific structural parameters.

[0032] S35 is designed as a joint optimization objective function to simultaneously optimize the breakdown voltage and on-resistance of two device models, aiming to bring the key performance parameters of the two device models close to their respective preset target values; the joint optimization objective function is:

[0033] ;

[0034] In the formula, Indicates the target breakdown voltage. The target on-resistance; and These represent the breakdown voltage and on-resistance of the double-thinned surface electric field, respectively. and These represent the breakdown voltage and on-resistance of the field plate, respectively.

[0035] S36. Based on the joint parameter vector from step S34, the performance is predicted by the intelligent design models of different terminal technology structural parameters trained in step S1. The differential evolution algorithm is called to update the joint parameter vector. After decomposition, the joint parameter vector is used to obtain the structural parameters of the forward model and the reverse model. The extreme value of the joint optimization objective function is continuously obtained through iteration to obtain the optimal solution vector and design the common structural parameters of the two device models.

[0036] Step S4 further includes:

[0037] S41, construct the first objective function J1 that favors breakdown voltage optimization and the second objective function J2 that favors on-resistance optimization respectively:

[0038] ;

[0039] ;

[0040] In the formula, Indicates the target breakdown voltage. V is the target on-resistance. B and R on These represent the breakdown voltage and on-resistance of devices corresponding to different junction termination technologies;

[0041] S42, the common structural parameters of different junction termination technologies after joint optimization are used as inherent parameters in the second-stage special optimization process. The unique structural parameters of the corresponding devices of different junction termination technologies are designed using the intelligent design model of structural parameters of different junction termination technologies. During the optimization process, if the on-resistance deviation in the optimization result is less than the preset deviation threshold, the unique structural parameters of the corresponding devices of different junction termination technologies are improved using the first objective function that tends to optimize the breakdown voltage to maximize the breakdown voltage. If the on-resistance deviation in the optimization result exceeds the preset deviation threshold, the unique structural parameters of the corresponding devices of different junction termination technologies are improved using the second objective function that tends to optimize the on-resistance to make the on-resistance tend to the target on-resistance.

[0042] S43, Repeat step S42 to perform specific optimization on the unique structural parameters of devices corresponding to different junction termination technologies until the number of automatic optimization designs reaches the number of design iterations, and output the optimized unique structural parameters and performance parameters of devices corresponding to different junction termination technologies respectively.

[0043] Furthermore, the preset deviation threshold is 5%.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] First, the optimal terminal structure selection method based on machine learning in this invention employs a dual-process approach of one-stage joint optimization and two-stage specific optimization. While maintaining the same common parameters, it utilizes R... on By constraining values ​​and optimizing unique junction termination parameters, the performance of devices can be improved, and ultimately, a better junction termination technology can be selected. The entire process requires no manual intervention and can provide users with better junction termination design technology and better device design solutions.

[0046] Second, the optimal junction terminal structure selection method based on machine learning of the present invention achieves joint optimization of common structural parameters and junction terminal parameters of two different junction terminal structures through intelligent design models of structural parameters of different junction terminal technologies. This overcomes the inefficiency of traditional manual parameter tuning and can quickly obtain the optimal junction terminal structure parameters and their performance that meet the requirements, thereby accelerating the design speed.

[0047] Third, the optimal junction termination structure selection method based on machine learning in this invention, and the intelligent design model for junction termination technical structure parameters based on deep neural networks and differential evolution algorithms, the model predicts V B and R on It has high accuracy and small error compared with TCAD simulation results, which can provide designers with more reliable guidance on structural parameters. Attached Figure Description

[0048] Figure 1This is a flowchart of the optimal terminal structure selection method based on machine learning according to the present invention. Detailed Implementation

[0049] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0050] This invention discloses a method for selecting optimal terminal structures based on machine learning, the method comprising the following steps:

[0051] S1, Training intelligent design models for different terminal technology structure parameters based on machine learning:

[0052] Based on deep neural networks and differential evolution algorithms, an intelligent design model for structural parameters of different junction termination technologies is constructed. This intelligent design model for structural parameters of different junction termination technologies consists of two parts: a forward model and a reverse model. The forward model predicts the breakdown voltage and on-resistance based on the structural parameters of the device, while the reverse model designs the structural parameters of the device based on the breakdown voltage and on-resistance.

[0053] S2, Constructing the model filter:

[0054] Based on the different terminal technology structure parameters in step S1, an intelligent design model is constructed to build a model filter, which integrates the functions of the forward model and the reverse model, and the Baliga static merit coefficient is defined as the model screening condition.

[0055] S3, Phase One Joint Optimization:

[0056] Obtain the common structural parameters of devices corresponding to different junction termination technologies, design a joint optimization objective function, use the intelligent design model of structural parameters of different junction termination technologies in step S1 to obtain the extreme value of the joint optimization objective function, and perform joint optimization on the common structural parameters of devices corresponding to different junction termination technologies in the entire parameter space so that the breakdown voltage and on-resistance of the forward model and the reverse model tend to their respective preset target values, and the common structural parameters of the forward model and the reverse model are consistent.

[0057] S4, Phase Two Specific Optimization:

[0058] The design incorporates a first objective function focused on breakdown voltage optimization and a second objective function focused on on-resistance optimization. Based on fixed common structural parameters, the unique structural parameters of devices corresponding to different junction termination technologies are optimized separately. During optimization, if the on-resistance deviation in the optimization result is less than a preset deviation threshold, the first objective function focused on breakdown voltage optimization is used to improve the unique structural parameters of devices corresponding to different junction termination technologies to maximize the breakdown voltage. If the on-resistance deviation in the optimization result exceeds the preset deviation threshold, the second objective function focused on on-resistance optimization is used to improve the unique structural parameters of devices corresponding to different junction termination technologies, making the on-resistance approach the target on-resistance. The optimized breakdown voltage and on-resistance are then calculated based on the optimized unique structural parameters.

[0059] S5, Selecting the optimal terminal structure:

[0060] Based on the optimized breakdown voltage and on-resistance obtained in step S4, compare the Baliga figure of merit (BFOM) of devices corresponding to different junction termination technologies, and output the junction termination technology corresponding to the device with the larger Baliga figure of merit.

[0061] In this embodiment of the invention, the semiconductor device is a silicon-on-insulator laterally diffused metal-oxide-semiconductor field-effect transistor. Two junction termination techniques are employed: a dual-reduced surface electric field (D-RESURF) and a field plate (FP). For example, see [link to relevant documentation]. Figure 1 The flowchart of the optimal terminal structure selection method based on machine learning of the present invention is as follows:

[0062] Step 1: Train an intelligent design model for different junction terminal technology structural parameters based on machine learning: Obtain the corresponding dataset and establish an intelligent design model for different junction terminal technology structural parameters based on deep neural networks and differential evolution algorithms. This model includes two parts: a forward model and a backward model. The forward model can perform V-model design based on the device structural parameters. B and R on The prediction, the inverse model can be based on V B and R on Design the device structure parameters.

[0063] Step 1 specifically includes the following sub-steps:

[0064] Step 11: Construct a relational dataset: For different terminal technologies, select those affecting V. B and R on The key structural parameters were determined, and reasonable value ranges for each parameter were identified. Based on the TCAD simulation tool, a relationship between the structural parameters and the electrical performance parameter V was established. B and Ron The relationship between the datasets. For D-RESURF, the structural parameters selected include the drift region concentration N. d Drift region length L d Top silicon thickness T s Buried oxygen layer thickness T ox P-type buried oxygen layer length L p Distance D from P-type buried layer to oxide layer, thickness T2 of P-type buried layer, and doping concentration P of P-type buried layer. bd The distance from the P-type buried layer to the drain is L1, and the N+ doping concentration at the drain is N. nd Ohmic contact P+ doping concentration P d Table 1 shows the range of values ​​for the D-RESURF structural parameters in this example. For FP, the structural parameters selected include the drift region concentration N. d Drift region length L d Top silicon thickness T s Buried oxygen layer thickness T ox Drain field plate length L dfp Thickness T of oxide layer under drain field plate dfp Gate field plate length L gfp Thickness T of oxide layer under gate field plate gfp Table 2 shows the range of values ​​for the FP structure parameters in this example.

[0065] Table 1. Range of D-RESURF structural parameters

[0066] Structural parameters Full name scope <![CDATA[N d ]]> Drift region concentration <![CDATA[[5×10 14 ,5×10 15 ]cm -3 ]]> <![CDATA[L d ]]> Drift zone length [3,33]μm <![CDATA[T s ]]> Top silicon thickness [6,16]μm <![CDATA[T ox ]]> Buried oxide layer thickness [0.1,4]μm <![CDATA[L p ]]> P-type buried oxide layer length <![CDATA[[0.33L d ,0.66L d ]μm]]> D Distance from P-type buried layer to oxide layer 0 <![CDATA[T2]]> P-type buried layer thickness [1,3.5]μm <![CDATA[P bd ]]> P-type buried layer doping concentration <![CDATA[[5×10 15 ,8×10 16 ]cm -3 ]]> <![CDATA[L1]]> P-type buried layer to drain distance <![CDATA[[0.1L d ,0.33L d ]μm]]> <![CDATA[N nd ]]> Drain N+ doping concentration <![CDATA[[1×10 18 ,1×10 20 ]cm -3 ]]> <![CDATA[P d ]]> Ohmic contact P+ doping concentration <![CDATA[[1×10 20 ,1×10 22 ]cm -3 ]]>

[0067] Table 2. Range of values ​​for FP structural parameters

[0068] Structural parameters Full name scope <![CDATA[N d ]]> Drift region concentration <![CDATA[[1×10 14 ,6×10 15 ]cm -3 ]]> <![CDATA[L d ]]> Drift zone length [20,60]μm <![CDATA[T s ]]> Top silicon thickness [0.2,10]μm <![CDATA[T ox ]]> Oxygen layer thickness [0.1,4]μm <![CDATA[L dfp ]]> Drain field plate length <![CDATA[[L d ×10%,L d ×50%]μm]]> <![CDATA[T dfp ]]> Oxide layer thickness under drain field plate [0.1,1.9]μm <![CDATA[L gfp ]]> Gate field plate length <![CDATA[[L d ×10%,L d ×50%]μm]]> <![CDATA[T gfp ]]> Gate field plate under oxide layer thickness [0.1,1.9]μm

[0069] Step 12: Train an intelligent design model for different junction terminal technology structural parameters based on machine learning: Establish an intelligent design model for different junction terminal technology structural parameters based on deep neural networks and differential evolution algorithms, consisting of a forward model and a backward model. The input parameters of the forward model are device structural parameters, and the output parameters are device performance parameters V. B and R on The inverse model takes device performance parameters as input and outputs device structural parameters as output. After training the intelligent design model with different junction terminal technology structural parameters, 30 sets of new data were collected using TCAD simulation software to verify the model accuracy. The verification showed that the prediction errors of both the forward and inverse models remained within 5%.

[0070] Step 2: Constructing a Model Filter: Based on the different terminal technology structure parameters trained in Step 1, intelligently design models, construct a model filter, integrate the functions of the forward and backward models, and define the Baliga static merit coefficient as the model selection criterion. This includes the following sub-steps:

[0071] Step 21: Define the model filter;

[0072] Step 22: Import the trained intelligent design models of different terminal technology structure parameters into the model filter as a library, and call its forward and reverse functions;

[0073] Step 23: The BFOM value is a key criterion for final terminal technology screening. The formula for calculating the BFOM value is as follows:

[0074] .

[0075] Step 3: One-stage joint optimization: Based on the input performance metric V B and R on Joint optimization is performed across the entire parameter space. The intelligent design model for different junction termination technologies, as described in step 1, is used to design common structural parameters for device models corresponding to different junction termination technologies. This ensures that the common structural parameters of each device model are consistent, while simultaneously predicting performance close to the input target value. In this embodiment, device models are constructed for both D-ReducedSurface (D-RESURF) and FieldPlate (FP) junction terminations. The first-stage joint optimization process includes:

[0076] Step 31: Input performance metric V B and R on , where V B 120V, R on 45mΩ·cm 2 Start optimizing the process;

[0077] Step 32, Parameter Boundary Design: Based on the relational dataset obtained in Step 1, each structural parameter adopts the boundary values ​​in the relational dataset as its maximum and minimum values;

[0078] Step 33, Common structure parameter processing: Both device models have their own junction terminal structure parameters, and there are also some overlapping parts, namely common structure parameters. Each common structure parameter obtains the corresponding value space according to step 32, and takes their intersection as the effective range of the common structure parameter. Tables 3 and 4 are the ranges of common structure parameters of the device models corresponding to D-RESURF and FP, respectively. Table 5 is the final selected range of common structure parameters.

[0079] Table 3 Common structural parameter ranges for D-RESURF device models

[0080] Structural parameters Full name scope <![CDATA[N d ]]> Drift region concentration <![CDATA[[5×10 14 ,5×10 15 ]cm -3 ]]> <![CDATA[L d ]]> Drift zone length [3,33]μm <![CDATA[T s ]]> Top silicon thickness [6,16]μm <![CDATA[T ox ]]> Oxygen layer thickness [0.1,4]μm

[0081] Table 4 Common Structural Parameter Range for FP Device Models

[0082] Structural parameters Full name scope <![CDATA[N d ]]> Drift region concentration <![CDATA[[1×10 14 ,6×10 15 ]cm -3 ]]> <![CDATA[L d ]]> Drift zone length [20,60]μm <![CDATA[T s ]]> Top silicon thickness [0.2,10]μm <![CDATA[T ox ]]> Oxygen layer thickness [0.1,4]μm

[0083] Table 5. Range of final selected common structural parameters

[0084] Structural parameters Full name scope <![CDATA[N d ]]> Drift region concentration <![CDATA[[5×10 14 ,5×10 15 ]cm -3 ]]> <![CDATA[L d ]]> Drift zone length [20,33]μm <![CDATA[T s ]]> Top silicon thickness [6,10]μm <![CDATA[T ox ]]> Oxygen layer thickness [0.1,4]μm

[0085] In step 33, because the parameter lists of the previously trained models were set differently, the positions of their common structural parameters were extracted to construct a mapping relationship between the common structural parameters of the two device models. This facilitates subsequent verification of the common structural parameters and the recombination of optimized structural parameters. For the common structural parameters, the minimum and maximum values ​​of the common structural parameters are obtained from the D-RESURF and FP device models, respectively, and their intersection is taken as the boundary. This step ensures that the common structural parameters satisfy the physical and design constraints of the two device models during the subsequent joint optimization process, avoiding result failure due to exceeding the model range.

[0086] Step 34: Construct a joint parameter vector: Construct a high-dimensional parameter space vector and divide the parameter space into three parts: common parameters, D-RESURF unique parameters, and FP unique parameters.

[0087] Step 35: Define the joint optimization objective: Ensure that the common structural parameters of the two device models are consistent, and design a joint optimization objective function to simultaneously optimize the breakdown voltage and on-resistance of the two device models. The joint optimization objective function aims to make the key performance parameters of the two device models as close as possible to the preset target values.

[0088] Specifically, let the target breakdown voltage be... The target on-resistance is The joint optimization objective function expression is as follows:

[0089] ;

[0090] Where, in the formula, Indicates the target breakdown voltage. The target on-resistance; and These represent the breakdown voltage and on-resistance of the double-thinned surface electric field, respectively. and These represent the breakdown voltage and on-resistance of the field plate, respectively.

[0091] Step 36: Obtain the first-stage joint optimization results: Based on the joint parameter vector from Step 34, predict the performance using the intelligent design models of different terminal technology structural parameters trained in Step 1, and update the joint parameter vector using the differential evolution algorithm. After decomposition, the structural parameters of the two device models are obtained. Iterate continuously to find the extreme value of the joint optimization objective function in Step 35, obtain the optimal solution vector, and design the common structural parameters of the two device models. The final first-stage joint optimization results are shown in Table 6, and the optimized common structural parameters are shown in Table 7.

[0092] Table 6 Results of Phase One Joint Optimization

[0093] Performance parameters Input performance metrics D-RESURF FP <![CDATA[V B (V)]]> 120 124.98 121.26 <![CDATA[R on (mΩ·cm 2 )]]> 45 47.31 43.56

[0094] Table 7 Common structural parameters of the one-stage joint optimization

[0095] Structural parameters Value <![CDATA[N d ]]> <![CDATA[1.01×10 15 cm -3 ]]> <![CDATA[L d ]]> 26.69μm <![CDATA[T s ]]> 6.74μm <![CDATA[T ox ]]> 2.09μm

[0096] Step 4: Two-Stage Specific Optimization: Based on the common structural parameters determined in Step 3, design the objective functions for the second-stage specific optimization, including a first objective function favoring breakdown voltage optimization and a second objective function favoring on-resistance optimization. With the common structural parameters fixed, optimize the unique structural parameters of the devices corresponding to the two junction termination technologies separately. During the optimization process, if the on-resistance deviation in the optimization result is less than a preset deviation threshold (e.g., 5%), use the first objective function favoring breakdown voltage optimization to improve the junction termination parameters to maximize the breakdown voltage. If the on-resistance deviation in the optimization result exceeds the preset deviation threshold, use the second objective function favoring on-resistance optimization to improve the junction termination structural parameters so that the on-resistance approaches the target on-resistance; specifically including the following steps:

[0097] Step 41: Obtain initial values ​​for the second-stage special optimization: Based on the structural parameters of the corresponding device models of D-RESURF and FP obtained in Step 3, use them as the starting point for the second-stage special optimization;

[0098] Step 42: Define the second-stage special optimization objective: Based on the common structural parameter values ​​designed in Step 3, fix the common structural parameters, and use different terminal structural parameter optimization models to design the unique structural parameters of the two device models respectively;

[0099] Step 43: Define the two-stage specific optimization objective function: Design the second-stage specific optimization objective function, including a first objective function favoring breakdown voltage optimization and a second objective function favoring on-resistance optimization. Based on fixed common structural parameters, optimize the unique structural parameters of the two junction termination technologies separately. During optimization, if the on-resistance deviation in the optimization result is less than 5%, use the first objective function favoring breakdown voltage optimization to improve the junction termination parameters to maximize the breakdown voltage. If the on-resistance deviation in the optimization result exceeds 5%, use the second objective function, which is more inclined towards on-resistance optimization, to improve the junction termination structural parameters to approximate the on-resistance.

[0100] The second-stage targeted optimization, based on fixed common structural parameters, focuses on optimizing the unique structural parameters of each device model, aiming to maximize V... B At the same time, ensure R on To keep the target value near the target value, two optimization objective functions with different focuses were designed:

[0101] When the device model's R on When the relative deviation from the target value is less than 5%, the optimization function uses V B Maximizing is the primary objective, while incorporating a moderate amount of R. on Bias penalty, the first objective function J1 is as follows:

[0102] ;

[0103] If the on-resistance deviation still exceeds 5% after the initial optimization of the device model, then switch to emphasizing R. on The objective function for precision control. Here, the optimization function emphasizes R... on Control, while retaining a slight V B The reward is used to maintain performance balance. The second objective function J2 is shown below:

[0104] ;

[0105] The above objective function is constructed by R on The relative deviation term from the target value and V B The normalized reward term enables comprehensive optimization and control of performance metrics. The first term is R. on The weighted penalty term for the deviation is used to suppress the R-value caused by the structural parameter shift. on Fluctuation; the second term is V B Positive incentive terms encourage optimizing the device's V-axis as much as possible during the optimization process. B The two objective functions are optimized by minimizing the objective function value. While maintaining controllable on-resistance accuracy, V is gradually increased. BThis ensures that the final optimization result achieves an effective balance between the two key performance indicators. When R... on When the deviation is small, the optimization function focuses more on V. B The improvement; when R on When the deviation exceeds a set threshold, the objective function switches to enhanced R. on This control method improves the overall stability and reliability of the optimization.

[0106] Step 44: Obtain the final optimization results: When the number of automatic optimization designs reaches the number of design iterations, output the structural parameters and performance parameters of the corresponding device models of the optimized D-RESURF and FP respectively.

[0107] Table 8 shows the performance parameters and V values ​​of the corresponding device models of D-RESURF and FP after the second-stage project optimization. B Improvement rate.

[0108] Table 8 Performance parameters for the second-stage special optimization

[0109] <![CDATA[V B (V)]]> <![CDATA[R on (mΩ·cm 2 )]]> <![CDATA[V B Improvement rate D-RESURF 195.62 45.01 63.02% FP 289.78 44.95 141.49%

[0110] The optimized structural parameters of the device model were put into the TCAD simulation software to obtain the TCAD simulation results. The comparison between the TCAD simulation results and the optimized device model results is shown in Table 9.

[0111] Table 9 Comparison of TCAD Simulation Results

[0112] <![CDATA[TCAD simulation V B (V)]]> <![CDATA[TCAD simulation R on (mΩ·cm 2 )]]> <![CDATA[V B Error (%) <![CDATA[R on Error (%) D-RESURF 199.93 41.15 4.13 8.51 FP 271.66 42.23 5.61 6.05

[0113] Step 5: Select the optimal terminal structure: Based on the V obtained in Step 4 B and R on The process compares the BFOM of device models corresponding to different junction termination technologies and outputs the device model corresponding to the junction termination technology with the larger BFOM. Specifically, it includes the following sub-steps:

[0114] Step 51: Obtain the optimized performance results: Obtain the optimized V from step 4. B and R on Value, V of the device model corresponding to D-RESURF B It is 195.62V, R on 45.01 mΩ·cm 2 FP corresponds to V in the device model B It is 289.78V, R on 44.95 mΩ·cm 2 ;

[0115] Step 52: Select the optimal junction termination technology: Based on the junction termination technology selection criteria defined in Step 2, and from the V in Step 51... Band R on The BFOM value is calculated for the device models corresponding to different junction termination technologies, and the device model with the larger BFOM is output. In this embodiment, the BFOM values ​​for the device models corresponding to the two junction termination technologies are as follows:

[0116] ;

[0117] ;

[0118] Therefore, FP is the optimal junction termination technology, and its corresponding device model is ultimately optimized to V. B It is 289.78V, R on 44.95 mΩ·cm 2 V B The improvement rate is 141.49%, and the final structural parameters are shown in Table 10.

[0119] Table 10 Structural parameters of the device model corresponding to the optimal junction termination technology FP

[0120] Structural parameters Value <![CDATA[N d ]]> <![CDATA[1.01×10 15 cm -3 ]]> <![CDATA[L d ]]> 26.69μm <![CDATA[T s ]]> 6.74μm <![CDATA[T ox ]]> 2.09μm <![CDATA[L dfp ]]> 11.12μm <![CDATA[T dfp ]]> 0.42μm <![CDATA[L gfp ]]> 8.01μm <![CDATA[T gfp ]]> 0.37μm

[0121] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0122] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A machine learning based optimal junction termination structure screening method, characterized in that, The method comprises the following steps: S1, training an intelligent design model of different junction termination technology structure parameters based on machine learning: An intelligent design model of different junction termination technology structure parameters is constructed based on a deep neural network and a differential evolution algorithm, the intelligent design model of different junction termination technology structure parameters comprises a forward model and a reverse model, the forward model is used to predict the breakdown voltage and the on-resistance according to the structure parameters of a device, and the reverse model is used to design the structure parameters of the device according to the breakdown voltage and the on-resistance; S2, constructing a model filter: A model filter is constructed according to the intelligent design model of different junction termination technology structure parameters in step S1, so as to integrate the functions of the forward model and the reverse model, and define a Baliga static merit factor as a model screening condition; S3, one-stage joint optimization: Common structure parameters of devices corresponding to different junction termination technologies are obtained, a joint optimization objective function is designed, the intelligent design model of different junction termination technology structure parameters in step S1 is used to obtain an extreme value of the joint optimization objective function, and the common structure parameters of the devices corresponding to different junction termination technologies are jointly optimized in the entire parameter space, so that the breakdown voltage and the on-resistance of the forward model and the reverse model tend to be respectively preset target values, and the common structure parameters of the forward model and the reverse model are consistent; S4, two-stage special optimization: A first objective function for breakdown voltage optimization and a second objective function for on-resistance optimization are designed, and the unique structure parameters of the devices corresponding to different junction termination technologies are respectively optimized on the basis of fixing the common structure parameters; in the optimization process, if the on-resistance deviation in the optimization result is less than a preset deviation threshold, the unique structure parameters of the devices corresponding to different junction termination technologies are improved by using the first objective function for breakdown voltage optimization, so as to maximize the breakdown voltage; if the on-resistance deviation in the optimization result reaches the preset deviation threshold, the unique structure parameters of the devices corresponding to different junction termination technologies are improved by using the second objective function for on-resistance optimization, so that the on-resistance tends to be a target on-resistance; and the optimized breakdown voltage and on-resistance are calculated according to the optimized unique structure parameters; Step S4 further comprises: S41, respectively constructing a first objective function J1 for breakdown voltage optimization and a second objective function J2 for on-resistance optimization: ； ； In the formula, represents the target breakdown voltage, is the target on-resistance, V B and R on are the breakdown voltage and on-resistance of the device corresponding to different junction terminal technologies, respectively. S42, taking the common structure parameters of different junction termination technologies after joint optimization as inherent parameters in the two-stage special optimization process, and using the intelligent design model of different junction termination technology structure parameters to design the unique structure parameters of the devices corresponding to different junction termination technologies; in the optimization process, if the on-resistance deviation in the optimization result is less than a preset deviation threshold, the unique structure parameters of the devices corresponding to different junction termination technologies are improved by using the first objective function for breakdown voltage optimization, so as to maximize the breakdown voltage; if the on-resistance deviation in the optimization result exceeds the preset deviation threshold, the unique structure parameters of the devices corresponding to different junction termination technologies are improved by using the second objective function for on-resistance optimization, so that the on-resistance tends to be a target on-resistance; and the optimized breakdown voltage and on-resistance are calculated according to the optimized unique structure parameters; S43, repeating step S42 to perform special optimization on the unique structure parameters of the devices corresponding to different junction termination technologies until the number of automatic optimization design reaches the iteration number of design, and then outputting the optimized unique structure parameters and performance parameters of the devices corresponding to different junction termination technologies respectively; S5, screening the optimal junction termination structure: According to the optimized breakdown voltage and on-resistance obtained in step S4, comparing the Baliga static merit coefficients of the devices corresponding to different junction termination technologies, and outputting the device corresponding to the junction termination technology with a larger Baliga static merit coefficient.

2. The machine learning based optimal junction termination structure screening method of claim 1, wherein, The different junction termination technologies include double-thinning surface electric field and field plate; For double-reduced surface electric field, the structural parameters include: drift region concentration N d , drift region length L d , top layer silicon thickness T s , buried oxygen layer thickness T ox , P-type buried oxygen layer length L p , P-type buried layer to oxide layer distance D, P-type buried layer thickness T2, P-type buried layer doping concentration P bd , P-type buried layer to drain distance L1, drain N+ doping concentration N nd , ohmic contact P+ doping concentration P d ; For field plate, its structure parameters include: drift region concentration N d , drift region length L d , top silicon thickness T s , buried oxide layer thickness T ox , drain field plate length L dfp , oxide layer thickness under drain field plate T dfp , gate field plate length L gfp , oxide layer thickness under gate field plate T gfp .

3. The machine learning based optimal junction termination structure screening method of claim 1, wherein, Step S1 further comprises: S11, selecting breakdown voltage and on-resistance as electrical performance parameters for different junction termination technologies, then selecting key structure parameters affecting the breakdown voltage and on-resistance respectively, determining the value space of each structure parameter, and establishing a relationship data set between the structure parameters and the electrical performance parameters based on a TCAD simulation tool; S12, establishing a different junction termination technology structure parameter intelligent design model based on deep neural network and differential evolution algorithm, which includes a forward model and a reverse model, the input parameters of the forward model are the structure parameters of the device, and the output parameters are the electrical performance parameters of the device; the input parameters of the reverse model are the electrical performance parameters of the device, and the output parameters are the structure parameters of the device; S13, training the different junction termination technology structure parameter intelligent design model using the relationship data set.

4. The machine learning based optimal junction termination structure screening method of claim 1, wherein, In step S2, the calculation formula of the Baliga static merit coefficient is: ； In the formula, V B and R on are the breakdown voltage and on-resistance, respectively.

5. The machine learning based optimal junction termination structure screening method of claim 3, wherein, Step S3 further comprises: S31, input performance index: breakdown voltage V B and on-resistance R on ; S32, extracting the maximum and minimum values of each structure parameter in the forward model and the reverse model according to the value space of each structure parameter; S33, determining the common structure parameters of the forward model and the reverse model, extracting the value space of each common structure parameter in the forward model and the reverse model respectively, and taking the intersection as the effective range of the common structure parameter; S34, constructing device models for double-thinning surface electric field and field plate respectively, and simultaneously constructing a high-dimensional parameter space vector as a joint parameter vector, and dividing the parameter space into three parts: common structure parameters, double-thinning surface electric field unique structure parameters and field plate unique structure parameters; S35, designing a joint optimization objective function for simultaneously optimizing the breakdown voltage and on-resistance of the two device models, which is used to make the key performance parameters of the two device models approach their respective preset target values; the joint optimization objective function is: ； wherein, represents the target breakdown voltage, is the target on-resistance; and respectively represent the breakdown voltage and on-resistance of the double-thinned surface field, and respectively represent the breakdown voltage and on-resistance of the field plate; S36, predicting the performance based on the joint parameter vector of step S34 through the different junction termination technology structure parameter intelligent design model trained in step S1, calling the differential evolution algorithm to update the joint parameter vector, and obtaining the optimal solution vector by continuously iterating the extreme value of the joint optimization objective function, and obtaining the common structure parameters of the two device models.

6. The machine learning based optimal junction termination structure screening method of claim 1, wherein, The preset deviation threshold is 5%.

Images