Method for establishing silicon carbide MOSFET TCAD physical model based on artificial intelligence technology

Abstract

The invention discloses a silicon carbide MOSFET TCAD physical model establishing method based on the artificial intelligence technology. The method comprises the following steps: constructing a device simulation structure and a physical mechanism model; key parameters are determined, and a data set is generated through TCAD simulation; training an agent model by using the data set, and carrying out key parameter optimization based on an improved genetic algorithm; and substituting the key parameter optimization value into TCAD simulation software, and combining a device simulation structure and a physical mechanism model to obtain a required model. According to the method, the complexity dilemma of traditional modeling parameters is broken through, key parameters such as oxide layer fixed charges, an interface trap charge distribution function and an interface roughness scattering factor are accurately locked, interface scattering and defect distribution core elements are focused, secondary parameter interference is abandoned, model construction accurately focuses on core driving factors, and the modeling efficiency and pertinence are greatly improved; the improved genetic algorithm is used to optimize and extract appropriate key parameters, so that the optimized key parameters are more accurate and reliable.

Description

Technical Field

[0001] The present invention relates to the field of computer-aided semiconductor design, and in particular to a method for establishing a silicon carbide MOSFET TCAD physical model based on artificial intelligence technology, which can be applied to the design of silicon carbide MOS devices. Background Art

[0002] In the early stages of device design, TCAD software simulation is generally needed to explore device performance and reduce development costs. In today's TCAD, the simulation of silicon-based devices has been relatively well developed. However, for other semiconductors, especially third-generation semiconductors, TCAD simulation can only qualitatively analyze device performance and cannot quantitatively provide reliable guidance for the optimization design of device performance.

[0003] As a representative of the third generation of semiconductors, silicon carbide is particularly prone to this problem in simulation design. Due to the serious lattice mismatch between the gate oxide of the silicon carbide MOSFET device and the silicon carbide interface during the growth process, a large amount of interface defect charges are introduced. The density of these defects is generally two orders of magnitude higher than that of silicon-based devices. These defects cause a series of problems in silicon carbide MOSFET devices, such as increased subthreshold slope, threshold voltage drift, and reduced saturation leakage current. To a certain extent, these defects directly determine the performance of silicon carbide MOSFET devices. Therefore, determining the correct interface state parameters in the simulation design stage of the device is crucial to the optimal design of the device.

[0004] Generally, establishing a reliable TCAD physical model that matches experimental data requires professional engineers to adjust it in combination with characterization methods such as the CV curve of MOS capacitors. S The interface states extracted by methods such as the method still have a large error with the actual distribution, and still need time-consuming and laborious debugging. In recent years, artificial intelligence technology has developed rapidly, and it has become increasingly closely connected with more and more industries. When faced with problems with complex and cumbersome mechanisms, artificial intelligence technology can provide great help. Therefore, the present invention proposes a method for quickly establishing a reliable silicon carbide MOSFET TCAD physical model based on artificial intelligence technology. Summary of the invention

[0005] The purpose of the present invention is to provide a method for establishing a TCAD physical model of a silicon carbide MOSFET based on artificial intelligence technology in view of the deficiencies in the prior art, which mainly involves the determination of parameters such as interface states and interface scattering coefficients in the TCAD modeling process. The method provided by the present invention successfully combines neural networks, optimization algorithms and TCAD simulation, thereby accelerating the process of establishing a device model through TCAD iterative simulation.

[0006] To achieve the above purpose, a method for establishing a TCAD physical model of a silicon carbide MOSFET based on artificial intelligence technology comprises the following steps: Step 1: Use TCAD simulation software to model the semiconductor device structure and obtain the device simulation structure; Step 2: Determine various physical mechanism models; The various physical mechanism models include a bandgap width model, an incomplete ionization model, a mobility model, and a carrier high field velocity saturation model in a conductive channel; Step 3: Screen out multiple key parameters from interface state distribution related parameters and interface scattering related parameters of SiC MOSFET devices; Step 4: Generate data set through TCAD simulation: Within the value range of multiple key parameters, multiple key parameter sets are generated by randomly selecting values; the parameter values ​​of each key parameter set are different from each other; Input multiple key parameter sets into TCAD simulation software, combine device simulation structure and various physical mechanism models, and obtain a data set of key parameters of drain-source output current through TCAD simulation; Step 5: Design the surrogate model and train it using the dataset generated by TCAD simulation: Step 6: Use the agent model to optimize key parameters based on the improved genetic algorithm; Step 7: Substitute the optimized values ​​of key parameters obtained in step 6 into the TCAD simulation software, and combine the device simulation structure and various physical mechanism models to obtain a TCAD physical simulation model of silicon carbide MOSFET that can reflect the actual working conditions of the device.

[0007] Preferably, the device simulation structure in step 1 uses a Gaussian distribution to simulate the distribution formed by ion implantation.

[0008] Preferably, the mobility model in step 2 includes an interface mobility degradation model Lombardi, an interface charge scattering model InterfaceCharge, an inter-carrier scattering and ionized impurity scattering model PhuMob, and an impurity doping mobility degradation model Arora.

[0009] Preferably, the key parameters in step 3 are composed of the oxide layer fixed charge (Qot), the interface trap charge (Qit) distribution function and the interface roughness scattering factor.

[0010] Preferably, the proxy model in step 5 adopts a neural network, the structure of the neural network is that the input layer contains N neurons, contains multiple hidden layers, each hidden layer has multiple neurons, the activation function is sigmoid, and the output layer has one neuron; N is the number of key parameters.

[0011] Preferably, the specific optimization process of the improved genetic algorithm described in step 6 is as follows: 1) First determine the evaluation function of the optimization process; 2) Initialize the population: Initialize the population, and set the genetic DNA of each individual in the population to X={key parameter}; 3) Select an operation: Calculate the fitness of each individual according to the evaluation function, and arrange the individuals in the population in ascending order of fitness; then divide the population into four groups according to four equal parts, denoted as X1, X2, X3, and X4; finally, pair the four groups according to {X1, X2}, {X1, X3}, {X1, X4}, {X2, X3}; 4) Crossover operation: Among the four pairs obtained after selection, each pair is subjected to crossover operation as the center; after the crossover is completed, the fitness is sorted from small to large; 5) Mutation operation; 6) Replacement and retention: Every K rounds, the best m individuals in the current population are retained and the remaining (nm) individuals are randomly initialized.

[0012] Preferably, if the key parameter is the oxide fixed charge (Q ot ), interface trap charge (Q it ) distribution function and the interface roughness scattering factor, then the evaluation function used in the improved genetic algorithm described in step 6 is determined as: Among them, I i is the current at the i-th point on the corresponding transfer characteristic and output characteristic curve, V GSi is the gate-source voltage corresponding to the i-th point on the characteristic curve, V DSi is the drain-source voltage corresponding to the i-th point on the curve, M is the total number of points on the characteristic curve, and the F function represents the proxy model; They represent the charge density of the oxide layer, the peak concentration of the exponential distribution, the attenuation factor of the exponential distribution, the uniform distribution concentration, the uniform distribution width, and the interface roughness scattering coefficient, respectively.

[0013] Preferably, the crossover operation in the improved genetic algorithm described in step 6 is specifically: Among them, X* i represents the individual obtained after the crossover of the i-th individual in each pairing, XA i represents the i-th individual in the former of each pairing, and r ijis a random number between 0 and 1, XA ij is the jth gene of the i-th individual in each pairing, and XB ij is the jth gene of the i-th individual in each pairing.

[0014] Preferably, the mutation rule in the improved genetic algorithm described in step 6 is as follows: Where X ik is the kth gene of the ith individual, X max and X min They are The upper and lower limits, r is a random number between [0,1], t is the current evolutionary generation, T max represents the maximum number of evolutionary generations.

[0015] Compared with the prior art, the present invention achieves multi-dimensional innovation in the method of establishing the SiC MOSFET TCAD physical model, reshaping the semiconductor design process and efficiency: (1) The key parameters of the TCAD physical simulation model for calibrating MOSFET devices of the third-generation semiconductor SiC with a large defect state density are clarified, breaking through the complexity of traditional modeling parameters and accurately locking key parameters such as oxide layer fixed charge, interface trap charge distribution function and interface roughness scattering factor. Therefore, the ineffective debugging of the remaining tens of thousands of parameters in the process of establishing a TCAD physical model consistent with experimental data is effectively avoided, the modeling process is accelerated, and the focus is on SiO 2 / SiC interface scattering and defect distribution core elements, abandoning the interference of minor parameters, so that model construction can accurately focus on the core driving factors, greatly improving modeling efficiency and pertinence.

[0016] (2) A neural network is used as a proxy model to deeply analyze the complex nonlinear relationship between the drain-source output current and key parameters, thereby replacing the TCAD simulation process. It can accurately reflect the relationship between the device's drain-source output current and key parameters, and greatly accelerate the iterative simulation optimization process, significantly shortening the design cycle.

[0017] (3) An improved genetic algorithm is used to optimize and extract appropriate key parameters, making the optimized key parameters more accurate and reliable. The unique evaluation function fits the device performance characterization. Innovative population operations and evolutionary strategies balance global search and local deep mining, effectively avoiding the premature and local optimal traps of traditional algorithms, ensuring that the global optimal parameter combination is obtained, ensuring that the model is highly consistent with the actual device performance, and comprehensively improving the model's accuracy, reliability, and universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is the structure diagram of the N-channel SiC MOSFET device in TCAD.

[0019] Figure 2 The effects of different oxide layer fixed charge densities on device performance, including (a) subthreshold region characteristics; (b) strong inversion region characteristics; (c) transconductance.

[0020] Figure 3 Figure 3. Effects of different donor-like defect interface state densities on device performance, including (a) subthreshold region characteristics and (b) strong inversion region characteristics.

[0021] Figure 4 Figure 3. Effects of different acceptor defect interface state densities on device performance, including (a) subthreshold region characteristics and (b) strong inversion region characteristics.

[0022] Figure 5 The effects of different interface roughness scattering on device performance, including (a) subthreshold region characteristics; (b) strong inversion region characteristics; (c) transconductance.

[0023] Figure 6 The effects of different interface phonon scattering on device performance, including (a) subthreshold region characteristics and (b) strong inversion region characteristics.

[0024] Figure 7 It is a schematic diagram of the structure of the proxy model of the present invention.

[0025] Figure 8 It is a process of optimizing key parameters based on surrogate model using improved genetic algorithm.

[0026] Fig. 9 It is the overall optimization and establishment process of TCAD physical model.

[0027] Fig.10 The transfer characteristics of the model finally optimized are shown in Figure 2, where (a) the subthreshold region characteristics of the transfer characteristics; (b) the strong inversion region characteristics of the transfer characteristics; and (c) the transconductance.

[0028] Fig.11 It is the output characteristic result of the final optimized model. DETAILED DESCRIPTION

[0029] To more clearly illustrate the purpose, scheme and advantages of the present invention, the following description is made in conjunction with the accompanying drawings.

[0030] As can be seen from the background technology, TCAD semiconductor simulation is an important method for semiconductor device design, but if the physical model parameters and device structure parameters are not set correctly, the simulation data is often very different from the experimental data, and the design and tape-out of the device cannot be effectively guided. Therefore, the TCAD modeling work of semiconductor devices based on experimental data is particularly important. The interface state density is higher in silicon carbide MOSFET, which has a strong impact on device performance, so it is necessary to focus on calibrating its simulation model to establish a reliable simulation model, so as to effectively guide the design of the device.

[0031] Therefore, this embodiment provides a method for establishing a TCAD physical model of a silicon carbide MOSFET based on artificial intelligence technology, comprising the following steps: Step 1: In order to speed up the simulation, the semiconductor device structure is modeled using TCAD simulation software to obtain the device simulation structure; The TCAD simulation software may adopt Sentaurus TCAD software, and utilize the SDE module to perform semiconductor device structure modeling.

[0032] In order to speed up the simulation and reflect the doping distribution in the real device, the present invention uses Gaussian distribution to simulate the distribution formed by ion implantation, and finally constructs the device structure as shown in FIG. Figure 1 .

[0033] Step 2: To establish a reliable TCAD physical simulation model, various appropriate physical mechanism models must be determined; First of all, according to the characteristics of SiC, the correct bandgap width model needs to be used. And because the ionization energy of the doped impurities is relatively large and cannot be completely ionized at room temperature, the incomplete ionization model needs to be activated.

[0034] The mobility model plays a key role in the performance of MOSFET devices, and mainly considers the carrier mobility in the conductive channel. In the channel, due to factors such as surface phonon scattering, interface roughness scattering, and Coulomb scattering caused by interface state charges, the mobility is severely degraded in the channel, and the Lombardi model and InterfaceCharge model are used here to describe it. In addition, the high-field velocity saturation effect of carriers caused by the increase in drain-source voltage is described by HighFieldSaturation, which uses the Cannali model by default. At the same time, the mobility degradation caused by impurity doping is considered, and the Arora model is used. In order to more comprehensively consider the mobility of the device, the PhuMob Philips unified model is activated to describe the influence of inter-carrier scattering and ionized impurity scattering on carrier mobility.

[0035] Therefore, the various physical mechanism models include a bandgap width model, an incomplete ionization model, a mobility model, and a carrier high-field velocity saturation model in a conductive channel; The mobility model includes an interface mobility degradation model Lombardi, an interface charge scattering model InterfaceCharge, an inter-carrier scattering and ionized impurity scattering model PhuMob, and an impurity doping mobility degradation model Arora; Step 3: Screen out the oxide layer fixed charge (Q) from the interface state related parameters and interface scattering related parameters of the SiC MOSFET device ot ), interface trap charge (Q it ) distribution function and interface roughness scattering factor as key parameters; The performance of SiC MOSFET devices is mainly related to the channel width-to-length ratio and oxide thickness geometric parameters, but also to the SiO 2 The scattering at the / SiC interface is closely related, so the key parameters mainly consider the geometric parameters, various scattering related parameters at the interface and the interface defect distribution parameters.

[0036] The fixed charge of the oxide layer in SiC MOSFET is mainly positive charge; the distribution of defects in the bandgap is generally a "semi-U-shaped" distribution. The defects in the bandgap are divided into acceptor-like defects and donor-like defects. The donor-like defects in N-channel MOSFET are generally around 10 11 Magnitude, combined Figure 3 The simulation analysis does not take it as a key parameter here; by changing the fixed charge of the oxide layer at the device interface (Q ot ), donor-like defect interface state distribution (DonorQ it ), acceptor-like defect interface state distribution (AcceptorQ it ), roughness scattering coefficient (δ) and interface phonon scattering coefficient (λ) parameters, and simulate the SiC MOSFET device to obtain the following Figure 2 Middle (a) to Figure 2 Middle (c), Figure 3 Middle (a) to Figure 3 Middle (b), Figure 4 Middle (a) to Figure 4 Middle (b), Figure 5 Middle (a) to Figure 5 Middle (c), Figure 6 Middle (a) to Figure 6 The results shown in (b) ultimately determine the oxide layer fixed charge (Q ot ), interface defect interface state distribution function and interface roughness scattering factor (δ) are used as key parameters to establish a reliable SiC MOSFET model.

[0037] In Sentaurus TCAD software, interface state distribution generally includes single energy level distribution, uniform distribution, exponential distribution and Gaussian distribution. Combining the interface distribution extracted by traditional high and low frequency methods and previous research reports, this modeling uses a model that combines uniform distribution and exponential distribution. Combined with the simulation results, it can be seen that both acceptor-like and donor-like interface states have a significant impact on the performance of MOSFET devices. Among them, in N-channel MOSFET, acceptor-like defects mainly affect the strong inversion region of the device, and donor-like defects mainly affect the subthreshold region of the device. However, in N-channel MOSFET, the density of donor-like defects is generally on the order of 10 to the 11th power, and the impact on the device is almost negligible. Therefore, the defects at the interface mainly consider the acceptor-like defect interface state (AcceptorQ it According to the “semi-U-shaped” distribution of defects, the exponential distribution and uniform distribution are combined in TCAD software to achieve it.

[0038] In this modeling, the exponential distribution model used is as follows: Among them, N 0 It means E=E 0 The defect state density at E 0 Here, E is set to the energy level at the conduction band. S Represents the attenuation factor.

[0039] In addition, the uniform distribution model used is as follows: Among them, N 0 Indicates the size of uniform distribution of defect state density, E 0 represents the center of the uniform distribution, E S Indicates the width of uniform distribution. In addition, since the thickness of the oxide layer of the device is different, it means that the oxidation growth time is different. As the growth time changes, the density of defect states formed by the accumulation of carbon clusters at the interface is also different, so the distribution of defect states is also closely related to the thickness of the oxide layer.

[0040] The key parameter finally determined is the oxide fixed charge (Q ot ), interface trap charge (Q it ) distribution function and interface roughness scattering factor, including the device channel width W, channel length L, oxide layer thickness T ox , oxide layer charge density N ot , exponential distribution peak concentration N E0 , the decay factor E of the exponential distribution ES , uniformly distributed concentration N U0 , uniformly distributed center E UM , uniform distribution width EUS and the interface roughness scattering coefficient δ.

[0041] Step 4: Generate data set through TCAD simulation: After determining the key parameters of the device, a large number of iterative simulation debugging are required to obtain the key parameters that can correctly characterize the actual device performance. However, during the debugging process, the coupling between the parameters is obvious and each simulation takes a lot of time, making the debugging process time-consuming and laborious.

[0042] As the development of artificial intelligence technology is increasingly integrated with the field of semiconductor design, more and more artificial intelligence algorithms are applied to semiconductor design technology. The present invention constructs a device characteristic curve about a key parameter data set, designs a reasonable proxy model, uses this proxy model to replace TCAD simulation, and then optimizes the appropriate key parameters through an optimization algorithm to establish a reliable TCAD physical simulation model that is consistent with the device experimental data.

[0043] Therefore, within the value range of each key parameter, multiple key parameter sets are randomly generated; the values ​​of each key parameter set are different from each other; multiple key parameter sets are input into TCAD simulation software, combined with device simulation structure and various physical mechanism models, and a data set of drain-source output current related to key parameters is obtained through TCAD simulation; For example, a SiC MOSFET device with a channel length of 10μm, a channel length of 5μm, and an oxide layer thickness of 50nm is taken as an example. Combining the traditional interface state extraction method and previous research reports, the setting values ​​of each key parameter are shown in Table 1:

[0044] Table 1 Values ​​of key device parameters

[0045] Finally, different combinations of the above key parameter values ​​are input into the Sdevice component for simulation to obtain the relationship between the key parameters and the drain output current.

[0046] Step 5: Design the surrogate model and train it using the dataset generated by TCAD simulation: After obtaining the data set of drain-source output current with respect to key parameters, it is necessary to obtain the relationship between drain-source output current and key parameters by training the proxy model, so as to proxy the TCAD simulation process for the subsequent rapid key parameter optimization process.

[0047] In order to quickly and accurately obtain the key parameters of the device, the proxy model used here must be able to accurately reflect the output characteristics of the device about the key parameter relationship. Commonly used regression algorithms now include ridge regression, decision tree regression, random forest regression and support vector machine regression, but their regression prediction accuracy is lower than that of the neural network that has developed rapidly in recent years. Therefore, in order to fit the relationship between the two to the greatest extent in the present invention, a neural network model is used to implement a proxy model. Here, the input layer of the neural network includes 9 neurons, contains 4 hidden layers, each hidden layer has a total of 10 neurons, the activation function is sigmoid, and the output layer has one neuron.

[0048] Build a basic neural network model ( Figure 7 ), the above-mentioned data set is used for training, and finally the relationship between the key parameters and the output current is obtained.

[0049] Step 6: Use the agent model to optimize key parameters based on the improved genetic algorithm: Use the proxy model combined with the improved genetic algorithm to obtain the key parameters that best fit the actual device characteristics; The specific optimization process of the improved genetic algorithm is as follows: 1) First, determine the evaluation function of the optimization process. Here, the evaluation function used is determined as: Among them, I i is the current at the i-th point on the corresponding transfer characteristic and output characteristic curve, V GSi is the gate-source voltage corresponding to the i-th point on the characteristic curve, V DSi is the drain-source voltage corresponding to the i-th point on the curve, M is the total number of points on the characteristic curve, and the F function represents the proxy model; They represent the charge density of the oxide layer, the peak concentration of the exponential distribution, the attenuation factor of the exponential distribution, the uniform distribution concentration, the uniform distribution width, and the interface roughness scattering coefficient respectively; 2) Initialize the population; The ranges of the parameters described in Table 1 are used as the upper and lower limits of the initialization of the parameters in the population and the population is randomly initialized in the form of real number coding. The gene DNA of each individual in the population is set to X={N ot , N E0 ,E ES , N U0 , E UM , E US ,δ}; For example, the population size is initialized to 200 and the maximum number of iterations is set to 2000.

[0050] 3) Selection operation: In the process of genetic optimization, choosing which individuals to perform subsequent operations plays a crucial role in the optimization process of the entire algorithm.

[0051] Therefore, the fitness of each individual is calculated according to the evaluation function, and the individuals in the population are arranged from small to large according to the fitness. The smaller the fitness, the closer the group parameters are to the actual device situation. Then the population is divided into four groups according to four equal parts, which are recorded as X1, X2, X3, and X4 respectively; finally, the four groups are paired according to {X1, X2}, {X1, X3}, {X1, X4}, {X2, X3};

[0052] 4) Crossover operation; In the four pairs obtained after selection, each pair was cross-operated with the former as the center; Among them, X* i represents the individual obtained after the crossover of the i-th individual in each pairing, XA i represents the i-th individual in the former of each pairing, and r ij is a random number between 0 and 1, XA ij is the jth gene of the i-th individual in each pair, and XB ij is the jth gene of the i-th individual in each pair; After the crossover is completed, sort them in ascending order of fitness; 5) Mutation operation; Mutation is to introduce more diversity into the population. Under the premise of not affecting the convergence of the algorithm, the following mutation rules are adopted: Where X ik is the kth gene of the ith individual, X max and X min They are The upper and lower limits, r is a random number between [0,1], t is the current evolutionary generation, T max represents the maximum number of evolutionary generations.

[0053] 6) Replacement and retention; In order to prevent the algorithm from premature maturation, it is necessary to introduce stronger diversity. Every K rounds, keep the best m individuals in the current population and randomly initialize the remaining (nm) individuals. In this example, m=50, K=9, n=200. Finally, in order to prevent the loss of high-quality individuals in the population, it is necessary to keep the best m individuals from the parent generation at the beginning and after crossover of each iteration, and replace the worst m individuals after mutation. The final optimization process is as follows Figure 8 shown.

[0054] After determining the evaluation function, the improved genetic algorithm used in the present invention can obtain the required key parameter values ​​that can reflect the device characteristics through a series of iterative operations such as population initialization, selection operation, crossover operation, mutation operation, retention and replacement on the key parameters.

[0055] Step 7: Model verification: Substitute the optimized values ​​of key parameters obtained in the previous step into the TCAD simulation software, and combine the device simulation structure and various physical mechanism models to obtain a TCAD physical simulation model of silicon carbide MOSFET that can reflect the actual working conditions of the device.

[0056] The overall TCAD physical modeling optimization process is as follows: Fig. 9 As shown in Table 2, through iterative optimization of the overall algorithm, the values ​​of the key parameters are finally obtained: Table 2 Optimization results of key device parameters

[0057] Substituting the key parameter values ​​obtained in Table 2 into TCAD, the final simulation results are as follows Fig.10 Middle (a) - Fig.10 Middle (c), Fig.11 The results shown prove that this method is effective for establishing the physical model of SiC TCAD MOSFET. Based on this method, a SiC TCAD MOSFET physical model that is consistent with the actual device working conditions can be established quickly and accurately.

[0058] In summary, the present invention focuses on the method for establishing a TCAD physical model of silicon carbide MOSFET, which has far-reaching significance in the field of semiconductor design, opens up a new path for solving the simulation problems of silicon carbide devices, and at least achieves: 1) Accurately anchor key parameters such as oxide layer fixed charge, interface trap charge distribution function, and interface roughness scattering factor to avoid complicated and ineffective debugging. By screening key parameters, the model is built in a targeted manner, which greatly shortens the modeling cycle, meets the industry's urgent need for efficient design processes, helps engineers focus on optimizing key factors, improves design efficiency and quality, accelerates product launch, and enhances market competitiveness.

[0059] 2) Introducing a neural network proxy model to replace traditional TCAD simulation, deeply exploring the intrinsic relationship between drain-source output current and key parameters, and accurately mapping device characteristics. After training with a large number of data sets, its prediction accuracy far exceeds that of conventional regression algorithms, and it can quickly respond to changes in design parameters, iteratively optimize models, reduce design costs and time consumption, and inject vitality into R&D. It is especially suitable for complex and changeable design scenarios, ensuring the unity of design flexibility and accuracy.

[0060] 3) Based on the improved genetic algorithm to optimize key parameters, the unique evaluation function, population operation and parameter setting work together to ensure that the algorithm converges to the global optimal solution. From the initialization of population diversity expansion, through selection, crossover, and mutation operations, to elite retention and population update balance, multiple links work closely together to improve the accuracy of key parameters, ensure that the model is close to the actual device performance, lay a solid foundation for the design of high-performance silicon carbide MOSFET, improve product performance consistency and reliability, meet high-demand application scenarios, such as electric vehicles, renewable energy conversion systems, etc., which have strict standards for the stability and efficiency of power devices, expand the scope of product application, and lead the industry's technology development towards precision and efficiency.

[0061] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for establishing a TCAD physical model of a silicon carbide MOSFET based on artificial intelligence technology, characterized in that The following steps are involved: Step 1: Use TCAD simulation software to model the semiconductor device structure and obtain the device simulation structure; Step 2: Determine various physical mechanism models; The various physical mechanism models include a bandgap width model, an incomplete ionization model, a mobility model, and a carrier high field velocity saturation model in a conductive channel; Step 3: Screen out multiple key parameters from interface state distribution related parameters and interface scattering related parameters of SiC MOSFET devices; Step 4: Generate data set through TCAD simulation: Within the value range of multiple key parameters, multiple key parameter sets are generated by randomly selecting values; the parameter values ​​of each key parameter set are different from each other; Input multiple key parameter sets into TCAD simulation software, combine device simulation structure and various physical mechanism models, and obtain a data set of key parameters of drain-source output current through TCAD simulation; Step 5: Design a surrogate model and train the surrogate model using the data set generated by TCAD simulation: Step 6: Use the surrogate model to optimize key parameters based on the improved genetic algorithm; Step 7: Substitute the optimized values ​​of key parameters obtained in step 6 into the TCAD simulation software, and combine the device simulation structure and various physical mechanism models to obtain a TCAD physical simulation model of silicon carbide MOSFET that can reflect the actual working conditions of the device.

2. The method according to claim 1, characterized in that: The device simulation structure described in step 1 uses a Gaussian distribution to simulate the distribution formed by ion implantation.

3. The method according to claim 1, characterized in that: The mobility models in step 2 include the interface mobility degradation model Lombardi, the interface charge scattering model InterfaceCharge, the inter-carrier scattering and ionized impurity scattering model PhuMob, and the impurity doping mobility degradation model Arora.

4. The method according to claim 1, characterized in that: The key parameters described in step 3 are composed of the fixed charge of the oxide layer, the interface trap charge distribution function and the interface roughness scattering factor.

5. The method according to claim 1, characterized in that: The proxy model described in step 5 adopts a neural network. The structure of the neural network is that the input layer contains N neurons, contains multiple hidden layers, each hidden layer has multiple neurons, the activation function is sigmoid, and the output layer has one neuron; N is the number of key parameters.

6. The method according to claim 1, characterized in that: The specific optimization process of the improved genetic algorithm described in step 6 is as follows: 1) First determine the evaluation function of the optimization process; 2) Initialize the population: Initialize the population, and set the genetic DNA of each individual in the population to X={key parameter}; 3) Select an operation: Calculate the fitness of each individual according to the evaluation function, and arrange the individuals in the population in ascending order of fitness; then divide the population into four groups according to four equal parts, denoted as X1, X2, X3, and X4; finally, pair the four groups according to {X1, X2}, {X1, X3}, {X1, X4}, {X2, X3}; 4) Crossover operation: Among the four pairs obtained after selection, each pair is subjected to crossover operation as the center; after the crossover is completed, the fitness is sorted from small to large; 5) Mutation operation; 6) Replacement and retention: Every K rounds, the best m individuals in the current population are retained and the remaining (nm) individuals are randomly initialized.

7. The method according to claim 6, characterized in that: If the key parameters are composed of the oxide layer fixed charge, the interface trap charge distribution function and the interface roughness scattering factor, the evaluation function used in the improved genetic algorithm described in step 6 is determined as:

8. Among them, I i is the current at the i-th point on the corresponding transfer characteristic and output characteristic curve, V GSi is the gate-source voltage corresponding to the i-th point on the characteristic curve, V DSi is the drain-source voltage corresponding to the i-th point on the curve, M is the total number of points on the characteristic curve, and the F function represents the proxy model; They represent the charge density of the oxide layer, the peak concentration of the exponential distribution, the attenuation factor of the exponential distribution, the uniform distribution concentration, the uniform distribution width, and the interface roughness scattering coefficient, respectively.

9. The method according to claim 6, characterized in that: The crossover operation in the improved genetic algorithm described in step 6 is specifically:

10. Among them, X* i represents the individual obtained after crossover of the i-th individual in each pairing, XA i represents the i-th individual in the former position in each pairing, r ij is a random number between 0 and 1, XA ij is the jth gene of the i-th individual in each pairing, and XB ij is the jth gene of the i-th individual in each pairing.

11. The method according to claim 6, characterized in that: The mutation rule in the improved genetic algorithm described in step 6 is as follows:

12. Among them X ik is the kth gene of the ith individual, X max and X min They are The upper and lower limits, r is a random number between [0,1], t is the current evolutionary generation, T max represents the maximum number of evolutionary generations.

Images