Performance Evaluation and Accelerated Optimization of 4H-SiC Power Devices Based on Neural Networks

Compared to traditional technology computer-aided design (TCAD) simulations, using neural networks to predict semiconductor device performance does not face convergence problems. This advantage is particularly significant when simulating devices made of materials like silicon carbide (SiC), which exhibit complex physical behaviors, making them difficult to converge in simulations. In addition, traditional TCAD software lacks the capability to deduce device structural parameters from device performance metrics. This article selects four critical structural parameters of 4H-SiC trench gate MOS devices: trench depth (D_t), gate oxide thickness (T_ox), drift region doping concentration (N_d), and P-region channel P-region length (L) as variables. Firstly, two types of neural network architectures were constructed and trained to serve as a classifier and a value predictor, respectively, among them, the breakdown mechanism classifier achieved an accuracy rate of 97% in the validation process. The average error of breakdown voltage prediction was 5.6%. In order to ensure the accuracy and stability of the prediction, we randomly selected 1000 sets of parameters within the value range for simulation to obtain a new dataset and improve the neural network structure. The improved neural network achieved average errors of 2.9% and 4.9% in the prediction of breakdown voltage and on-resistance, respectively. Subsequently, we built an optimizer based on the improved neural network, achieving an automated design process for device structural parameters according to target breakdown voltage and on-resistance. In the accuracy validation of the optimizer, the average error between target values and actual values of breakdown voltage and on-resistance is 2.5% and 7.9%, respectively.