Design Optimization of Multiple Stepped Oxide Field Plate Trench MOSFETs with Machine Learning for Ultralow On-resistance

We propose a machine learning approach for power device design that consists of TCAD simulation and line Bayesian optimization. It has the advantage of multivariable optimization of field plate (FP) MOSFETs, and we thus discover an optimal structure for 100-V-class 8-step oxide FP-MOSFET. Compared with gradient FP-MOSFET, 8-step FP-MOSFET reduces the specific on-resistance (R ON A) by 20% and power losses by 6.2%–7.9%. Further, these losses can be reduced up to 11.4%–17.1% by considering parasitic charges in the objective function. We analyze the optimal 8-step structure, and discuss the reason why R ON A reduces.