CNN-Based Rapid Co-Optimization of BV and R_ON,_sp for 4H-SiC SJ MOSFET

The structural design of power devices tends to be more complex, which leads to the problem of gradient explosion in the high-dimensional parameter design space. It is more difficult to discover the inherent physical laws among numerous design parameters. Using emerging artificial intelligence technologies to assist the structural design of power devices has become a research hotspot. Taking the 4HSiC super junction metal oxide semiconductor field effect transistor (4H-SiC SJ MOSFET) power device as an example, this paper uses the Sentaurus TCAD software to simulate the breakdown voltage (BV), special on-resistance (R<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ON,sp</inf>)of the SJ MOSFET. Based on the convolutional neural network (CNN), an multi-scale one- dimensional convolutional neural network (MS-1DCNN) model suitable for 4H-SiC SJ MOSFET is further proposed, which can assist the structural design of the SJ MOSFET power device and the multi-objective optimization of BV and R<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ON,sp</inf>. The MS-1DCNN model can not only achieve a prediction accuracy of over 97% for BV, R<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ON,sp</inf>, but also automatically and intelligently provide optimized device design schemes, thereby significantly reducing the device design cost and improving the design efficiency.