计算机科学
主动学习(机器学习)
反向
控制理论(社会学)
电子工程
反问题
缩小
工程类
控制工程
算法
钥匙(锁)
数学
数学优化
工作(物理)
作者
Huanhuan Zeng,Yijiang Liu,Ruifeng Yue,Yan Wang
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-06-01
卷期号:: 72-74
标识
DOI:10.1109/edssc70455.2026.11736499
摘要
This paper proposes a unified, data-efficient inverse design framework for 1200V planar 4H-SiC MOSFETs, which integrates active learning (AL) with multi-task surrogate modeling for performance prediction, together with multi-objective optimization for inverse design. Query-by-Committee (QBC)-Core-set sampling strategy is implemented to improve data efficiency in TCAD data acquisition, then by using the dataset constructed via AL, the proposed multi-task surrogate model with hard-parameter sharing (HPS) achieves excellent prediction accuracy and the mean percentage error (MAPE) of 3.5% by only 64 samples. Based on the trained surrogate model, multi-objective optimization is performed in a 9-dimensional design space searching for optimal device configuration under multiple metric constraints. The reverse design of the device is completed in less than one minute, and TCAD validation confirms that prediction errors are within 6.5%. The representative solutions identified from the Pareto set demonstrate a 32% improvement in the figure of merit compared to the best design in the original dataset.
科研通智能强力驱动
Strongly Powered by AbleSci AI