Interpreting artificial neural network-based modeling of 4 H-SiC mosfets using explainable AI.

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Science

Background:

• Wide bandgap (WBG) semiconductors like 4H-SiC MOSFETs are crucial for advanced power electronics, offering high efficiency and thermal stability.
• Traditional Technology Computer-Aided Design (TCAD) simulations for these devices are computationally intensive and lack scalability.
• Process variations significantly impact the electrical performance of WBG semiconductors.

Purpose of the Study:

• To develop a novel, explainable machine learning framework for accurate and interpretable device modeling of 4H-SiC MOSFETs.
• To address the computational limitations and scalability issues of conventional TCAD simulations.
• To provide a transparent, data-driven approach for understanding and optimizing semiconductor device design.

Main Methods:

• Integration of artificial neural networks (ANNs) with explainable artificial intelligence (XAI) techniques, specifically SHapley Additive exPlanations (SHAP).
• Training the ANN on extensive TCAD-generated datasets encompassing diverse structural and doping parameters.
• Utilizing SHAP to quantify the influence of individual design parameters on device electrical characteristics.

Main Results:

• The proposed model achieved a Pearson correlation coefficient greater than 0.99 for predicting on-state current.
• SHAP analysis demonstrated physically consistent relationships, such as the inverse correlation between drain current and oxide thickness/channel length.
• The framework successfully provided interpretable insights into design-performance correlations in SiC MOSFETs.

Conclusions:

• This research establishes a transparent and data-driven framework for understanding and optimizing SiC MOSFETs using explainable AI.
• The methodology offers a scalable and accurate alternative to conventional TCAD for device modeling.
• The approach is adaptable to other semiconductor technologies requiring both high accuracy and interpretability in device modeling.