Explainable artificial intelligence-based prediction of electrical characteristics in GAA MOSFETs

Abstract

In this paper, explainable artificial intelligence (XAI) based prediction is used for gate all around (GAA) MOSFET to predict its electrical behaviour. The data set is produced using TCAD simulations by changing device parameters like channel length (L g ), Radius of silicon pillar (R), work function (Φ m ), doping concentration (N d ), oxide thickness (t ox ) and drain bias (V ds ). The Extreme Gradient Boosting (XGBoost) algorithm is considered for training. The R 2 of 99.5% is achieved for threshold voltage prediction. The contribution of each device parameter to the model predictions at the global and local levels is determined by using Shapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) for the purpose of model interpretability. SHAP measures the relative relevance of input factors for the overall dataset, while LIME explains individual predictions through the construction of local surrogate models. SHAP analysis revealed that gate length was the most influential parameter with a mean absolute SHAP value of 0.1. This was followed by the metal gate work function, drain bias, nanowire radius, channel doping concentration and oxide thickness with mean absolute SHAP values of 0.065, 0.05, 0.041, 0.026 and 0.015 respectively. The concordance between the SHAP, LIME and the well-established electrostatic principles of GAA MOSFETs confirm that the proposed XAI framework delivers accurate, transparent and physically interpretable predictions, making it a potential tool for AI-assisted design and optimization of semiconductor devices.