A Scalable Machine Learning based Device Model with Mixture-of-Expert Neural Networks for Enhanced Accuracy and Efficiency

This paper proposes a novel global modeling framework based on a Mixture of Experts (MoE) architecture to address the challenge that conventional compact global models face in capturing the size-dependent operational characteristics of gate-all-around field-effect transistor (GAAFET). The proposed framework dynamically activates specialized expert models through an adaptive gating network, effectively integrating device geometric parameters with bias voltage conditions to enable precise prediction of device current characteristics. Comprehensive experimental results demonstrate that the proposed MoE-based model can significantly improve prediction accuracy while ensuring computational efficiency, especially when dealing with high-dimensional and complex current voltage (IV) characteristics, exhibiting strong adaptability and generalization ability. Compared to conventional global models with equivalent parameter counts, the proposed method reduces root mean square error (RMSE) by 20.97% while accelerating inference speed by 53.57%. When benchmarked against models with comparable inference speeds, it achieves superior performance with 54.26% lower RMSE and 48.88% reduced mean absolute percentage error (MAPE). These results highlight the framework's strong adaptability to device scaling variations and enhanced generalization capabilities across diverse operating regimes.