A Physics-Related Parameter Decoupling and Self-Attention-Based Framework for NBTI Aging Prediction

Negative bias temperature instability (NBTI) is a key factor limiting the reliability of advanced semiconductor devices. However, accurate aging modeling faces the dual challenges of complex parameter extraction in traditional physical models and the heavy data dependency of pure data-driven models. To address these issues, this article proposes a neural-network framework based on physics-related parameter decoupling and an attention mechanism, named the attention-based physics-related parameter decoupling network (APPD-Net). By structurally embedding the power-law model and utilizing multitask learning (MTL) together with a self-attention mechanism, this framework constructs a high-accuracy prediction model for NBTI degradation. Validated on experimental data from 0.18 —m pMOSFETs, APPD-Net outperforms classic neural network models such as the artificial neural network (ANN), convolutional neural network (CNN), and Transformer, even when using as little as 20% of the available training samples. Furthermore, the framework autonomously decouples the intermediate physics-related parameters A and n, which can be used for compact aging modeling, while maintaining strong long-term extrapolation capability. APPD-Net combines the predictive power of deep learning with the structural advantages of physical models, providing a practical approach for developing highfidelity and explainable aging models for SPICE-oriented circuit simulation.