Device Modeling Based on Cost-Sensitive Densely Connected Deep Neural Networks

Abstract

Engineers used TCAD tools for semiconductor devices modeling. However, it is computationally expensive and time-consuming for advanced devices with smaller dimensions. Therefore, this work proposes a machine learning-based device modeling algorithm to capture the complex nonlinear relationship between parameters and electrical characteristics of gate-all-around (GAA) nanowire field-effect transistors (NWFETs) from technology computer-aided design (TCAD) simulation results. This method utilizes a densely connected deep neural networks (DenseDNN), which establishes direct connections between layers in the neural networks, provides stronger feature extraction and information transmission capabilities. By incorporating cost-sensitive learning methods, the proposed model focus more on the critical data that determines device characteristics, leading to accurate prediction of key device characteristics under various parameters. Experimental results on a test dataset of 116 NWFETs demonstrate the effectiveness of this method. The DenseDNN model with cost-sensitive learning exhibits better performance than traditional deep neural networks (DNN) with various widths and depths, with a prediction error below 1.62%. Moreover, compared to TCAD simulation results, the model can speedup $10^{6}\times$ .