New Neural Network Framework Enhances Quantum Transport Predictions for Field-Effect Transistors

Researchers have developed a physics-integrated neural network framework aimed at improving the prediction of quantum transport in field-effect transistors (FETs). This new approach, detailed in the paper titled "Physics-integrated Neural Network for Quantum Transport Prediction of Field-effect Transistor" by Xiuying Zhang, Linqiang Xu, Jing Lu, Zhaofu Zhang, and Lei Shen, addresses the challenges posed by traditional machine learning methods in this domain, particularly the need for high-fidelity datasets and the integration of physical knowledge.

The study introduces a high-fidelity database that includes transport curves for sub-5-nm gate-all-around (GAA) FETs, which were compiled from existing literature and first-principles calculations. The researchers expanded their analysis beyond silicon to include materials such as indium arsenide, indium phosphide, and selenium nanowires, which are vital for advancing FET technology.

A key innovation of this research is the development of a physical-knowledge-integrated hyper vector neural network (PHVNN). This model incorporates five new physical features into its inputs, achieving a mean absolute error of just 0.39 in its predictions. Notably, approximately 98% of the current prediction residuals fell within one order of magnitude, indicating a high level of accuracy.

The PHVNN was particularly effective in screening symmetric p-type GAA FETs that exhibit similar performance metrics to n-type counterparts, which is essential for the production of homogeneous complementary metal-oxide-semiconductor (CMOS) circuits. The automatic differentiation analysis performed in the study also provided insights into the contributions of the new input parameters, enhancing the reliability of the predictions made by the PHVNN.

This research offers a promising method for rapidly screening suitable GAA FETs, potentially accelerating the design process for next-generation electronic devices. The findings are significant for the semiconductor industry, where optimizing FET designs is crucial for advancing technology in various applications.