Physics-Integrated Machine Learning for Efficient Design and Optimization of a Nanoscale Carbon Nanotube Field-Effect Transistor

We propose an efficient framework for optimizing the design of Carbon Nanotube Field-Effect Transistor (CNTFET) through the integration of device physics, machine learning (ML), and multi-objective optimization (MOO). Firstly, we leverage the calibrated TCAD model based on experimental data to dissect the physical mechanisms of CNTFET, gaining insights into its operational principles and unique physical properties. This model also serves as a foundation, enabling multi-scale performance evaluations essential for dataset construction. In the ML phase, a chain structure of Support Vector Regression (SVR Chain) guided by a comprehensive statistical analysis of the design metrics is utilized to predict the design metrics. The surrogate model based on the SVR Chain achieves an average mean absolute percentage error (MAPE) of 1.59% across all design metrics without overfitting, even with limited data. The established ML model exhibits its competence in rapidly producing a global response surface for multi-scale CNTFET. Remarkably, an anomalous equivalent oxide thickness (EOT) and ON-state current (I _on ) relationship is observed in CNTFET behavior due to extreme gate length scaling in long channel devices. This intriguing observation is further elucidated through a physics-based explanation. We further compare shallow and deep learning-based TCAD digital twins for model selection guidance. Using the Non-Dominated Sorted Genetic Algorithm-II (NSGA-II) in MOO, we harmonize metrics at both device and circuit levels, significantly reducing the design space. The closed-loop framework expedites the early-stage development of advanced transistors, overcoming the challenges posed by limited data.