Predictive modeling of drain current in advanced FET architectures using ML-based TCAD calibration

Area of Science:

• Semiconductor device physics and engineering
• Computational electronics and materials science
• Machine learning applications in nanotechnology

Background:

• Advanced semiconductor devices like multi-gate MOSFETs are crucial for integrated circuit miniaturization.
• Accurate modeling of transistor behavior is essential but computationally challenging for sub-10nm geometries.
• Existing methods often require extensive Technology Computer-Aided Design (TCAD) simulations.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) driven approach for predicting the drain current (Id) of nanowire and nanoribbon field-effect transistors (FETs).
• To compare the performance of various ML models, including tree-based and deep learning methods, using TCAD-extracted datasets.
• To reduce the computational burden of traditional TCAD simulations for device characterization.

Main Methods:

• Extracted datasets from TCAD simulations were used to train six ML models: XGBoost, Random Forest, 1D Convolutional Neural Network (CNN), Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and TabNet.
• Model performance was evaluated using metrics such as R-squared (R²) and Root Mean Squared Error (RMSE).
• A web-based application was developed to deploy the best-performing models for real-time prediction and visualization.

Main Results:

• Ensemble tree-based methods, XGBoost and Random Forest, demonstrated superior performance in predicting drain current.
• XGBoost achieved an R² score of 0.9926 and RMSE of 2.1×10⁻⁴ A, while Random Forest achieved an R² of 0.9549 and RMSE of 2.0×10⁻⁴ A.
• Deep learning models like LSTM (R² = 0.6778) and 1D CNN (R² = 0.6997) showed significantly lower accuracy, indicating tree-based models generalize better for structured TCAD data.

Conclusions:

• Tree-based ML models are highly effective for predicting semiconductor device behavior using TCAD data, challenging the assumption that deep learning is always superior.
• The proposed ML framework significantly accelerates the prediction of device characteristics, reducing reliance on iterative TCAD simulations.
• The developed web application provides a practical tool for real-time drain current prediction and visualization, aiding in rapid device design and optimization.