Real Time TCAD Calibration via Transfer Learning

This study introduces a novel calibration framework of deep learning model, called Real Time TCAD (RTT) calibration. The proposed method comprises two steps: the pre-training step and the transfer learning step. In the pre-training step, a deep learning model is trained with a large amount of data generated from the uncalibrated TCAD simulation to learn the underlying semiconductor physics. During the transfer learning step, the pre-trained model is fine-tuned with actual measurements to minimize the discrepancy between the model predictions and the actual measurements. To prevent overfitting to the experimental data, we introduce a novel loss function that enables the model to learn the modified physics while preserving the underlying TCAD knowledge. The proposed method significantly reduces both calibration and data generation time while maintaining high model accuracy. Furthermore, our method tackles the critical problem of missing values in the data, which is commonly encountered in practice. Experimental results on a 28- nm logic process demonstrate the effectiveness of the proposed method.