FET modeling with deep neural networks and GAN-augmented small measurement dataset

Official URL: https://dx.doi.org/10.1109/SMACD65553.2025.11092163

We present FET modeling using Deep Neural Networks with a small but augmented dataset. NMOS-HV measurement data from the IHP SG13G2 OpenPDK repository, including FET dimensions, VGS, VDS, and VBS, were used for DC drain current prediction. To improve accuracy, we employed Generative Adversarial Networks (GANs) for data augmentation. Using only 78,972 measured data points with 80-20% training-test split, we achieved 11.16% MAPE across 27 FET sizes. Using DL Model on GAN-based augmented dataset further reduced the error to 9.84%, demonstrating that accurate FET models can be developed with significantly fewer data points than traditional ANN approaches. Our results highlight the potential of data-efficient Deep Learning and GAN augmentation for automated transistor modeling.