Accelerating the Self-consistent Quantum Transport Simulation Model With Advanced Trained Neural Networks

As semiconductor devices continue to scale into the nanometer regime, the non-equilibrium Green’s function (NEGF) formalism, coupled with Poisson’s equation, remains one of the most reliable methods for quantum transport simulations. However, NEGF simulation model is known to be computationally intensive due to the repetitive matrix operations and energy integrations involved, which limits their applications in device and circuit design. In this paper, we present a machine learning (ML)-based NEGF framework that uses convolutional neural networks (CNNs), long-short-term memory (LSTM) networks, and gated recurrent unit (GRU) networks to significantly accelerate computation time. Our approach replaces the computationally expensive iterative NEGF-Poisson loop by directly predicting electrostatic potentials through the ML model. By training on labeled datasets, our framework enhances computational efficiency while maintaining high accuracy. The results show that the CNN model outperforms in training accuracy and root mean square error, with only a marginally increased computation time, even when using smaller batch sizes, compared to the LSTM and GRU models.