Quantum transport in atomistic devices is governed by both localized quantum coupling and long-range Coulomb interactions. However, existing deep learning models for device simulation struggle to capture these inherently nonlocal effects. To address this challenge, we propose the Position-Aware Global Attention Network (PGA-Net), a Transformer-based architecture for quantum transport simulation. PGA-Net introduces a multiscale physical field learning framework that integrates local physical encoding with global attention decoding, along with a trainable two-dimensional (2D) positional encoding to enhance spatial structure awareness. By learning an end-to-end mapping from device inputs to the atomistic electrostatic potential outputs, PGA-Net achieves up to a $10\times$ simulation speedup while maintaining consistency with the nonequilibrium Green’s function method. On a dataset of 6015 2D atomistic MOSFET samples, PGA-Net attains a mean absolute error below 0.02 V for potential prediction. Moreover, the learned physical representations can be further used to recover other key transport-related quantities, such as the charge density and the local density of states. Overall, this work provides an efficient route for atomic-scale device simulation under similar physical assumptions and boundary conditions and offers insights into extending physics-aware global attention mechanisms to broader device-level modeling.

Close

PrevNext

View figure in article

• Quantum transport
• Graphene
• MOSFET
• Nanoribbon
• Deep learning
• Multiscale modeling
• Neural network simulations