10× Faster Quantum Transport Simulation With Position-Aware Attention

Researchers at Wuhan University have developed a new deep learning model capable of simulating quantum transport in atomic devices with a tenfold increase in speed. The Position-Aware Global Attention Network, or PGA-Net, utilizes a Transformer-based architecture with trainable two-dimensional positional encoding to better capture the long-range interactions critical to accurate simulation. Validated against a dataset of 6015 2D atomistic MOSFET samples, PGA-Net achieved a mean absolute error below 0.02 V in predicting electrostatic potential. Published in Applied Physics Letters, this work provides an efficient route for atomic-scale device simulation and demonstrates the potential of physics-aware global attention mechanisms for broader device modeling, according to the research team.

The architecture, developed by a team at Wuhan University, addresses a critical limitation of existing methods: accurately capturing the long-range Coulomb interactions governing electron behavior in nanoscale devices. PGA-Net achieves this through a multiscale physical field learning framework, integrating local physical encoding with a global attention decoding process. Crucially, the model incorporates a trainable two-dimensional positional encoding, enhancing its ability to understand spatial relationships within the simulated device. The model maintained an accuracy of 0.02 V when predicting electrostatic potential while simultaneously delivering the reported 10-fold speedup compared to the established nonequilibrium Green’s function method. Researchers found that the learned physical representations extend beyond potential prediction, enabling the recovery of other vital transport characteristics like charge density and local density of states. The team believes their approach offers insights into broadening the application of physics-aware global attention mechanisms to more complex device modeling scenarios, potentially accelerating materials discovery and device optimization efforts.