Position-aware quantum transport simulation 10x faster

Researchers at Wuhan University have developed a new deep learning model that can simulate quantum transport in atomic devices 10 times faster. Position-Aware Global Attendant Network (PGA-Net) leverages a Transformer-based architecture with trainable 2D position encoding to better capture long-range interactions that are important for accurate simulation. Validated on a dataset of 6015 2D atomic MOSFET samples, PGA-Net achieved an average absolute error of less than 0.02 V in predicting the electrostatic potential. The research, published in Applied Physics Letters, provides an efficient route for atomic-scale device simulation and demonstrates the potential of physics-aware global attention mechanisms for broader device modeling, the researchers said.

The architecture, developed by a team at Wuhan University, addresses a critical limitation of existing methods in accurately capturing the long-range Coulomb interactions that govern the behavior of electrons in nanoscale devices. PGA-Net accomplishes this through a multiscale physical field learning framework that integrates local physical encoding and global attention decoding processes. Importantly, this model incorporates trainable 2D positional encoding, enhancing its ability to understand spatial relationships within a simulated device. The model achieved a reported 10-fold speedup compared to the established nonequilibrium Green’s function method while maintaining an accuracy of 0.02 V in predicting the electrostatic potential. The researchers found that the learned physical representation extends beyond potential predictions, allowing recovery of other important transport properties such as charge density and local density of states. The research team believes their approach provides insights for extending the application of physics-aware global attention mechanisms to more complex device modeling scenarios, potentially accelerating materials discovery and device optimization efforts.