High-Accurate and -Efficient Potential for BCC Iron based on The Physically Informed Artificial Neural Networks

Atomic forces and energies, calculated by interatomic potential, are fundamental components of molecular dynamics (MD) and Monte Carlo (MC) simulations. Compared with traditional potential, machine learning (ML) potentials trained on extensive DFT databases offer enhanced accuracy in predicting physical and chemical properties of materials, but their transferability often faces constraints. To address this limitation, physically informed neural network (PINN) potentials have emerged. These models synergistically combine the strengths of ML with physics-based bond-order interatomic potentials, aiming for both improved accuracy and broader applicability. However, a major hurdle remains: the low performance of PINN potentials, hindering large-scale simulations. This work introduces a novel approach that significantly improves the performance of PINN potentials. The developed PINN potential for body-centered cubic (BCC) iron demonstrates exceptional accuracy in property predictions while boasting remarkable computational efficiency. Its performance overcomes both 12-MPI CPU-only ML potentials and GPU-accelerated ML potentials by achieving speedups of 168x and 22x, respectively. The proposed approach has a potential to provide a powerful way to develop high-performance and high-accuracy potentials even in the other systems.