Automated discovery of a robust interatomic potential for aluminum

Accuracy of molecular dynamics simulations depends crucially on the interatomic potential used to generate forces. The gold standard would be first-principles quantum mechanics (QM) calculations, but these become prohibitively expensive at large simulation scales. Machine learning (ML) based potentials aim for faithful emulation QM drastically reduced computational cost. accuracy and robustness an ML is primarily limited by quality diversity training dataset. Using principles active (AL), we present a highly automated approach dataset construction. strategy use under development sample new atomic configurations and, whenever configuration reached which uncertainty sufficiently large, collect data. Here, seek push limits automation, removing as much expert knowledge from AL process possible. All sampling performed using MD starting initially disordered configuration, undergoing non-equilibrium driven time-varying applied temperatures. We demonstrate this building aluminum (ANI-Al). After many iterations, ANI-Al teaches itself predict properties like radial distribution function in melt, liquid-solid coexistence curve, crystal such defect energies barriers. To transferability, perform 1.3M atom shock simulation, show that predictions agree very well with DFT calculations local environments sampled nonequilibrium dynamics. Interestingly, appearing appear have been dataset, way illustrate visually.