Batch active learning for accelerating the development of interatomic potentials

Classical molecular dynamics (MD) has been widely used to study atomistic mechanisms and emergent behavior in materials at length time scales beyond the capabilities of first-principles approaches. The success classical MD simulations relies on ability interatomic potentials accurately map complex many-body interacting systems electrons nuclei into effective few-body atoms. In practice, development is a nontrivial process requires considerable amount effort. Recently, machine learning become promising approach accelerate potential development. However, these approaches are often computation data intense, as they require large training from calculations, such total energies, atomic forces, stress tensors many structures. Here we propose an active combined with theory calculations expedite potentials. particular, develop batch method which combines both energy uncertainty structure similarity metrics efficiently sample highly uncertain structures that difficult predict. This sampling maximizes utility dataset each generates accurate robust model coefficients achieve conventional To demonstrate this method, for monolayer GeSe, two-dimensional ferroelectric-ferroelastic material, compare quality robustness obtained random sampling. Batch opens up avenues accelerating using small set will be valuable computational materials, physics, chemistry community. • developed Active diverse configurations reduces required structural space. consistent, reliable,