Ushering in a new era of atomic potentials with engineering accuracy
Atomic potentials have long been used as the first of several bridges between the first principles and engineering length scales. Traditional atomic potentials (EAM, MEAM, ADP, COMB, ReaxFF, etc.) have enabled the identification of the key physical processes involving the interactions between point, line and extended crystalline defects, which dictate the plastic deformation, fracture and other important properties of materials.
However, traditional atomic potentials, developed based on limited experimental data, generally have lacked the predictive accuracy required to quantitatively inform engineering models of material and component behavior.
The AMSET Project, a three-year multinational academic, commercial and government collaboration, endeavors to create a versatile development tool, integrated with LAMMPS within the Materials Design, Inc. MedeA software user environment, for creating machine learning atomic potentials (MLPs) with dramatically increased quantitative reliability. The goal of AMSET is to automate the development of new, high-quality MLPs, and to make this crucial capability directly available to engineering development organizations through MedeA.
AMSET is directed at providing tools for generating large training sets for MLP development using optimized DFT and selected beyond-DFT methods, and for bringing to bear the latest, state-of-the-art methods for MLP optimization, all in automated fashion.
AMSET development partners include Materials Design, Inc.; the Naval Nuclear Laboratory (the project sponsor), operated for the US Naval Reactors Program; Prof. Georg Kresse, a principal architect of the VASP code, at the University of Vienna; Prof. Bill Curtin at EPFL Lausanne; Prof. Gus Hart at Brigham Young University (Provo, Utah); Prof. Chris Wolverton at Northwestern University (Evanston, Illinois); and Prof. Jeffrey Grossman at MIT (Cambridge, Massachusetts).