Machine learning for metallurgy IV: A neural network potential for Al-Cu-Mg and Al-Cu-Mg-Zn

Most metallurgical properties, e.g., dislocation propagation, precipitate formation, can only be fully understood atomistically but most phenomena and quantities of interest cannot measured experimentally. Accurate simulation methods are essential first-principles density functional theory (DFT) is prohibitively expensive while empirical interatomic potentials rarely sufficiently accurate for alloys. Machine learning (ML) emerging as an approach to create computationally-efficient atomistic achieving near-DFT accuracy. Building on recent work binary Al-Cu ternary Al-Mg-Si, here a family neural network (NNPs) Al alloys Al-Cu-Mg Al-Cu-Mg-Zn developed assessed using the Behler-Parinello formulation. Training uses robust set metallurgically-relevant structures including intermetallic phases, stacking faults, solute/solute solute/stacking fault interactions, solute clusters, matrix/precipitate interfaces. The accuracy these NNPs then demonstrated across comprehensive properties derived from training and, moreover, many important not represented in such generalized energy (GSFE) surface critical S-phase Al-Mg-Cu, antisite vacancy formation energies intermetallics. broader also have high subtle Cu substitutional ${\ensuremath{\eta}}^{\ensuremath{'}}$ T phases small Al-Zn-Mg clusters. Together with earlier results, this paper shows how increasingly complex multicomponent alloy systematically by expanding database, leading broad family, technological Al-2xxx, Al-5xxx, Al-7xxx