Validation of moment tensor potentials for fcc and bcc metals using EXAFS spectra

Abstract

Machine-learning potentials for materials, namely the moment tensor potentials (MTPs), were validated using experimental EXAFS spectra for the first time. The MTPs for four metals (bcc W and Mo, fcc Cu and Ni) were obtained by the active learning algorithm of fitting to the results of the calculations using density functional theory (DFT). The MTP accuracy was assessed by comparing metal K-edge EXAFS spectra obtained experimentally and computed from the results of molecular dynamics (MD) simulations. The sensitivity of the method to various aspects of the MD and DFT models was demonstrated using Ni as an example. Good agreement was found for W, Mo and Cu using the recommended PAW pseudopotentials, whereas a more accurate pseudopotential with 18 valence electrons was required for Ni to achieve a similar agreement. The use of EXAFS spectra allows one to estimate the MTP ability in reproducing both average and dynamic atomic structures.

References in corpus (5)

• On-the-fly machine learning force field generation: Application to melting points
• An Accurate and Transferable Machine Learning Potential for Carbon
• Elinvar effect in $\beta -$Ti simulated by on-the-fly trained moment tensor potential
• Ab initio molecular dynamics simulations of negative thermal expansion in ScF3: the effect of the supercell size
• X-ray absorption and Raman spectroscopy studies of tungstates solid solutions ZncNi1-cWO4 (c=0.0-1.0)