Machine-learned force model simulates silicon-oxygen structures from glass surfaces to monoxide grains

The work

Technical · from the record

The authors built a machine-learning interatomic potential for the full binary silicon–oxygen system, using an active-learning loop in which high-uncertainty atomic environments found in large-scale simulations were cut out, embedded in an amorphous matrix, and labelled with SCAN density-functional theory. The final database contains 11,428 structures with a total of about 1.3 million atoms, and the fitted atomic cluster expansion model reaches test-set errors of 16.7 meV per atom for energies and 306 meV per ångström for forces. Simulations with the potential covered high-pressure silica polymorphs and compressed amorphous silica, quartz and amorphous silica surfaces, porous structures, and melt-quenched models of amorphous silicon monoxide, which segregated into silicon-rich and silica-like regions with estimated average grain diameters between 24 and 54 ångström, compared with 30–40 ångström from published transmission electron microscopy. Heating the silicon monoxide models to 1400 K and quenching to 1200 K produced crystallisation in the silicon-rich regions.

How AI was used

Two families of machine-learned interatomic potentials were fitted to density-functional-theory energies and forces. During database construction, moment tensor potentials implemented in the MLIP package were fitted to the current database and then used to drive molecular dynamics in three separate tracks (high-pressure silica, silica surfaces, non-stoichiometric SiOx); structures exceeding an extrapolation threshold were selected, labelled with SCAN DFT in VASP, and added to the database, and the cycle was repeated until the threshold was no longer exceeded. In a further stage, committees of two to four moment tensor potentials trained on the same database gave per-atom committee errors in large-scale simulations; environments of high-uncertainty atoms were cut into DFT-sized cubes, the atoms inside the potential cut-off were held fixed, and the surrounding region was melted in a machine-learning molecular-dynamics run to form an amorphous matrix before DFT labelling. The merged database was then used to fit the final potential in the nonlinear atomic cluster expansion framework with PACEMAKER, using 600 basis functions, 5700 parameters, Bessel radial functions, a force-to-energy weight ratio of 0.01 and 2000 BFGS steps; linear and Finnis–Sinclair-like variants were fitted to the same data for comparison. The final potential then drove large-scale molecular dynamics and static calculations in LAMMPS, with ASE and the OVITO Python interface used for protocols and analysis, for melt-quenching, compression, aerogel densification, surface relaxation, thermodynamic integration in calphy and thermal treatment of the silicon monoxide models.