Neural Network-Accelerated Molecular Dynamics Simulations for Next-Generation Energy Materials Discovery: An Integrated AI-Physics Framework

Silicon anodes promise exceptional energy density for lithium-ion batteries but suffer from capacity degradation mechanisms requiring atomic-scale understanding across compositional and structural complexity that exceeds ab initio simulation capabilities. An integrated computational framework combining E(3)-equivariant neural network architecture respecting rotational symmetries with uncertainty-driven active learning protocols achieves quantum mechanical accuracy (5.2 meV/atom energy, 82 meV/Å force) using 6,500 strategically selected training configurations spanning crystalline, amorphous, and liquid phases across the complete compositional range. Nanosecond molecular dynamics simulations spanning pristine silicon to fully lithiated Li22Si5 reveal three-stage volume expansion with inflection points at Li/Si ratios of 0.6 and 2.0, amorphization onset at x ≈ 1.4–1.6 matching synchrotron X-ray diffraction observations, and lithium diffusion coefficients spanning 3 orders of magnitude from 10–10 cm2/s to 10–7 cm2/s with activation energies decreasing systematically from 0.38 to 0.22 eV as frameworks progressively lose rigidity. Predicted densities, voltage profiles, and transport properties exhibit quantitative agreement with experimental measurements within typical uncertainties, establishing a generalizable paradigm for accelerating computational materials discovery in energy storage applications where multiphase coexistence renders empirical potentials inadequate.