Classical DFT energies and equilibrium configurations obtained by GNN-driven Monte Carlo simulations for solvent-free polymer-grafted nanoparticles

Abstract

This deposit contains the configuration–energy datasets and equilibrium configurations supporting the manuscript "Probing the many-body energy landscape of a soft glass with graph neural networks." We train an equivariant graph neural network (NequIP) on classical density functional theory energies of solvent-free polymer-grafted nanoparticles, a model soft glass in which grafted polymers uniformly fill the interstitial space and generate strong angular-dependent many-body interactions between the cores. The training configurations were sampled from hard-sphere dynamics without any energy bias, and reproduce none of the measured structural signatures of the equilibrium states. The trained network nonetheless recovers equilibrium configurations, validated against single-point classical DFT and against small-angle X-ray scattering measurements. Contents: out-of-equilibrium configurations with classical DFT energies (5,000 per sample, five sets of design parameters); equilibrium configurations from GNN-driven Monte Carlo with both GNN and classical DFT energies (2,000 per sample). All configurations are in extended XYZ format. Energies are total system energies in units of k_B T; lengths are in units of the core diameter. See README.md for full details.