Accuracy, Reliability, and Efficiency of a MACE Machine-Learning Interatomic Potential for Aspirin: A Reproducible Benchmark for Molecular Simulation

Machine-learning interatomic potentials (MLIPs) can reproduce density functional theory (DFT) energies and forces at a small fraction of the computational cost, yet an average benchmark error is an incomplete measure of their reliability in molecular simulation. This study presents a reproducible workflow that assesses prediction accuracy, error heterogeneity, and dynamical stability together, using aspirin (C9H8O4), a widely used pharmaceutical compound, as an initial molecular benchmark en route to metal-containing systems. A MACE equivariant graph neural network potential is trained on the MD17 dataset and deployed for molecular dynamics through the Atomic Simulation Environment (ASE). Across 210 762 held-out configurations the potential attains a force root-mean-square error of 15.30 meV Å−1, within the published MACE benchmark range, with a validation-to-test generalization gap of 7.6%. Force error correlates positively with configuration energy (Pearson r = 0.51), the carboxyl carbon emerges as the least reliable atomic site, and 90% of force predictions lie within 23.5 meV Å−1 of the DFT reference. Microcanonical simulations conserve total energy and canonical trajectories remain stable, while single-configuration evaluation is approximately four orders of magnitude (a roughly ten-thousand-fold) faster than the corresponding DFT calculation. By combining aggregate accuracy with configuration-and atom-resolved error analysis and explicit dynamical tests, the workflow exposes behavior that a single error metric conceals. The resulting aspirin benchmark establishes a documented baseline for subsequent validation on metal-containing systems and illustrates the value of assessing error distributions and dynamical behavior alongside mean prediction accuracy.