Machine Learning Force Fields Data Collection Docs ========================================================== Project overview ---------------------------------------------------------- Machine learning force fields (MLFFs) can accelerate atomic simulations by serving as computationally efficient surrogates for density functional theory (DFT). Their usefulness, however, depends on whether they remain reliable when atomic structures move beyond familiar training configurations. This project asks: **When an atomic structure is pushed away from equilibrium, can the model recover its original structure?** The study applies controlled adversarial perturbations to 20 two-dimensional Materials Project structures and 20 Li-C-O configurations. Recovery is evaluated across CHGNet, MACE, MACE-MH, UMA, and MTP using forces, atomic displacements, relaxation convergence, symmetry, and structural topology. FGSM, I-FGSM, and PGD are compared with non-adversarial contour exploration. The experiments also examine perturbation magnitude, numerical precision, random seeds, and supercells. What we found ------------- Perturbations approaching 5% of the minimum lattice parameter increasingly prevent structures from recovering their pre-attack configurations. At this magnitude, the largest post-relaxation force changes observed for the LiCOHPF-trained MACE and MTP models reach approximately five orders of magnitude above those observed for the pretrained general-purpose models. This work provides a reproducible framework for testing MLFF reliability inside an iterative physical workflow, where every force prediction affects the next atomic configuration. `View the project repository `_ .. toctree:: :maxdepth: 2 :hidden: Motivation Concepts Experiment Design Getting Started Equations Outputs Plot Methods Results Findings References