How Long Is Long Enough? Extrapolation of Machine-Learning Interatomic Potentials for Oligomeric and Polymeric...

Machine-learning interatomic potentials (MLIPs) have surged in popularity due to their promise of expanding the spatiotemporal scales possible for simulating molecules with high fidelity. The accuracy of any MLIP is dependent on the data used for its training; thus, for large molecules like polymers and biomolecules, where ab initio training data is prohibitively difficult to obtain, it becomes necessary to use smaller, analogous chemical systems to construct MLIPs. Here, we perform a control study using n = 1-8 linear alkanes to determine when MLIPs trained on small molecules can accurately extrapolate to longer chains and more complex architectures. By combining MLIP performance analysis with environment-resolved Smooth Overlap of Atomic Positions descriptors and Principal Covariates Classification, we show that reliable extrapolation emerges when the relevant local chemical environments have converged between the training and target systems. We further demonstrate that careful construction of the neighbor list substantially improves the learnability of intermolecular energetics, the component most critical for polymeric behavior. Together, these results provide a practical, data-driven blueprint for designing transferable MLIPs for macromolecular materials, whether built from bespoke training data or adapted from emerging universal MLIP frameworks.