MACHINE LEARNING POTENTIALS FOR MOLECULAR THERMODYNAMICS

Abstract

Machine-learning interatomic potentials represent a turning point in molecular simulation, promising to extend the accuracy of quantum chemistry into the collective and time- dependent domain of liquids. This study has tested that promise by employing the MACE potential to explore the thermodynamic behaviour of liquid 1-pentanol at 300 K. Through simulations of systems containing 128 and 1024 molecules in both NVT and NPT ensembles, a range of bulk properties were calculated, including density, isothermal compressibility, heat capacities, and enthalpy of vaporisation, together with structural descriptors derived from radial distribution functions. The results disclose a pattern of both achievement and limitation. Compressibility, when evaluated in sufficiently large systems, converges to experimental values, attesting to the capacity of MACE to capture long-wavelength fluctuations. Yet other observables - most notably density, heat capacities, and enthalpy of vaporisation - are systematically displaced from experiment, reflecting an over-structured hydrogen-bond network and the neglect of quantum effects in classical dynamics. The computational performance of MACE, while tractable, remains slower than traditional force fields, underscoring the familiar trade-off between accuracy and efficiency. Taken together, these findings show that MACE, though not yet a complete substitute for experiment or for carefully parameterised classical models, nonetheless marks a significant advance in the simulation of molecular liquids. By offering a systematic evaluation of a machine-learning potential for a linear alcohol, this work provides both a benchmark of present capabilities and a guide to the methodological and conceptual refinements required for the next generation of molecular thermodynamics.