Learning Accurate and Transferable Force Fields for Physical Property Predictions of Organic Liquids

Machine learning interatomic potentials (MLIPs) have emerged as efficient surrogates for quantum mechanical calculations, offering substantial acceleration in ab-initio atomistic simulations. While most applications of MLIPs have focused on isolated molecules, extending their accuracy and transferability to predict condensed-phase properties of liquids remains a major challenge. In this work, we present a novel framework for constructing transferable and data-efficient force fields for organic liquids using a dual-space active learning (AL) strategy. This approach enables an efficient AL workflow across both configurational and chemical space by coupling a query-by-committee method with an explicitly constructed target chemical space, generated using computationally inexpensive classical methods. Our approach employs a Euclidean transformer architecture to train the Neural Network Potential for Liquid Simulations (NPLS). As a proof of concept, we target the complete alkane family and train NPLS using high-level DFT data. The resulting model is rigorously benchmarked against experimental measurements and the widely used classical force field, i.e., OPLS-AA, across thermodynamic, dynamic, and phase transition properties. Remarkably, the NPLS model demonstrates strong generalization to larger systems of polyolefins and accurately captures liquid-to-solid transitions. However, systematic deviations are observed, particularly in the predicted liquid densities. We attribute these discrepancies to the non-negligible influence of nuclear quantum fluctuations (NQF), which are not captured with classical molecular dynamics (MD) sampling. To address this, we perform path-integral (PI) MD simulations to incorporate NQF, predictions of which show significantly improved agreement with experimental measurements. The NLPS-PIMD results outperform the OPLS-AA force fields in accuracy. We believe this methodology provides a practical and extensible route for developing highly accurate and transferable bespoke or universal MLIPs for complex liquids.