Abstract
<jats:p>Accurate modelling of chemical reactions is key to optimising synthetic procedures and molecular discovery. Yet doing so represents one of the grand challenges in computational chemistry, especially when considering solvent and non-statistical dynamic effects. Pre-trained machine-learned interatomic potentials (MLIPs) have emerged as a promising alternative to traditional quantum chemistry approaches, offering stable out-of-the-box performance across chemical domains. Recent studies have shown that fine-tuning these MLIPs on task-specific datasets can substantially improve their accuracy. Nonetheless, it remains unclear how the composition and size of fine-tuning datasets influence model performance in reaction modelling, and to what extent fine-tuning compensates for limitations in pre-training datasets. In this work, we address these questions, focusing on organic reactions in solution, using three Diels–Alder reactions of increasing complexity as benchmark systems. We evaluate three pre-trained MACE models, each covering a different domain: a bespoke model for graphene-oxide (GO-MACE-23) and two `foundation' models for organic molecules (MACE-OFF-23) and materials (MACE-MP0), alongside the recently released and substantially larger MACE-OMOL model. We demonstrate that models can be fine-tuned using 50 or fewer samples representing vacuum and solvated clusters to accurately reproduce potential energy barriers, transition state geometries, and reaction dynamics in solution. We also identify key factors in model pre-training data that affect fine-tuning performance, including differences in levels of theory and limited coverage of non-equilibrium structures. While MACE-OMOL delivers competitive out-of-the-box performance, its improvement upon fine-tuning is less substantial, and its size and memory requirements currently limit its use in long molecular dynamics simulations required for free energy sampling. Overall, this study presents a systematic analysis of MLIP fine-tuning for modelling organic reactions in solution, offering practical guidelines that facilitate their use and move these approaches closer to routine applications.</jats:p>