MLIP Toward Complex Fluid Systems: From Training Strategies to Reaction Dynamics in Aqueous and Supercritical Environments

Machine learning interatomic potentials (MLIPs) bridge first-principles accuracy with classical MD scalability for atomistic study of complex fluids. We present a unified perspective spanning training strategies for transferable MLIPs to physicochemical phenomena in aqueous and supercritical environments.

We first show that diversifying density in training data is far more effective than temperature diversification for transferable MLIPs, as density variation alters local coordination topology more efficiently [1]. A two-phase active learning framework yields a deep potential for moisture-induced degradation in tin-based metal-oxide resists, revealing a facile dimerization pathway (~13.2 kJ/mol) with atomistic insight into EUV photoresist stability [2].

We then demonstrate MLIPs capturing complex water dynamics: a SCAN-trained model reproduces the dielectric spectrum including high-frequency H-bond resonances, with many-body interactions governing relaxation [3]. DPMD uniquely reproduces anomalous diffusion acceleration by chaotropic salts [4], and for aqueous AlCl₃, uncovers a hydration pathway where the sixth water penetrates from beyond the second shell via cooperative deshielding [5].

Finally, comparing NequIP and ReaxFF for acetic acid oxidation in supercritical water, NequIP more faithfully reproduces radical pathways and complete oxidation [6]. These studies establish MLIPs as a versatile framework for structure, dynamics, and reactivity of complex fluids