Machine learning interatomic potentials with accurate long-range interactions for molecular dynamics collision simulations of atmospherically-relevant molecules

Recent advances in machine learning interatomic potentials have enabled the simulation of cluster formation from precursor vapor at a high level of theory. However, performing these simulations requires verifying that the models accurately describe cluster formation dynamics, particularly collision processes. In this work, we study the performance of two distinct machine learning (ML) architectures, AIMNet2 and PaiNN, against GFN1-xTB and ωB97X-3c reference data for atmospherically relevant collision systems (H₂SO₄–H₂SO₄, H₂SO₄–HSO₄^-, and H₂SO₄–NH(CH₃)₂).

We evaluate the models' ability to reproduce one-dimensional potentials of mean force (PMFs) and collision probabilities. Both models achieve excellent agreement with reference PMFs, yielding RMSEs at least an order of magnitude lower than chemical accuracy (1 kcal mol^-1). Notably, PaiNN achieves lower errors in the binding region.

However, we observe significant differences in collision probabilities. While AIMNet2 accurately reproduces these probabilities, PaiNN fails to capture long-range interactions beyond its local cutoff (10 Å). For the charged H₂SO₄–HSO₄^- system, this leads to a complete loss of collision probability beyond 14 Å and an underestimation at shorter distances.

Our results demonstrate a clear trade-off: while PaiNN offers superior accuracy for equilibrium properties and binding energies, its local nature makes it unsuitable for collision kinetics in systems with strong long-range interactions. Conversely, AIMNet2's ability to model these long-range interactions makes it the necessary choice for simulating collisions in such systems.