Thermodynamics of Molecular Binding and Clustering in the Atmosphere Revealed through Conventional and ML-Enhanced Umbrella Sampling

TY - JOUR

T1 - Thermodynamics of Molecular Binding and Clustering in the Atmosphere Revealed through Conventional and ML-Enhanced Umbrella Sampling

AU - Kubečka, Jakub

AU - Knattrup, Yosef

AU - Trolle, Georg Baadsgaard

AU - Reischl, Bernhard

AU - Lykke-Møller, August Smart

AU - Elm, Jonas

AU - Neefjes, Ivo

N1 - © 2025 The Authors. Published by American Chemical Society.

PY - 2025/9/2

Y1 - 2025/9/2

N2 - Accurately modeling the binding free energies associated with molecular cluster formation is critical for understanding atmospheric new particle formation. Conventional quantum-chemistry methods, however, often struggle to describe thermodynamic contributions, particularly in systems exhibiting significant anharmonicity and configurational complexity. We employed umbrella sampling, an enhanced-sampling molecular dynamics technique, to compute Gibbs binding free energies for clusters formed from a diverse set of new particle formation precursors, including sulfuric acid, ammonia, dimethylamine, and water. By performing umbrella sampling along the evaporation coordinate, using forces computed at the semiempirical GFN1-xTB level of theory, we effectively capture entropic effects such as vibrational anharmonicities and transitions between different configurational minima, while avoiding errors from symmetry overcounting. In addition, we explored machine-learning-enhanced umbrella sampling simulations using neural network potentials trained on higher-level quantum chemistry data, demonstrating the feasibility of this approach for improving accuracy while maintaining computational efficiency. Our results show improved agreement with experimental values compared to conventional methods. We also present examples of gas-to-particle uptake processes, providing insights into cluster and aerosol-surface chemistry using first-principles approaches rather than commonly used molecular-mechanics force fields. This study demonstrates the importance of accounting for dynamics in predicting molecular binding thermodynamics in complex environments and highlights the potential of combining physics-based simulations with machine learning for reliable and scalable predictions.

AB - Accurately modeling the binding free energies associated with molecular cluster formation is critical for understanding atmospheric new particle formation. Conventional quantum-chemistry methods, however, often struggle to describe thermodynamic contributions, particularly in systems exhibiting significant anharmonicity and configurational complexity. We employed umbrella sampling, an enhanced-sampling molecular dynamics technique, to compute Gibbs binding free energies for clusters formed from a diverse set of new particle formation precursors, including sulfuric acid, ammonia, dimethylamine, and water. By performing umbrella sampling along the evaporation coordinate, using forces computed at the semiempirical GFN1-xTB level of theory, we effectively capture entropic effects such as vibrational anharmonicities and transitions between different configurational minima, while avoiding errors from symmetry overcounting. In addition, we explored machine-learning-enhanced umbrella sampling simulations using neural network potentials trained on higher-level quantum chemistry data, demonstrating the feasibility of this approach for improving accuracy while maintaining computational efficiency. Our results show improved agreement with experimental values compared to conventional methods. We also present examples of gas-to-particle uptake processes, providing insights into cluster and aerosol-surface chemistry using first-principles approaches rather than commonly used molecular-mechanics force fields. This study demonstrates the importance of accounting for dynamics in predicting molecular binding thermodynamics in complex environments and highlights the potential of combining physics-based simulations with machine learning for reliable and scalable predictions.

UR - https://www.scopus.com/pages/publications/105014804130

U2 - 10.1021/acsomega.5c05634

DO - 10.1021/acsomega.5c05634

M3 - Journal article

C2 - 40918357

SN - 2470-1343

VL - 10

SP - 39148

EP - 39161

JO - ACS Omega

JF - ACS Omega

IS - 34

ER -