Speaker
Description
Modern ML force fields approach the accuracy of Density Functional Theory (DFT) calculations for the properties of organic molecules and their reactions.[1,2] The vast majority of them do not model the effects of the environment, specifically the effect of the solvent.[3] However, correctly modeling interaction of molecules with the solvent is crucial for modeling biological systems and organic reactivity. Therefore, we set out to explore whether we could model these effects with electrostatics-based implicit solvation models, such as the COSMO model.[4] To that end, we reimplemented a linear-scaling variant of this model (ddCOSMO[5]) in the PyTorch framework. Also, to model non-electrostatic effects, we decided to reimplement the OpenCOSMO-RS solvation model[4,6] in a similar fashion. These can be then coupled to models that produce electrostatic potentials, such as MACE-POLAR[7] or MACE-MDP.[8] We will discuss the performance of the resulting hybrid models and outline possible improvements.
References
[1] Yuan, E. C. -.; Liu, Y.; Chen, J.; Zhong, P.; Raja, S.; Kreiman, T.; Vargas, S.; Xu, W.; Head-Gordon, M.; Yang, C.; Blau, S. M.; Cheng, B.; Krishnapriyan, A.; Head-Gordon, T. Foundation Models for Atomistic Simulation of Chemistry and Materials. arXiv 2025. doi:10.48550/arXiv.2503.10538
[2] Kalita, B.; Gokcan, H.; Isayev, O. Machine learning interatomic potentials at the centennial crossroads of quantum mechanics. Nat. Comput. Sci. 2025, 5, 1120-1132. doi:10.1038/s43588-025-00930-6
[3] Zhang, L.; Zavadlav, J. ConSolv: Solvent-Conditional Machine Learning Implicit Solvent Potential. arXiv 2026. doi:10.48550/arXiv.2606.24983
[4] Klamt, A. The COSMO and COSMO‐RS solvation models. WIREs Comput. Mol. Sci. 2018, 8, e1338. doi:10.1002/wcms.1338
[5] Stamm, B.; Lagardère, L.; Scalmani, G.; Gatto, P.; Cancès, E.; Piquemal, J.; Maday, Y.; Mennucci, B.; Lipparini, F. How to make continuum solvation incredibly fast in a few simple steps: A practical guide to the domain decomposition paradigm for the conductor‐like screening model. Int. J. Quantum Chem. 2019, 119, e25669. doi:10.1002/qua.25669
[6] Müller, S.; Nevolianis, T.; Garcia-Ratés, M.; Riplinger, C.; Leonhard, K.; Smirnova, I. Predicting solvation free energies for neutral molecules in any solvent with openCOSMO-RS. Fluid Phase Equilib. 2025, 589, 114250. doi:10.1016/j.fluid.2024.114250
[7] Batatia, I.; Baldwin, W. J.; Kuryla, D.; Hart, J.; Kasoar, E.; Elena, A. M.; Moore, H.; Gawkowski, M. J.; Shi, B. X.; Kapil, V.; Kourtis, P.; Magdău, I.; Csányi, G. MACE-POLAR-1: A Polarisable Electrostatic Foundation Model for Molecular Chemistry. arXiv 2026. doi:10.48550/arXiv.2602.19411
[8] Gönnheimer, N.; Reuter, K.; Kapil, V.; Margraf, J. T. MACE-MDP: A General Dipole and Polarizability Model for Organic Molecules and Materials. ChemRxiv 2026. doi:10.26434/chemrxiv.15000716/v2