CC BY 4.0Strickstrock, RobinRobinStrickstrockHagg, AlexanderAlexanderHaggReith, DirkDirkReithKirschner, Karl N.Karl N.Kirschner2025-09-152025-09-152025https://publica.fraunhofer.de/handle/publica/495200https://doi.org/10.24406/publica-541310.1002/cphc.20250035310.24406/publica-54132-s2.0-105014915761Molecular modeling plays a vital role in many scientific fields, ranging from material science to drug design. To predict and investigate the properties of those systems, a suitable force field (FF) is required. Improving the accuracy or expanding the applicability of the FFs is an ongoing process, referred to as force-field parameter (FFParam) optimization. In recent years, data-driven machine learning (ML) algorithms have become increasingly relevant in computational sciences and elevated the capability of many molecular modeling methods. Herein, time-consuming molecular dynamic simulations, used during a multiscale FFParam optimization, are substituted by a ML surrogate model to speed-up the optimization process. Subject to this multiscale optimization are the Lennard–Jones parameters for carbon and hydrogen that are used to reproduce the target properties: n-octane's relative conformational energies and its bulk-phase density. By substituting the most time-consuming element of this optimization, the required time is reduced by a factor of ≈20, while retaining FFs with similar quality. Furthermore, the workflow used to obtain the surrogate model (i.e., training data acquisition, data preparation, model selection, and training) for such substitution is presented.entrueforce-field optimizationgradient-based optimizationLennard–Jones parametersmachine learningneural networksjournal article