Enhancement of classical force field predictions of hydrogen bonding in water, methanol and their crossed dimers using machine learning

• Published in: Molecular simulation Vol. 51; no. 7; pp. 478 - 492
• Main Authors: Obeid, Emil, Topcu, Ahmet E., Khader, Lina, Murshid, Nimer, Abu Samha, Mahmoud
• Format: Journal Article
• Language: English
• Published: Taylor & Francis 03.05.2025
• Subjects: hydrogen bonds strength; Machine-learning (ML) algorithms; machine-learning force fields (MLFFs) energies; methanol dimers and Hydrogen life-span; water
• ISSN: 0892-7022; 1029-0435
• DOI: 10.1080/08927022.2025.2501382

In this study, we use machine-learning (ML) algorithms as implemented in python 3.0 to study hydrogen bonding in water, methanol and water:methanol dimers. Our goals are to use ML to improve the performance of classical force fields at predicting H-bond interaction energies, and applying the resulting machine-learning force fields (MLFFs) to assess the H-bonds strength based on combined geometric criteria and MLFF dimer energies. Compared to DFT calculations using the WB97X-D/6-311G++(d,p) level, the MLFF models enhance the prediction accuracy for water, methanol, and water:methanol crossed dimers, with notable increases in the coefficient of determination ( $R^{2}$ R 2 ), averaging a 24% improvement across the three studied dimers. Amongst the studied ML algorithms, Gradient Boosting Classifier (GBC) is the most effective in classifying bond strengths for all studied dimers. MLFFs reduce the overestimation of strong bonds compared to classical methods (reducing the average overestimation from 10.5% to 5.2%, across the three systems). Our approach is more accurate than traditional force field methods and provides a basis for studying larger clusters and non-additive effects in hydrogen-bonded systems.