Machine learning driven advances in molecular dynamics of aqueous, interfacial systems

Molecular dynamics (MD) simulations have been widely applied to investigate various physical and chemical processes in aqueous and interfacial environments, which are crucial for the design of energy materials and for understanding several chemical processes at the heart of life itself. However, the applicability of MD simulations has been constrained by several inherent challenges, including the accuracy of force fields, limitations in simulation size and timescales. One promising solution to these challenges is the integration of machine learning (ML) methods, both for improved description of the nature of interactions in aqueous systems as well as for enhanced sampling. In this review, we discuss the principles, implementation, and applications of ML force fields (MLFFs) and ML enhanced sampling methods to the study of aqueous, interfacial systems. We discuss five key categories of applications that use MLFFs, ML-enhanced sampling, and ML-driven data analytics. We first discuss how MLFFs are enabling quantum level accuracy at classical level cost for large scale simulations of complex aqueous and interfacial systems, and then highlight how coupling them with enhanced sampling and advanced data analytics, especially graph based approaches for featurizing such systems, can be used both for enhancing simulations and for understanding them by yielding reliable low dimensional reaction coordinates that improve the interpretation of high dimensional MD data. The discussed applications include investigations into the structure and dynamics of bulk water and aqueous interfaces, proton transfer, catalysis, phase transitions, and the prediction of vibrational spectra. In each case, we highlight how ML-based methods enable simulations that were previously computationally prohibitive and provide new physical insights into aqueous solutions and interfaces. For instance, MLFFs allow nanosecond-scale simulations with thousands of atoms while maintaining quantum chemistry accuracy. Additionally, ML-enhanced sampling facilitates the crossing of large reaction barriers and enables the exploration of extensive configuration spaces. Moreover, ML models trained on simulation data uncover previously overlooked factors, such as the role of solvent dynamics in phase transitions. The combination of MLFFs with enhanced sampling techniques makes the calculation of high-dimensional free energy surfaces feasible, significantly improving our understanding of chemical reactions. Finally, we discuss the current challenges in this field and outline potential future research directions to further advance the integration of ML in MD simulations.