(356b) Combining Molecular Dynamics and Machine Learning to Predict Interfacial Properties

Authors: 
Kelkar, A., University of Wisconsin-Madison
Dallin, B. C., University of Wisconsin-Madison
Van Lehn, R. C., University of Wisconsin-Madison
Molecular dynamics (MD) simulations are a powerful tool for analyzing the properties and behavior of materials in solvent environments by explicitly representing atomistic degrees of freedom and accounting for intermolecular interactions. However, MD simulations can require lengthy runtimes that limit the high-throughput screening of system properties. Moreover, the analysis of MD simulation output typically requires the implementation of techniques that rely on human insight to relate the properties of a statistical ensemble to relevant macroscopic properties. Alternatively, machine learning has emerged as a valuable tool to interpret molecular simulation data. Here, we present computational techniques that use machine learning to analyze molecular simulations of complex solvent environments. We focus on quantifying the hydrophobicity of functionalized interfaces by utilizing deep learning to analyze MD-derived interfacial water configurations. By selecting an appropriate representation of the MD output data, we demonstrate that a 2-dimensional convolutional neural network (2D CNN) can infer spatial correlations between solvent molecules and map these features to thermodynamic observables. We systematically vary the representation of the MD input data to determine the impact on the accuracy of the ML predictions. Once trained, we find that the 2D CNN can predict dewetting free energies with minimal simulation input, facilitating the rapid screening of interfacial hydrophobicity for a range of surface compositions. We further compare the accuracy of alternative ML architectures, including 3D CNNs that incorporate temporal correlations in our deep learning framework and graph neural networks that incorporate the network connectivity of water hydrogen bonds. Together, these methodologies demonstrate that ML analysis of spatial and temporal correlations that are encoded within the solvent environment can rapidly predict interfacial hydrophobicity to guide materials design.