(364g) Combining Molecular Dynamics Simulations and Active Learning to Study the Hydrophobicity of Chemically Heterogeneous Surfaces

The hydrophobicity of a material is a key property that influences behavior in aqueous environments. While hydrophobicity can be quantified experimentally and predicted computationally for small molecules and larger, chemically uniform surfaces, predicting the hydrophobicity of chemically heterogeneous surfaces – surfaces with polar and nonpolar groups spatially arranged in different patterns on nanometer length scales – remains an outstanding challenge. While classical approaches consider additive effects of various chemical groups on the hydrophobicity based on their area coverage, these relations often fail to capture trends in the hydrophobicity of chemically heterogeneous surfaces with the same mole fractions of different chemical groups but vastly different hydration responses. Probing chemically heterogeneous systems using experimental methods is also challenging due to the length scale of the heterogeneity (on the order of nanometers). Instead, computational methods like traditional molecular dynamics (MD) simulations and enhanced sampling techniques have made progress in understanding the effect of chemical heterogeneity on hydrophobicity. However, two major challenges in using computational approaches to explore the hydrophobicity of chemically heterogeneous surfaces are the large sampling times required for enhanced sampling techniques and the vast number of possible spatial patterns and combinations of polar/nonpolar groups possible in even small molecular systems.

In this work, we present a hierarchical active learning approach to efficiently study the effect of patterning of polar and nonpolar groups on the hydrophobicity of self-assembled monolayers (SAMs). We employ atomistic molecular dynamics simulations and indirect umbrella sampling (INDUS) to label each chemically heterogeneous SAM with a hydration free energy (HFE). We then combine INDUS simulations and 3D Convolutional Neural Networks (CNNs) in an active learning framework. The active learning framework uses a Gaussian Process Regression model to predict the HFE of a SAM and an associated uncertainty in the HFE prediction based on the pattern of polar and nonpolar groups embedded on the SAM. The active learning model combines the labels and their associated uncertainty into an acquisition function which is designed to efficiently explore regions of pattern space with high uncertainty and find patterns which show large deviations from ideal behavior. Using the hierarchical active learning approach, we explore the relationship between patterning and hydrophobicity and recreate an “envelope” of HFEs associated with various mole fractions of the polar component. We also identify patterns with large positive and negative deviations from a linear combination of HFEs at the given mole fraction. Comparison of the HFE envelope for different polar end groups further reveals how deviations from a linear combination of HFEs depend on the end group chemistry. Finally, we analyze structural metrics of interfacial water near the SAM surface to understand how chemical context plays a role in determining hydrophobicity and why some motifs on patterned SAMs are more hydrophobic than others.