Molecular dynamics (MD) simulations are a powerful tool for analyzing the thermodynamics of complex solvent environments. However, MD simulations can be computationally expensive, and the analysis of simulation output requires human insight to relate simulation observables to relevant macroscopic properties. Alternatively, machine learning techniques are promising methods to rapidly interpret the data output from molecular simulations. In this talk, I will discuss our efforts to combine MD simulations and machine learning to study solvent effects relevant to liquid-phase reactions and interfacial properties. I will show that a convolutional neural network (CNN), a machine learning model developed to analyze images, can infer spatial correlations between solvent molecules that can be mapped to thermodynamic observables. I will discuss the use of classical MD simulations to predict the rates of acid-catalyzed biomass conversion reactions in mixtures of water and polar aprotic cosolvents. Unbiased MD simulations show that human-derived descriptors quantifying the enrichment of water near hydrophilic reactants correlate with reactant rates, but a trained 3D CNN produces more accurate predictions that are transferrable across a range of solvent compositions. I will then discuss how MD can be used to quantify the hydrophobicity of functionalized interfaces in good agreement with experiments. We further train a CNN to predict interfacial hydration free energies based on water molecule positions, facilitating the screening of materials hydrophobicity. This work demonstrates how data-centric methods can complement MD simulations for the analysis of liquid-phase systems.