A 3d convolutional neural network autoencoder for predicting solvent configuration changes in condensed phase biomass reactions

Abstract

Configuration changes in the solvent or melt-phase (condensed phase reactions) molecules impact reaction thermodynamics and kinetics, making it vital to assess if solvent/melt-phase molecules need to be considered explicitly in first principles-based reactive molecular simulations. A basis for these configuration changes is established using MD simulations of melt-phase cellobiose decomposition. A 3d CNN autoencoder is trained to extract spatio-temporal features from coordinates of atomic positions in the MD trajectories of cellobiose decomposition. The differences between the encoded reactant and product features were fit to probability distributions, where larger configuration changes were found to be more probable at lower temperatures. The machine learning model then predicts changes in solvent orientation by using a distance-based classifier to assess the closeness between encoded features from reactant trajectories of cellobiose systems with larger configuration changes and those from the following systems: (i) fructose protonation in water–DMSO and, (ii) glucose isomerization via hydride transfer in water and methanol. The extent of solvent configuration changes in the fructose systems was predicted to increase with DMSO concentrations and was validated using trends in the difference between reaction free energies. For glucose isomerization, configurational changes in pure methanol were predicted to be higher than that in water, consistent with the high polarizability of methanol due to which the reaction free energy barrier is ∼50 kJ mol⁻¹ higher than that in water. This work demonstrates a machine learning framework that has the potential to limit the computational cost and accelerate the deployment of molecular simulations in screening solvents for reactive chemical transformations.