Comparison of Machine-Learning and Classical Force Fields in Simulating the Solvation of Small Organic Molecules in Acetonitrile

Machine learning force fields (MLFFs) have emerged as a new method for molecular simulation that combines the accuracy of ab initio approaches with the computational efficiency of classical force fields. However, the performance of MLFFs in describing the solvation configuration has yet to be explored. Here, we compare and contrast the performance of ANI-1ccx MLFF, the GAFF classical force field, and the ab initio molecular dynamics (AIMD) in simulating nine organic solutes in acetonitrile solvents. We examine the solvent-solute interaction described by these methods from four aspects: the solute conformation landscape, the solvation shell structure, the structure and dynamics of the O-H⋯N hydrogen bond, and the dynamics of the first solvation shell. For solute conformation description, ANI-1ccx and GAFF both yield minima that agree with density functional theory optimization for rigid solutes. However, their results diverge for flexible solutes. For solvation shell structure description, ANI-1ccx agrees better with AIMD on the location of the first solvent shell than GAFF does. For the description of the O-H⋯N hydrogen bond formed between acetonitrile and the solute, ANI-1ccx generates stronger hydrogen bonds with shorter bond lengths, wider bond angles, and longer hydrogen bond lifetimes, agreeing better with DFT-optimized structure. ANI-1ccx also describes a more frequent exchange of acetonitrile molecules in and out of the first solvation shell than GAFF. Our study demonstrates the potential benefits of utilizing MLFF for simulating solution-phase dynamics and generating solvation configurations.