Fluid density regulation under nanoconfinement revealed by machine learning.

Fluids under nanoscale confinement differ - and often dramatically - from their bulk counterparts. A notorious feature of nanoconfined fluids is their inhomogeneous density profile along the confining dimension, which plays a key role in many fluid structural and transport phenomena in nanopores. Nearly five decades of theoretical efforts on predicting this phenomenon (fluid layering) have shown that its complexity resists purely analytical treatments; as a consequence, nearly all current approaches make extensive use of molecular simulations, and tend not to have generalizable predictive capabilities. In this work, we demonstrate that machine-learning-based models (in particular, a random forest model), trained upon large molecular simulation data sets, can serve as reliable surrogates in lieu of further molecular simulation. We show that this random forest model has excellent interpolative capabilities over a wide range of temperatures and confining lengthscales, and even has modest extrapolative ability. These results provide a promising pathway forward for developing models of nanoconfined fluid properties that are generalizable, lower cost than "pure" molecular simulation, and sufficiently predictive for fluids-in-nanopores practitioners.