Application of transfer learning to predict diffusion properties in metal-organic frameworks

Abstract

Transfer learning (TL) facilitates the way in which a model can learn well with small amounts of data by sharing the knowledge from a pre-trained model with relatively large data. In this work, we applied TL to demonstrate whether the knowledge gained from methane adsorption properties can improve a model that predicts the methane diffusion properties within metal-organic frameworks (MOFs). Because there is a large discrepancy in computational costs between the Monte Carlo (MC) and molecular dynamics (MD) simulations for gas molecules in MOFs, relatively cheap MC simulations were leveraged in helping to predict the diffusion properties and we demonstrate performance improvement with this method. Furthermore, we conducted a feature importance analysis to identify how the knowledge from the source task can enhance the model for the target task, which can elucidate the process and help choose the optimal source target to be used in the TL process.