Predicting the Solubility of Gases, Vapors, and Supercritical Fluids in Amorphous Polymers from Electron Density using Convolutional Neural Networks

Abstract

A twin convolutional neural network is proposed to predict the pressure and temperature-dependent sorption of gases, vapors, and supercritical fluids in amorphous polymers based solely on spatial electron density distribution. Quantum chemical data as 3D tensors (3D images) is derived from DFT calculations. A dataset of almost 15000 experimentally measured uptakes (0.01-50 wt%) of 79 gases in 102 different polymers under pressures from 1E-3 – 7E+2 bar range and temperatures from 233-508 K range is collected from over 250 literature sources. The dataset includes measurements on almost 500 solvent-polymer systems spanning from typical low-pressure sorption in membrane glassy polymers to high-pressure solubility of supercritical fluids in molten polymers. Irreducible mean absolute percentage error (MAPE) is approximately estimated to be ~20%. The sources of inherent data variability are briefly discussed. In 100 epochs, the model achieved 31% MAPE on a test set of 1600 experimental measurements concerning 22 polymers previously unseen by the model.