Synergizing machine learning, molecular simulation and experiment to develop polymer membranes for solvent recovery

Abstract

Organic solvent nanofiltration (OSN) is a robust membrane technology for solvent recovery and molecular separation in harsh conditions. However, the current OSN membranes are largely produced through trial-and-error methods. In this study, machine learning (ML), molecular simulation (MS) and experiment are synergized for the development of OSN membranes. Using three different learning strategies, ML models are first constructed to identify critical gross properties (i.e., solvent viscosity, membrane thickness and water contact angle) and establish a phenomenological relationship for permeability prediction. Subsequently, ML models based on molecular representation via concatenated fragments are developed to predict methanol permeabilities in three polymer of intrinsic microporosity (PIM) membranes (PIM-A1, CX-PIM-A1 and PIM-8). The methanol permeability predicted in PIM-A1 is the highest among the three and also higher than that in archetypal PIM-1. Next, MS is conducted to provide microscopic insights into swelling behavior and methanol permeation in the three PIM membranes. Finally, the PIM-A1 membrane is experimentally fabricated and found to exhibit nearly complete solute rejection and methanol permeability of 2.33 × 10−6 L·m/m2·h·bar, which validates the ML prediction. This study demonstrates that the synergy of ML, MS and experiment can fundamentally elucidate and quantitatively predict solvent permeation in polymer membranes, and the holistic approach may advance the development of new membranes for solvent recovery and other important separation processes.