Multi-Fidelity Machine Learning Predictors for Gas Permeability Through Polymers: An Example of Experimental and Simulation Data Fusion

Polymer based gas separation membranes have a wide variety of applications ranging from carbon capture, drug delivery, and food packaging. Such technology, when extended to biodegradable or chemically recyclable polymers has the potential to advance green sustainable energy and applications. Polymer membrane discovery and design has been accelerated using machine learning (ML) prediction models trained on experimental gas transport property data. Predictions in the known scope of experimental chemical spaces are reputable; however, they are less reliable when extrapolating to new spaces. In this work, we use a high-throughput molecular dynamics (MD) simulation pipeline to explore new chemical spaces. This pipeline we built produces simulation data for gas diffusivity, solubility, and permeability. Subsequently, we create multi-task deep learning models trained on both “high-fidelity” experimental and “low-fidelity” simulation data. The new multi-task model shows a significantly improved performance over models trained with only experimental data, particularly in data-scarce situations. The algorithm ascertains the correlation between low- and high-fidelity data, and this basis allows for a more informed decision when making predictions in new chemical spaces. This approach demonstrates the benefit of high-throughput classical MD simulations with data fusion to produce best-in-class property predictors especially when experimental data is not as accessible for properties of interest.