(191a) Energy Fingerprints for Machine Learning Prediction of Adsorption in Nanoporous Materials

The ability of nanoporous materials (NPMs) to adsorb various molecules of interest makes them promising candidates for applications in energy storage, chemical separation, sensing, catalysis, and drug delivery. However, finding the top-performing NPMs for a given application is impeded by the exponentially increasing chemical and structural complexities in the materials space. While molecular simulations can be applied to evaluate the adsorption properties under relevant conditions, they are still considered the bottleneck in discovering the best candidate materials, given thousands or more potential materials to be evaluated.

In this talk, I will first discuss our recent development of a new physics-inspired machine learning (ML) model based on two-dimensional (2D) energy histogram features, which are obtained from the probe-adsorbent energies and energy gradients given the material’s atomic structure and a suitable force field. The 2D energy histogram features encode both energetic and structural information of the material, and they lead to highly accurate ML models for predicting single-component adsorption of both spherical and chain molecules in various crystalline and amorphous NPMs. I will also show that it is possible to extract energy fingerprints for ML given an experimental adsorption isotherm for a material. This approach allows us to rapidly evaluate the separation capability of synthesized NPMs without knowing their atomic structures and accurate force fields. We will demonstrate the use of the method for adsorptive separation of binary mixtures and show how the ML model leads us to design rules for better NPMs.