(420g) Active Learning Enhanced Screening of CO2 Capture Materials

In this work, we have developed and demonstrated an end-to-end modeling framework for screening of solid sorbent materials for CO₂ capture. This framework is a wholistic consideration of all the scales from which the performance of a sorbent material arises, starting from molecular dynamics, adsorption, and diffusion phenomena to process related interactions such as gas composition and contamination. Each of these components is tied to structural motifs of the sorbent material with the aid of machine learning models in an active learning process. Multiple elements of unsupervised and supervised learning are folded into this approach where the clusters of data, aggregates of features, model structure and parameters are continuously learnt starting from a small set of diverse adsorbent structures. Molecular simulations and process models are directed to evaluate better performing, synthetically promising structures by a multi-objective screening procedure that also weighs in the variable uncertainty in machine learning (ML) model predictions across different structures. This stratified approach allows us to accelerate identification of promising candidate sorbents that satisfy multi-scale end-to-end criteria starting from a very limited dataset. The accuracy and confidence of the ML-models are also improved through each cycle. We believe this approach of incrementally blending in knowledge through an integrated active learning cycle can be applied successfully to a wide range of material discovery problems.