CDS&E: Transferable machine learning (ML) potentials to enable ML-based screening of adsorbents for separations involving chemisorption

About This Grant

Separating a mixture of many chemical species into its constituents is a common operation in the chemical industry. One separation method is to adsorb molecules of the desired species onto the surface of an adsorbent. Later, the same molecules can be desorbed and recovered. The efficiency of the separation could be improved by modifying the adsorbent for each desired separation. Promising adsorbents can be identified by screening vast “libraries” of possible candidates using a computer. However there is a major hurdle: the attractive forces between the desired molecule and each candidate adsorbent in the library are unknown. This project will develop machine learning tools to rapidly predict the forces. These predictions will be used to identify optimal adsorbents for some industrially important separation processes. The results of this project will improve the efficiency of adsorptive separation processes. Additional benefits will come from outreach to high school students through summer research programs, to professionals via online courses, and to broad audiences through an educational video. Chemisorption is the adsorption of a molecule on a surface accompanied by the formation of chemical bonds. Chemisorption is relevant to many separation processes that are based on adsorption/desorption using an adsorbent surface. A major ongoing research effort is the computational design of adsorbents that can selectively adsorb desirable species out of a mixture. A crucial step in such computations is to develop interaction potentials to describe the chemisorption between the molecule and the surface. This project seeks to reduce the computational effort to develop such interaction potentials. Simulations will be conducted to obtain interaction potentials between specific molecules and adsorbent surfaces. These data will be fed to a machine learning model to predict the interaction potential between the same molecules and other adsorbent surfaces. The central hypothesis to be tested is that adsorbent surfaces can be treated as “building blocks” so that the interaction potentials are transferrable from one adsorbent to another with similar building blocks. This project will dramatically boost data generation capabilities to train artificial intelligence frameworks for adsorbent discovery. The project will rapidly screen hundreds of thousands of metal organic frameworks to identify adsorbents that can separate mixtures of small molecules such as ethane/ethylene, ammonia/nitrogen/hydrogen, etc. Separation of such mixtures is immensely important in the chemical industry. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.