(297c) Towards a Universal Adsorption Model Using Machine Learning: Motivation, Progress and Challenges

Adsorption is central to innumerable processes of biological, physical and chemical nature. Hence, adsorption has concerned scientists for long. Perhaps more notably in the first half of the 20^th century, when efforts to develop models to the predict the adsorption isotherm (the “heart” of adsorption studies) flourished. Some consider the Brunauer-Emmet-Teller (BET) model as the first attempt to create a universal adsorption theory, followed by models such as the Dubinin-Radushkievich (DR) model. However, it quickly became clear that an analytical equation including all factors affecting adsorption would have such complexity that it would be impractical, given the mathematical tools of the time.

More than 70 years later, adsorption continues to be critical in a diversity of applications. For example, for chemical separations, adsorption predictions are required to accelerate the discovery of adsorbents among a “search space” currently spanning millions of adsorbents, including “new age” materials such as metal-organic frameworks (MOFs). Although molecular simulation have helped navigate the above search space, the cost of adsorption predictions with this method has limited the “bandwidth” of the research community to a very limited number of chemical separations at specific conditions, while only considering limited points in the adsorption isotherm to estimate relative performance among adsorbents.

A fast universal model to predict adsorption isotherms would enable i) searching for suitable adsorbents for the ~80% of current chemical separations that could potentially see their energy cost reduced 10-fold by switching from a thermal method to an adsorptive method, and ii) coupling adsorbent and process optimization, as well as facilitating process modeling-based screening for the top adsorbents. Here it is proposed that now it is the time to retake the development of universal adsorption models by leveraging the mathematical tools of the “age of data.” As a first step, it was demonstrated that the same deep neural network (DNN) trained with molecular simulation data for 200 alchemical molecules in 2,000 MOFs can predict individual full adsorption isotherms in the 0.01 to 100 bar range for CH₄, Ar, Xe, Kr, H₂, N₂, O₂, C₂H₆ in 400 MOFs. Moreover, the DNN model was found to have good transferability when tested on MOFs of a separate database previously unknown to the model.

To illustrate the screening possibilities that a universal DNN-based adsorption model opens, the model predictions were combined with ideal adsorption solution theory (IAST) to screen a ~50,000 MOF database for their capability to separate four industrially relevant mixtures (Xe/Kr, CH4/CH6, N2/CH4, and Ar/Kr) at multiple compositions and pressures. For this framework to work with sufficient accuracy, it was found critical for the DNN to make accurate predictions at low pressures (0.01–0.1 bar). After training a model with this capability, it was found that MOFs in the 95th and 90th percentiles of separation performance determined from DNN+IAST calculations were 65% and 87%, respectively, the same as MOFs in the simulation-predicted 95th percentile across several mixtures at diverse conditions (on average). Finally we discuss currently pursued strategies and related challenges in extending prediction capabilities of machine learning-based towards more diverse types of molecules.