(572f) An Efficient Featurization Scheme for Machine-Learning Predictions of Diverse Molecules in Metal-Organic Frameworks

For the past decade, machine learning (ML) studies of adsorption in metal-organic frameworks (MOFs) have proliferated based on the numerous existing databases covering tens of thousands of structures and rapidly improving computer resources. However, many of these studies are focused on individual adsorbates and thus share a similar limitation: transferability of the resulting ML models between different adsorbates. Here, we propose a single ML model that overcomes this limitation and predicts an adsorption property, the Henry’s constant, of a diverse set of > 40 adsorbates in > 400 MOFs. Host-guest interaction energies are known to be critical to determining the Henry’s constant of adsorbates in MOFs. Here, we develop an efficient new descriptor scheme called Gaussian Integral (GI) descriptors in which the energy landscape of host-guest interactions is transformed into a vector with fixed size. GI descriptors are especially suitable to numerically expressing the dispersion potential landscape. With this novel scheme, we show that ML predictions of Henry’s constants across many materials and adsorbates are improved and achieve mean absolute errors within chemical accuracy (1 kcal/mol) as well as an orders-of-magnitude increase in speed as compared to conventional molecular simulations.