(582a) Accelerated Discovery of Metal-Organic Frameworks for Gas Separations Using Bayesian Optimization

High-throughput screening has been increasingly important for computational discovery of metal-organic frameworks (MOFs); such an approach is useful for gathering material design insights but requires vast computational expense. Machine learning models can provide faster estimations of desired properties compared to rigorous calculations and can accelerate screening studies; however, machine learning methods typically require on the order of 10³ datapoints at a minimum for training. Here, we show that Bayesian optimization can identify promising candidate gas adsorbents after training on only a fraction of a given dataset. Using ~51,000 hypothetical MOFs and publicly available CH4, CO2, and N2 adsorption data from Wilmer et al. [1], we demonstrate that our Gaussian-process surrogate model is able to reliably find top 10 candidates for methane uptake at 35 bar while training on less than 0.1% of the dataset. We further show that the Gaussian process provides interpretable design insights while finding optimal structures without the need for large-scale screening. We discuss how our proposed workflow could be extended to other gas separation objective functions, such as selectivity between CO2 and N2.

This work is funded by the National Energy Technology Laboratory (NETL) under the Discovery of Carbon Capture Substances and Systems (DOCCSS) Initiative. Computational resources provided by National Energy Technology Laboratory and Lawrence Berkeley National Laboratory.

[1] Wilmer, C., Leaf, M., Lee, C. et al. Large-scale screening of hypothetical metal–organic frameworks. Nature Chem 4, 83–89 (2012).