(690d) A Literature-Derived Data Set and Machine Learning-Enabled Prediction of Metal-Organic Framework Water Stability

The water stability of a metal-organic framework (MOF) determines its viability in applications entailing water exposure. Consequently, it is useful to predict whether a MOF is water-stable before investing time and resources into synthesis. While heuristics for designing water-stable MOFs are well-known, these can lack generality and artificially limit the diversity of explored chemistry. Thus, we instead look to machine learning (ML) models informed by experimental data extracted from the literature. We first expand on a prior MOF water stability data set by over 400% to better train these models. The additional data is shown to improve model performance over varied chemistry, as evidenced by test set ROC-AUC scores above 0.8, for the prediction of water stability and stability in harsher conditions, such as in acidic aqueous solution. We then illustrate how the expanded data set and models can be used to screen for not-yet-synthesized, application-ready MOFs through genetic algorithms that consider three different types of stability jointly. The genetic algorithms enable efficient assessment of thousands of MOFs generated through in silico recombination of secondary building units (SBUs) and linkers. The data set and models are expected to contribute toward the synthesis of novel, water-stable MOFs for gas capture and water treatment, among other applications.