Machine-Learning Based Optimization of Sorbent Materials for Energy Storage—A Case Study on Metal Organic Frameworks–MOFs

Metal Organic Frameworks - MOFs are relatively new compounds consisting of metal ions/clusters and organic linkers, and are characterized by unique properties, such as tunable porosity and incredibly high surface area [1]. Those materials are therefore employable in the thermal energy storage field by exploiting physisorption phenomena, which can be accompanied by significant amount of energy exchange.<br/>In this work, we considered adsorption/desorption-based heat pumps as possible application. In such a context, this activity aimed at understanding which properties, at the elementary cell level, influence the performances of a heat pump based on MOFs allowing adsorption/desorption of water, and then, aware of those descriptors, to design an optimal theoretical MOF for that application. Another target of this work was to achieve a comprehensive comparison of algorithms for sequential MOFs optimization, assuming to ignore the values of relevant figures of merit. Such black-box function optimizations can be extremely demanding, especially when a high-dimensional parameter space has to be explored. In such cases, efficient exploration algorithms are provided by Sequential Learning (SL), aiming at reducing the number of evaluations to be performed [2].<br/>In order to achieve those targets, we have first featurized the Crystallographic Information Files (CIFs) of roughly 8000 potential MOFs [3], ending up with 1557 CFID (Classical Force-field Inspired Descriptors); the latters take into account chemical descriptors, such as averages of chemical properties of the elements in the cell, and structural descriptors, such as radial distribution function, nearest neighbor distribution, angle and dihedral distributions. Furthermore, for each compound, Boyd <i>et al.</i> [3] made available three relevant properties: Henry coefficient for CO₂, Henry coefficient for H₂O, working capacity for CO₂. We have therefore fitted and validated three regression models for those figures of merit. Moreover, we have ranked the importance of the 1557 features over the models’ outputs, getting the few dozens of effectively relevant descriptors for each of the three target properties.<br/>For the SL procedure, we have compared more methodologies for efficiently choosing the next material to be tested, aiming at the maximization of the three DFT-based figures of merit above; as result, it turned out that employing the subset of the most important features, in general does not ensure a faster convergence of the SL methodologies. We finally investigated the use of those potential MOFs for closed water sorption thermal energy storage application. More specifically, over the database above, we identified the material maximizing the released heat – dependent on the Henry coefficient for H₂O - under specific operating and environmental conditions. This optimal compound outperforms many of the known microporous materials in terms of energy density [4]. Then, in the hypervolume of the parameter space around this optimum, we set up a SL-based exploration, to find a theoretical combination of (the most relevant) features performing better than the best compound from the database.<br/>[1] S. Kitagawa <i>et al</i>., Metal–organic frameworks (mofs), Chemical Society Reviews 43, 5415 (2014)<br/>[2] E. Brochu, V. M. Cora, and N. De Freitas, A tutorial on bayesian optimization of expensive cost functions, with application to active user modeling and hierarchical reinforcement learning, arXiv preprint arXiv:1012.2599 (2010).<br/>[3] Boyd, Peter G., <i>et al</i>. "Data-driven design of metal–organic frameworks for wet flue gas CO₂capture." <i>Nature</i> 576.7786 (2019): 253-256.<br/>[4] De Lange, M. F., Verouden, K. J., Vlugt, T. J., Gascon, J., & Kapteijn, F. (2015). Adsorption-driven heat pumps: the potential of metal–organic frameworks. Chemical reviews, 115(22), 12205-12250.