Physics-Based Molecular Modeling and Machine Learning for Accelerated Discovery of Metal-Organic Frameworks

Metal-organic frameworks (MOFs) are a class of crystals, whose chemically and structurally tunable pores have made them exciting prospects for a myriad of applications across chemistry, engineering, and materials science. MOF tunability stems from a modular structure that can drastically (or subtly) vary by “mix and matching” potential constituent building blocks. This tunability, however, gives rise to an overwhelmingly large material design space that cannot be efficiently explored by experiments, or even by relying solely on traditional computational methods such as molecular simulation. In this presentation, I discuss some of the ways molecular simulation and machine learning have been combined to accelerate the navigation of databases of MOF computational “prototypes” for applications seeking to exploit selective sorption of molecules into MOF pores.<br/><br/>First, I discuss how the switch to a physics-based approach allowed us to address the issue of (sorption) data scarcity by allowing us to create “synthetic data” that led to the training of an effective multitasking machine learning model to predict sorption of a diversity of molecules in MOFs at a variety of thermodynamic conditions. The use of this model to find promising MOFs from a 50,000+ MOF database for removal of Xe and Kr in nuclear fuel reprocessing is demonstrated.<br/><br/>Second, I discuss the combination of physics-based MOF representations, traditional descriptor-based hierarchical screening, and iterative machine learning model training to efficiently explore a 10,000+ MOF database to find promising MOFs for product extraction in plasma reactors used for synthesis of NH3. The efficacy of the above combination of domain-knowledge and “off the shelf” machine learning tools is compared against more data-science involved approaches such as diversity-driven Bayesian optimization.