Abstract:

One of the main challenges in the molecular sciences is identifying promising candidates for a target application given the virtually limitless number of possibilities. Computer simulations and machine learning (ML) are integral tools to survey the vast molecular landscape and the thermodynamic conditions to evaluate their performance. Broadly, a typical pipeline for molecular characterization, design, and discovery with ML involves data curation and generation, featurization and/or descriptor development, and a model or algorithm for the final prediction. However, these tools are currently severely limited due to the lack of available data, because of the time and resource expense to generate it. In this seminar, I will discuss our research efforts focused on metal-organic frameworks (MOFs), which are nanoporous, crystalline materials composed of inorganic and organic components. I will begin by showcasing our use of large-scale, high-throughput computational screening techniques to generate gas adsorption data. Following this, I will illustrate how transfer learning and active learning can significantly reduce the data requirements of ML models, uncover insights into the material space, and enhance model capabilities when combined with thermodynamic information. Finally, I will present our recent work on utilizing sigma profiles—descriptions of molecular electrostatic surface charges—as chemical descriptors to efficiently navigate and optimize molecular landscapes.