DMREF: Accelerated Design and De-risking of Complex Organic Molecular Solid Forms Using New Theoretical and Experimental Methods in a Closed-loop Approach

Non-technical Description: A team of researchers from New York University, UC Riverside, and the University of Iowa tackles one of the most challenging problems in materials science: accurate computational prediction of the structure of organic molecular crystals starting from nothing more than a chemical diagram of the crystal’s constituent compound or compounds. Current structure prediction protocols require substantial high-performance computing resources and specialized expertise, which are not generally available to most research teams. The team proposes creating user-friendly, open-source software for predicting molecular crystal structures that can be used by non-experts. The proposed approach is based on a mathematically derived structure generation workflow that accurately accounts for patterns observed in all available molecular crystal systems, an Artificial Intelligence (AI)-driven accurate ranking of structures, and experimental solid-form screening techniques that will provide critical assessments of the computational approaches. Organic molecular crystals are ubiquitous materials used in pharmaceutical formulations, electronic devices, smart materials, dyes and pigments, insecticides, and military applications. The team proposes building strong relationships with K-12 schools across all three states to introduce students to age-appropriate workshops on concepts in the science of molecular materials, including symmetry, machine learning, crystallization, quantum science, and big data. Technical Description: Predicting computationally structures and polymorphs of organic molecular crystals remains a significant challenge. Current approaches involve searching for an intermolecular interaction energy model to identify candidate structures. While reasonably effective for relatively simple crystals containing only moderately flexible molecules, highly flexible molecules, crystals with multiple symmetrically distinct molecules, cocrystals, and hydrates/solvates remain out of reach of these techniques. Moreover, the high computational cost of these methods for simple crystals renders them inaccessible to many groups interested in molecular crystal structure prediction. Finally, interpreting the output results is complicated by the fact that this approach predicts far more potential structures than are experimentally observed. More recent attempts to create generative AI schemes reveal that these methods are even less predictive than random search using an interaction energy model. This project aims to develop a new paradigm for predicting molecular crystal structures by leveraging a mathematically and data-driven framework, known as ‘CrystalMath’, for the rapid generation of candidate structures without overprediction, combined with machine-learning-based frameworks for evaluating the energetic and thermodynamic stability of candidate structures. Accurate solid-form screening experiments will be used to validate the computational protocol. The workflow and corresponding software tools to be developed will reduce the associated computational cost and resource requirements of predicting molecular crystal structures by 1-2 order of magnitude over existing methodologies, thereby lowering the barrier to the adoption of these tools by a wide range of groups needing such molecular crystal structures and opening the door to accelerated design and deployment of complex organic molecular crystals for a wide range of applications. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.