(2hr) Machine Learning Guided Discovery of Organic and Polymeric Materials for Energy and Environmental Applications

Discovery of new polymeric and organic systems oftentimes demands traversing rich chemical, configurational, and combinatorial design spaces. My research interests are centered on utilizing emerging artificial intelligence (AI), machine learning (ML), and molecular simulation techniques to probe these design spaces and discover high-performance materials based on unique macromolecular building blocks. The Anstine Research Laboratory will pioneer AI-based computational frameworks to elucidate the physical and chemical underpinnings that define the structure-property-performance relationships of organic and polymeric systems. We will leverage these understandings, in combination with high-throughput molecular simulations, to pursue next-generation adsorbents for clean energy, polymeric membranes for water purification, and robust thermomechanical materials. We will specifically target connecting macromolecular building block chemistry and sequence to macroscopic observables, allowing us to efficiently identify promising material by performing function-guided optimization via cheminformatics strategies. As an overarching theme, my proposed research program will seek to develop techniques that overcome challenges traditionally regarded as “having too large a design space” or “being too computationally laborious” by integrating AI/ML with principles from chemistry, physics, chemical engineering, and materials science.

Doctoral Research University of Florida, Department of Chemistry and Department of Materials Science and Engineering, Advisor: Coray M. Colina

My dissertation research was centered on screening polymers of intrinsic microporosity (PIMs) for their performance as an adsorbent in the presence of diverse adsorbate species. Standard industry methods for processing chemical mixtures, such as fractional distillation, can carry significant energy costs, which motivates their retrofitting or replacement with adsorption-based separation strategies. The vast majority of molecular simulation studies aimed at screening microporous adsorbents apply a “static framework” assumption, but this limits the range of uptake that microporous polymers can reliably be assessed. Instead, my research contributed a high-throughput computational strategy for evaluating the critical role of adsorbate-induced polymer restructuring and its effect on dictating adsorbent performance in separation and storage applications. A power-law model, based on species critical temperature and molar mass, was shown to accurately predict the ability of an adsorbate to act as a plasticizing agent, allowing for quick estimation of the amount of expected swelling at target loadings. From an adsorbent perspective, adsorption-induced polymer swelling was demonstrated to be reduced when chemical substitution led to stronger intermolecular electrostatic interactions. This finding affirmed the importance of monomer scale design principles, particularly selection of chemical functionality, for synthesizing microporous polymers that have both exceptional adsorption properties and resistance to induced swelling. As a result of several external collaborations, the high-throughput flexible adsorption simulation framework was extended in the latter half of my doctoral research to study metal-organic frameworks, porous polymer networks, and zeolites.

Postdoctoral Research Carnegie Mellon University, Department of Chemistry, Advisor: Olexandr Isayev

Machine Learning Models for Chemical Reactions. Reactive events, ubiquitous among chemical engineering processes, typically require expensive quantum mechanical (QM) methods to study their molecular-level details. My recent research plan has focused on efficiently exploring reaction pathways by leveraging millions of high-throughput QM calculations to build ML models capable of ~10⁶ times faster characterization than density functional theory. Our current ML-based workflow is applicable to diverse organic reactions, supports high-fidelity transition state structure identification, and accurately reproduces interaction energy and forces. These studies are allowing for a deeper understanding of reaction mechanisms, degradation/formation thermodynamics, and stability of organic complexes.

Neural Network Potentials with Long-Range Interactions. An outstanding challenge in the field of molecular simulations is to push the use of neural network potentials to the interaction length scales needed to simulate macromolecules, supramolecular assemblies, and heterogenous materials. Molecular-level insight for these systems is typically obtained using simple force fields, where reparameterization needs to be performed for every system that falls outside the application space: a practice neither sustainable nor efficient. In contrast, we have built a neural network potential that incorporates long-range interactions by propagating information between atoms with a modern iterative message passing scheme. As a case study, we recently applied our model to molecular crystal structure predictions and found that the message passing neural networks are generalizable, have accuracy less than 1.0 kcal mol^-1, and can identify experimentally observed polymorphs. This outcome is encouraging considering the persistent difficulties in crystal structure prediction and that the training set did not contain the constituent molecules or anything resembling their crystals.

Teaching and Mentoring

As an educator and mentor, I will be committed to pursuing meaningful communication, fostering an inspiring academic environment, and advocating for inclusion. My diverse background in physical chemistry, adsorption science, and polymeric materials has prepared me to teach many core subjects in chemical engineering. My specific interests are to instruct undergraduate and graduate level courses in organic chemistry, materials science, and thermodynamics. Beyond the core curricula courses, I will develop polymer physics, machine learning, and molecular simulations electives that contribute to training a new generation of computational and polymer scientists. During my assistantship experience at the University of Florida, I sought to broaden my teaching qualifications by electing to lead laboratory courses: as a computational scientist, this decision was initially disorienting but yielded lasting impact on my abilities as an educator. I experienced the importance of establishing a strong connection between key principles and real-world examples, and this will be central in my future lesson plans. I will augment these examples with molecular modeling visualizations to ensure that my lectures are delivering a holistic view of chemical engineering topics. I am enthusiastic about the opportunity to inspire early career scientists and look forward to putting these plans in motion.