(2jh) Machine-Learned Committor Functions for Reactive Molecular  Dynamics

Computational materials science is an agile field, adept at filling gaps in our understanding of chemical space. However, due to the push-and-pull of accuracy vs. scale when modeling materials with finite resources, computational insight often amounts to a rough sketch. Thus far, I have dedicated my research career to strengthening the interaction between in silico and in-the-lab materials science. As a NASA postdoctoral fellow who started just after the pandemic began, i.e., after the labs were shut down, my postdoc experience has been a case study of the relationship between theory and experiment. I developed a reactive molecular dynamics model to predict the charring behavior of materials used for thermal protection systems
during hypersonic flight, such as on the leading surface of a spacecraft. After more than a year of fine-tuning while waiting for the labs to reopen, the model was confirmed to accurately predict the char yield for several newly proposed chemistries. Since then, I have used the model to help guide NASA scientists in the lab, which has been an inspiring testament to the predictive power of modern modeling techniques and of the faith that experimentalists have in them. These types of successes are what drive me to continue to improve both the accuracy and scale of atomic level models of our physical world, a discipline that has provided major contributions to our understanding of materials from ionic liquids to perovskites.

While at NASA, I developed the REACTER protocol, a method for modeling chemical reactions in classical molecular dynamics simulations. Implemented in the open-source LAMMPS software package, it has established a user base that is growing quickly in the fields of polymer and soft matter sciences. However, REACTER remains underutilized in many adjacent modeling communities, such as those studying electrochemical components of next-generation batteries or complex reactive pathways within the cell. New features of REACTER enable competing reaction pathways to be considered while reactivity is informed by quantum mechanics. At this AIChE conference, I will describe the first effort to use machine learning to provide reaction probabilities in real time, a step toward semi-automated incorporation of quantum-level information into these highly scalable atomic-level simulations. The method has a long way to go and will benefit greatly from closer integration with data science, e.g., libraries of reactions. As a faculty candidate, I would like to share my vision of how large-scale reactive molecular dynamics can be applied to systems from batteries to biology, as well as how it can be used as an instructive visual tool to teach students about chemistry and kinetics.

Finally, I would like to comment on a few topics outside of science that are important to me. Like too many graduate students, I have personally experienced how the power dynamic between advisors and students can be abused, leading the mistreatment of grad students (or postdocs). Secondly, I have witnessed that those who routinely exploit this power dynamic can also exhibit patterns of discrimination, often related to gender. Between this experience as well as the number of second-hand anecdotes I have heard, it is clear to me that systemic problems exist in academia. It will be of foremost importance to me to culture an atmosphere that is positive, accepting, and encouraging. I believe that actively taking a step back and cultivating such an atmosphere should be done not in spite of, but because of, the highly competitive nature of research. Welcoming new ideas and diverse perspectives is a key ingredient to scientific progress and is only sustainable if we make ourselves aware of existing issues, listen to those who raise concerns, and work to prevent them from recurring.