Gathered data on a set of reactions that have gone unpublished as they failed to produce a desirable product. Goal of data to lower the cost of and accelerate new material discovery. Currently doing research with unlabeled reaction generators, model interpretability, and active learning. For my coding contributions, see the Dark Reactions Github (https://github.com/darkreactions/DRP).• Built an active learning pipeline for recommendations of reaction syntheses to perform to better understandorganically templated crystal formation. Implemented in Django and uses the RDKit• Built and optimized parameters for hundreds of machine learning models for supervised and semi-supervisedlearning to compare performance across models, feature sets, and hyperparameters• Developed a theoretical framework for understanding why individual active learning recommendations are made• Assisted in developing a measure for bias in model confidence based in the legal precedent of disparate impact• Developed unsupervised strategy for characterizing trends in uncertainty for any dataset• Used adversarial modeling, such as in GANs, to transform images to help domain experts understand classifier predictions

• Contributions to linear regression models for efficient high-throughput redox potential calculations Created a model from scratch as a part of an independent study.• Ran density functional theory calculations for over 30,000 cpu hours on four NERSC supercomputers• Data analysis and reduction on large datasets• Wrote program in C++ and Python to extract bond data and to manipulate large-scale (functional group) features using Cartesian coordinates of molecular geometries

Performed tests to detect 65 EPA designated water contaminants as well as bacteria and pesticide contamination• Organized electronic data collection and meta-analysis for efficiency in result reporting and error detection• Supervised five undergrad and high school interns. This included training, scheduling batches and bench time and problem-solving unusual results.