Many synthesized chemicals in current use, based on carbon-intermediates, are byproducts from the crude oil refining industry. Due to the industry's highly efficient infrastructures along with cost-effective methods, it has been hard to think of any chemicals other than petrochemicals. However, with a recently growing desire for alternative carbon sources, more and more research has started to target the biomasses, which are potentially richer sources full of diverse intermediates. Such biomass-derived intermediates convertible to various chemicals are defined as "bioprivileged" by previous work in the field (Shanks, Broadbelt). Still, there is no apparent method to identify the bioprivileged molecules; the existing experimental approach is slow because only a small number of structures may be evaluated at a time. Thus, we use de novo molecular design models trained on a variety of commercially available chemical compounds databases, such as ZINC and PubChem, to generate a massive set of potentially bioprivileged molecules having a set of desired functional moieties. We measure the likelihood of a molecule being bioprivileged by the R0 and R1 reactivity indices, a recursive quantification of the number of reactive moieties used to estimate the reactivity of corresponding species. The reactivity indices serve as a target for optimization within a de novo design framework and allow us to generate candidate molecules that are enriched with the desired functional groups. This method may serve as the foundation for future work on discovering bioprivileged molecules including retrofitting synthetic pathways from biowaste precursors, optimizing reaction kinetics/selectivity through catalyst design, and analyzing the physiochemical properties of downstream products with respect to a set of desired functionalities. It may also be used in more general contexts to generate a set of molecular structures that are enriched with any set of desired substructures.

The viability of redox-flow (RF) batteries has, in recent years, become an increasingly prevalent point of interest in the chemical research community. RF batteries make use of the reversible electrochemical conversion of active redox species as a form of long-term energy storage. Currently, the most practical versions of these batteries utilize a vanadium-based solution, which is both costly and difficult to manufacture on a large scale. To solve this issue, researchers explored the possibility of using organic-based solutions and natural solvents. However, most of these batteries are limited to specific classes of organic molecules. Through the development of a generalized predictive model, we will create an accurate method of predicting the redox potential of a wide assortment of organic molecules which can be used to improve downstream generative AI algorithms for molecular design. To create our predictive model, we construct a set of experimental and computational redox potentials, which train our model. After compiling a database of roughly 100 organic molecules, we use our model to find correlations between the molecules’ measured redox potential and additional properties, which are calculated using various cheminformatics packages. We expect to find an approximate correlation within an acceptable range of error, which our model can base its predictions on. The limitations of our predictive model stem from our small sample size—larger data sets directly correlate to more accurate results. The successful development of a predictive model with a bounded range of error largely improves our ability to accurately find candidate molecules with high redox potentials, molecules which could potentially be used in large-scale redox flow battery systems.