D3SC: CDS&E: Collaborative Research: Machine Learning Modeling for the Reactivity of Organic Contaminants in Engineered and Natural Environments

With support from the Environmental Chemical Sciences Program of the NSF Division of Chemistry, Professors Huichun Zhang of Case Western Reserve University and Dong Wang of University of Illinois Urbana—Champaign will develop machine learning models to predict the reactivity of thousands of organic contaminants (OCs) in engineered (water) and natural (soil and sediment) environments. To assess and mitigate risks associated with this vast number of OCs, accurate predictive models are needed to readily provide reasonable estimates of their reactivity, both during important water treatment processes and in the environment.

However, existing models rely heavily on conventional statistical methods. They have multiple limitations such as small numbers and narrow scopes of OCs involved and lengthy calculations of molecular properties. The project will employ advanced machine learning algorithms to predict contaminant reactivities.

The obtained machine learning models will help identify OCs of concern and optimize the treatment processes. In addition, environmental data science will be developed as a new educational track at the pilot scale. Graduate, undergraduate and high school students with diverse backgrounds will be engaged in interdisciplinary research, including modeling and experimental work.

The project also plans hands-on activities on OCs for girls in grade 6-12 and underrepresented college students.

This study will systematically develop comprehensive and accurate machine learning models for predicting the reactivity of thousands of OCs in advanced oxidation processes (AOPs), adsorption onto engineered adsorbents, sorption onto soils and sediments, and biodegradation. The objectives of this research are to 1) mine the literature and available databases to obtain the largest datasets of contaminant reactivity in AOPs, (ad)sorption and biodegradation; 2) experimentally quantify the reactivity of selected OCs in AOPs, (ad)sorption and biodegradation; 3) develop confidence-aware machine learning models for the reactivity of OCs based on the data from the above two objectives; and 4) interpret the obtained models to make them trustable and define their applicability domains.

OCs will be modeled by new chemical representations including molecular fingerprints, molecular images, and different combinations of them with molecular descriptors. Including (ad)sorbent properties in the (ad)sorption models will be a major step to expand the model applicability to diverse (ad)sorbent structures and properties. Properly interpreting and modifying the obtained models and calculating model confidence bounds will make the obtained models trustable.

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.