III: Small: Developing A Trustworthy Toolbox for Double-Correct Predictive Modeling in Sciences

The increasing reliance on Artificial Intelligence (AI) and Machine Learning (ML) in critical scientific domains requires models that both predict with high accuracy but also derive their conclusions from valid and justified rationales, particularly for data with deviations from known patterns. This project addresses the critical gap in current AI/ML practices, where models are often trained on historical data insufficient to capture the full spectrum of complex, evolving phenomena, such as extreme weather events. The project introduces a generic, trustworthy toolbox aimed at enhancing the validity, explainability, and scalability of existing predictive models. By focusing on both the accuracy of predictions and the validity of their underlying rationales, the toolbox ensures that AI/ML systems remain reliable in the face of rare or even unprecedented scenarios, resulting in decisions that domain experts can trust. Resources will be made publicly available, ensuring that the broader scientific and technology communities can access, leverage, and build upon the advances. This project develops a trustworthy toolbox for making accurate predictions that are backed by scientifically grounded rationales. This is an important step towards trustworthy AI, ensuring that models used in critical scenarios are both precise and rationally transparent. The project includes three major areas of research. The first focuses on making the best predictions possible, even when faced with data patterns that are very unusual. The second maximizes the reliability of the reasoning supporting the models. The third deals with improving the models’ ability to scale and address the growing volume of scientific data. Research outcomes will be translated into open-source software, workflows that can be made widely applicable, and AI-ready scientific datasets. The principal investigators (PIs) will integrate research findings at multiple educational levels by collaborating with both academic and non-academic researchers. Initiatives include developing Vertically Integrated Projects (VIP) to engage undergraduate research, organizing ML in science hackathon events, presenting at the University of Delaware's annual DARWIN symposium, and reaching out to K-12 teachers and students at local high schools. 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.