ATD: Algorithmic Threat Detection and Mitigation with Robust Machine Learning

With the recent explosion of scientific data, machine learning (ML) is permeating many areas of science and engineering. Even areas that are traditionally dominated by differential equation models are incorporating ML into their models. However, confidence in and broader adoption of ML in science and engineering face a number of challenges, in large part because existing ML models were developed mostly with non-scientific applications in mind.

This project seeks to broaden the appeal of ML in scientific applications (hereafter called scientific ML) by addressing three key issues with existing ML models that are hindering their broader adoption in scientific applications.

This project will support one graduate student each year of the three year project. First, we consider how to tailor existing ML models to scientific applications by incorporating (scientific) domain knowledge. In most scientific applications, there is a rich body of domain knowledge to draw upon.

If we can properly integrate this knowledge with ML models, it allows the machine to focus on learning less well-understood aspects of the underlying theoretical and physical processes. Ultimately, this not only leads to ML models that can outperform human engineered models, but also to new insights into the underlying physical processes. Second, we consider how to train ML models that are stable and reliable under perturbations in the training data, model choice, and computational errors.

ML models are known to be unstable under such perturbations, which undermines their credibility in scientific applications. Third, we focus on training ML models that are capable of predicting the effects of interventions on a system. This is an area in which ML models are lacking compared to traditional differential equation models.

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.