Matt Menickelly, an assistant computational mathematician, is developing better adversarial machine learning (ML) methods through robust optimization. Adversarial learning refers to a learning scenario that simulates, and attempts to bypass, an attacker trying to force a ML classifier to misidentify something. Such misidentification can have serious consequences for cybersecurity and scientific discovery. Menickelly is exploring and designing new training algorithms intended to defend against such attacks. The research exploits techniques of optimization under uncertainty, particularly distributionally robust optimization.

Anirudh Subramanyam, a postdoctoral appointee, is developing advanced algorithms for optimizing critical infrastructure that are affected by high-impact rare events such as natural disasters and cyber attacks, which can cause catastrophic network failure. His aim is to formulate methodologies that combine optimization techniques with modern statistical tools from large deviations theory to enable accurate quantification – and subsequent mitigation – of these high-impact rare events.

Ashwin Renganathan, a postdoctoral appointee, is applying statistics and machine learning to improve the predictive accuracy of reduced-order models (ROMs) of transonic flows. He is also developing information theoretic acquisition methods that scientists can use to design optimal experiments and then train the accurate ROMs with as few simulations as possible. His research is motivated by aircraft design problems, where the millions of variables in the mathematical models currently limit the number of costly simulations that engineers can run.

LDRD Seed Awards are made in the amount of $25,000 to the laboratory’s postdoctoral and early career researchers as they pursue their independent research initiatives. A total of 15 awards were made for 2020. For further information about the SEED Awards, see this article (requires myArgonne login).