AOSC Seminar by Antonios Mamalakis, 3/12/2026
AOSC Seminar
Antonios Mamalakis
University of Virginia
Title
Has the AI Model Learned the System? Using Explainable AI to Diagnose Knowledge and Ignorance in Climate Prediction
Abstract
AI models are increasingly used to predict atmospheric and oceanic phenomena, from precipitation extremes to large-scale climate variability. While these models can achieve impressive predictive skill, a central scientific question remains unresolved: when has an AI model actually learned the underlying physical system?
In this talk, I address this question using a controlled synthetic benchmark inspired by climate prediction problems, where the true drivers of the target variable are explicitly known. This framework enables rigorous evaluation of explainable AI (XAI) methods under varying signalto-noise ratios and data availability; conditions that closely mirror those encountered in atmospheric and oceanic science.
Two key insights emerge. First, explanation reliability is tightly linked to physical learnability: XAI methods recover the correct drivers only once the model has genuinely learned the underlying signal. Second, agreement across different explanation methods or independently trained models provides a practical and quantitative indicator of whether an AI system is operating in a regime of physical understanding or in one of epistemic ignorance.
Together, these results suggest a new role for explainable AI in atmospheric and oceanic science; not merely as an explanation or visualization tool, but as a diagnostic framework for physical learning and scientific trust. More broadly, this work points toward AI systems that help scientists assess not only what and how a model predicts, but whether it truly knows the system it is trained to approximate.
Bio
Antonios Mamalakis is an Assistant Professor jointly appointed in the Department of Environmental Sciences and the School of Data Science at the University of Virginia. His research interests include hydroclimatic variability and predictability, artificial intelligence for geoscience, explainable AI, and the development of data-driven tools for improving prediction and understanding of environmental systems. Antonios received the Diploma and the M.S. degrees in Civil and Environmental Engineering from the University of Patras, Greece, in 2014 and 2016, respectively, and a Ph.D. degree in Civil and Environmental Engineering from the University of California, Irvine, USA, in 2020. Until 2023, he was a Postdoctoral Researcher and a Research Scientist at the Department of Atmospheric Science at Colorado State University
Contact
If you are not subscribed to seminar announcements and need a link to this online seminar please contact aosc-helper. You can also subscribe to weekly links and announcements below.
AOSC Seminar
Pre-seminar refreshment: N/A
Seminar: 3:30-4:30pm, Room: ATL 2400(only when in-person)
Meet-the-Speaker: 4:30-5:00pm, Room: ATL 3400(only when in-person) [For AOSC Students only]
[ Subscribe to receive seminar announcements | Email aosc-deptseminar@umd.edu to give a seminar ]