Dipartimento di Ingegneria informatica, automatica e gestionale

Abstract: Neural networks have shown great promise across geoscientific applications, yet their complex, nonlinear nature often hinders interpretability and limits scientific insight and model trust. Explainable AI (XAI) methods aim to attribute a model's prediction to specific input features, but their evaluation typically relies on image-based benchmark datasets like MNIST or ImageNet, which are lacking objective ground truth for attribution.

In this work, we introduce a framework to generate benchmark datasets with known attribution ground truth, using additively separable functions. We construct a large synthetic dataset, train fully connected networks to learn the target functions, and evaluate the performance of various XAI methods by comparing their attribution maps to the known ground truth. Our results highlight when and where specific XAI methods succeed or fail. This benchmark approach provides a much-needed foundation for rigorous, objective assessment of XAI tools in the geosciences, paving the way for more trustworthy models and deeper scientific discovery.


Short Bio: Professor Antonios Mamalakis is an environmental data scientist whose research focuses on applying statistical methods, machine learning, deep learning, and explainable AI to tackle key challenges in environmental science. His work explores topics such as hydroclimate prediction, climate teleconnections, causal discovery, and climate change impacts. Before joining the University of Virginia, he was a research scientist at Colorado State University, where he led efforts to evaluate explainable AI tools in geosciences. His research has been featured in high-impact journals such as Nature Communications, Nature Climate Change, and Geophysical Research Letters. Professor Mamalakis holds a Ph.D. in Civil and Environmental Engineering from the University of California, Irvine, and serves as Associate Editor for the AMS journal Artificial Intelligence for the Earth Systems.