Speaker
Description
Artificial intelligence has rapidly become an integral part of high-energy physics, evolving from specialized machine-learning tools for individual analysis tasks toward increasingly general and large-scale models for scientific data. As this field enters a new stage, an important question is not only how to build more powerful models, but also what we can learn from the models themselves. In this talk, I will review recent developments in AI for high-energy phenomenology, with a particular focus on collider physics and the emerging paradigm of foundation models. I will then discuss ongoing efforts to probe the internal representations learned by these models, using geometric approaches to study how their representations evolve during training and across architectures. More broadly, I will argue for a perspective on Scientific AI in which AI is not merely a computational tool for solving predefined physics tasks, but can increasingly serve as a new instrument for uncovering physical structures from the data itself.
