AI for Physics Seminar

Abstract

Foundation models have transformed modern AI by learning reusable representations that can be adapted across many tasks, datasets, and data types. Similar ideas are now beginning to enter physics and astrophysics, where scientific information is naturally multimodal and may include images, spectra, time series, simulation outputs, and physical parameters. In this talk, I will first give a pedagogical introduction to the key ideas behind transformers and foundation models, then review several recent efforts in astronomy and cosmology, including AstroCLIP, AstroLLAMA, and AION-1B. I will then discuss our recent work on a multi-modal foundation model for cosmological simulation data, which jointly models heterogeneous galaxy properties and supports flexible cross-modal prediction through masked training. I will conclude with a broader perspective on how foundation-model ideas may enable more general AI systems for scientific inference and prediction, including recent work extending multimodal modeling to plasma state reconstruction from sparse and incomplete observations.

Abstract

In this seminar we discuss a unified framework that combines explainable deep learning, deep reinforcement learning (DRL) and foundation models to advance both understanding and control of turbulence, with direct implications for accelerated design and discovery. First, we will show how explainable deep learning techniques can be used to identify the flow features that are truly responsible for key turbulent processes in wall-bounded flows. By systematically interrogating trained neural networks, we uncover the most influential coherent structures driving momentum transport and drag. Our results reveal that classically studied structures (while important) provide only a partial and sometimes misleading perspective, motivating a more data-driven and physics-aware view of turbulence organization. Building on these insights, we will demonstrate how deep reinforcement learning can be used to actively control turbulent flows by targeting the dynamically relevant structures identified through explainability. This approach achieves over 30% drag reduction in canonical wall-bounded turbulence and extends naturally to more complex configurations, including turbulent wings, highlighting the scalability of learning-based control strategies. Finally, we will introduce a foundation-model-based framework for accelerated design, optimization and scientific discovery. By learning compact, interpretable latent representations of high-dimensional flow physics, these models (combined with agentic-AI systems) enable rapid exploration of design spaces, causal reasoning and closed-loop optimization, bridging the gap between expensive simulations, control and engineering decision making. Together, these results illustrate how explainable and agentic AI are becoming essential for turbulence physics, flow control and next-generation engineering design.