Sandeep Madireddy is a computer scientist in the mathematics and computer science division at Argonne National Laboratory. His research spans multimodal and large-scale generative AI applied to diverse scientific data modalities, evaluation and safety of frontier AI models, inference-time and adaptive compute for resource-efficient reasoning, and uncertainty quantification. He leads the Multimodal Scientific Reasoning Models team in ModCon within the Genesis Mission, serves as the AI Co-Lead for the RAPIDS3 ASCR-SciDAC Institute. His work received a Best Paper Award at the Climate Change in AI Workshop at ICLR, and 2025 Climate Gordon Bell Prize finalist.
Presentation Title:
Scalable Flow based Generative Models for Science
Presentation Abstract:
Flow-based generative models, including diffusion and flow matching, provide a flexible framework for learning complex, high-dimensional scientific data. We present scalable flow-based approaches for efficient probabilistic learning from noisy data and long-horizon spatiotemporal generation, enabling future trajectories to be modeled jointly rather than autoregressively. We also demonstrate how hard physics constraints can be incorporated directly into the generative process to ensure physically consistent and feasible predictions. Finally, we highlight emerging multimodal scientific foundation models that jointly reason over language and scientific structures, paving the way toward unified AI models for scientific discovery.