Workshop contact:
Bobbie Scott, csedsi@umn.edu
Location:
McNamara Alumni Center
200 Oak St. SE
Minneapolis, MN 55455
GenAI is evolving at an exceptional pace and offers significant potential to accelerate scientific discovery. At the same time, scientific applications pose unique and often unmet challenges for current models. Addressing these challenges will not only expand the role of GenAI in science but also help drive its future development in response to the needs of scientific inquiry—particularly through the integration of scientific knowledge into GenAI frameworks.
This workshop will focus on advancing the methodological foundations of GenAI for scientific discovery, with a particular emphasis on approaches that incorporate domain knowledge, physical laws, and theoretical constraints.
Key goals include:
Analyzing the relative strengths and limitations of current GenAI approaches—including diffusion models, autoregressive models, VAEs, normalizing flows, and GANs—to organize the rich landscape of GenAI approaches and guide model selection across disciplines and scientific tasks.
Highlighting best practices for developing and deploying foundation models in science and engineering, including data collection, architecture design, and pretraining/fine-tuning based on the needs of scientific disciplines.
Addressing challenges, particularly in out-of-distribution settings, and exploring strategies for principled uncertainty quantification and to enhance robustness and alignment with scientific principles.
Identifying opportunities in GenAI for fundamental advances in their ability to perform symbolic reasoning, generate and validate scientific hypotheses, support principled decision-making, or provide scientifically meaningful explanations—capabilities critical for many scientific discovery processes.
These discussions will be grounded in application domains such as materials design and manufacturing, agriculture, environmental science, life sciences and health care, where GenAI holds significant promise and where the integration of scientific knowledge is especially critical.
This workshop will build on two prior events focused on the transformative potential of AI in science. The first workshop, held at NSF headquarters in Arlington, VA (March 8–9, 2023), led to this report, whose recommendations informed the recent multi-year NSF solicitation ACED: Accelerating Computing-Enabled Scientific Discovery (NSF 24-541). The second workshop (see report), hosted at the University of Minnesota on August 6, 2024, gathered feedback on the initial report and revisited its recommendations in light of recent developments—especially the rapid emergence of Generative AI (GenAI).
Slide show of images from the second NSF-sponsored workshop on the AI-enabled scientific revolution in August 2024.