The national Science Foundation convened a second workshop in August 2024 at the University of Minnesota to reassess how generative AI can accelerate scientific discovery. Building on a March 2023 meeting, participants took stock of rapid model advances and tested prior recommendations against new capabilities and failures. The convening framed generative AI as both a tool for automating laborious tasks and a driver of new scientific questions that, in turn, push method development. The report synthesizes objectives, participant perspectives, technical gaps, and a set of actionable recommendations intended to guide research funding and community priorities.
The workshop brought together 31 experts spanning computational biology, neuroscience, climate science, physics, and related fields, with representation from academia, industry, and government. Programmatic activity combined lightning talks and panel discussions, aimed at connecting domain problems to model capabilities. Organizers prioritized four goals: evaluate emerging technologies for discovery, identify limitations to reliability, highlight gaps where generative approaches fall short, and outline strategies for integrating these models into scientific workflows. Half the attendees had participated in the 2023 workshop, enabling a reflective dialogue about progress and persistent challenges.
Participants reviewed how generative systems rest on self-supervised learning and context-sensitive architectures like transformers, which let models adapt outputs to complex inputs and learn from massive unlabeled datasets. The workshop cataloged current scientific applications: protein structure prediction, drug discovery, materials design, and climate modeling among them. These use cases show promise in exploring chemical spaces, predicting molecular properties, and simulating systems that were previously intractable. Yet success stories coexist with hard limits; models can generate plausible but incorrect results, and they are often resource intensive to train and run.
The report lists technical gaps blocking broader scientific adoption: robustness and generalization, rare class handling, explainability and trustworthiness, uncertainty quantification, energy and computational efficiency, reasoning under uncertainty, and symbolic reasoning. To address these, experts recommended integrating domain knowledge into training, developing general uncertainty quantification methods for generative models, and combining generative approaches with reinforcement learning, planning, and symbolic techniques. They also called for an ecosystem of curated, AI-ready benchmark datasets, standardized evaluation metrics, cyberinfrastructure to deploy tools at scale, trans-disciplinary training programs, and collaborative frameworks that link AI researchers with domain scientists. Together these steps aim to make generative systems dependable partners in discovery rather than black box curiosities.
