100 Researchers Map How AI and Physical Sciences Should Evolve Together
A hundred researchers across astronomy, chemistry, materials, math, and physics map out how the AI-science relationship should evolve in both directions.
Takeaways
- Over 100 researchers across MPS domains published a strategic vision paper for AI+Science from an NSF workshop.
- Argues AI and science are a two-way street, citing the 2024 Nobel Prizes in Physics and Chemistry as evidence.
- Identifies six priority AI techniques: simulation-based inference, multi-scale simulations, uncertainty quantification, foundation models, experimental control, data-efficient methods.
- Calls for a dedicated Science of AI research program to crack open the black box using physics techniques.
- Recommends centralized compute and data infrastructure, expanding programs like the NAIRR pilot.
- Pushes universities to hire interdisciplinary centaur scientists and reform tenure evaluation for AI+Science work.
A sweeping community report from over a hundred researchers spanning astronomy, chemistry, materials science, mathematics, and physics has just landed on arXiv, laying out a coordinated vision for how the mathematical and physical sciences should engage with the AI wave. The paper emerged from an NSF-sponsored workshop and reads less like a research result and more like a policy blueprint, but it offers a rare consolidated view of where working scientists think the AI-plus-science relationship is headed.
The framing is deliberate. The link between AI and the Mathematical and Physical Sciences is becoming increasingly inextricable, as evidenced by the 2024 Nobel Prizes in Physics and Chemistry. One prize recognized the statistical physics roots of modern neural networks, the other went to computational protein design. The authors argue that these are not one-offs but symptoms of a deeper structural relationship that funding agencies, universities, and researchers should now formalize.
The two-way street thesis
The central argument is that AI and science are mutually reinforcing, and treating the relationship as one-directional (AI as a tool for science) misses half the value. On the science-benefits-from-AI side, the authors catalogue concrete wins across domains:
- Astronomy has driven progress in real-time control systems, scalable image processing, anomaly detection, and uncertainty quantification, motivated by the analysis of massive sky surveys and rare astronomical events.
- Chemistry has advanced AI through the development of graph neural networks and generative models for molecular design, significantly accelerating tasks such as materials and chemical matter discovery.
- Materials science has spurred innovations in inverse design, autonomous experimentation frameworks, and multiscale surrogate modeling.