MIT researchers are charting a transformative course for the intersection of artificial intelligence and the mathematical and physical sciences (MPS). Led by Professor Jesse Thaler, a new white paper advocates for a "two-way bridge" where AI doesn't just accelerate scientific discovery, but where scientific principles are used to refine and explain AI itself. This symbiotic relationship, termed the "Science of AI," focuses on how scientific reasoning can inform foundational model approaches and push the development of more robust algorithms.

The initiative highlights three specific domains: science driving AI, science inspiring AI, and science explaining AI. By treating neural networks as complex systems—much like particles or galaxies—scientists can uncover the underlying principles that govern machine intelligence. This approach aims to move beyond "black box" models toward systems that are interpretable and physically grounded, ensuring that AI tools used in high-stakes fields like chemistry and materials science are reliable and predictable.

To sustain this momentum, the report stresses the necessity of "centaur scientists"—polymaths who possess deep expertise in both their scientific discipline and advanced computing. MIT is already implementing these recommendations through joint faculty searches between the Department of Physics and the Schwarzman College of Computing. By fostering interdisciplinary PhD programs and shared data infrastructures, the goal is to create a new generation of researchers capable of navigating both the abstract world of algorithms and the concrete laws of physics.

MIT researchers are charting a transformative course for the intersection of artificial intelligence and the mathematical and physical sciences (MPS). Led by Professor Jesse Thaler, a new white paper advocates for a "two-way bridge" where AI doesn't just accelerate scientific discovery, but where scientific principles are used to refine and explain AI itself. This symbiotic relationship, termed the "Science of AI," focuses on how scientific reasoning can inform foundational model approaches and push the development of more robust algorithms.

The initiative highlights three specific domains: science driving AI, science inspiring AI, and science explaining AI. By treating neural networks as complex systems—much like particles or galaxies—scientists can uncover the underlying principles that govern machine intelligence. This approach aims to move beyond "black box" models toward systems that are interpretable and physically grounded, ensuring that AI tools used in high-stakes fields like chemistry and materials science are reliable and predictable.

To sustain this momentum, the report stresses the necessity of "centaur scientists"—polymaths who possess deep expertise in both their scientific discipline and advanced computing. MIT is already implementing these recommendations through joint faculty searches between the Department of Physics and the Schwarzman College of Computing. By fostering interdisciplinary PhD programs and shared data infrastructures, the goal is to create a new generation of researchers capable of navigating both the abstract world of algorithms and the concrete laws of physics.