Machine learning is already recovering equations, exposing unexpected forces and designing experiments. The harder question is whether AI can move from finding patterns to creating genuinely new physical theories.
THE UNIVERSAL RECORD
Sourced reporting. No opinions.
Brad Socha | September 23, 2026 | 5:19 AM EST
Artificial intelligence can already perform something remarkably close to scientific discovery. Give specialized AI experimental measurements, physical constraints and enough computational freedom, and some systems can infer mathematical relationships that researchers did not explicitly provide.
That is no longer purely hypothetical. Researchers have used machine learning to recover known equations, identify unexpected behaviour in laboratory plasma, solve an open mathematical problem connected to theoretical physics and explore experimental designs that human researchers might not have considered.
But none of those achievements means an AI has independently produced the next theory of relativity.
The distinction goes to the heart of how physics advances. Finding an equation that accurately describes data is one task. Recognizing a previously unknown principle of nature, explaining why it works and devising experiments capable of proving it wrong is considerably harder.
AI Can Already Extract Physics From Data
One of the clearest demonstrations came from MIT researchers Silviu-Marian Udrescu and Max Tegmark. Their AI Feynman system tackled a problem known as symbolic regression: finding mathematical equations that explain numerical data.
When tested against 100 equations drawn from The Feynman Lectures on Physics, AI Feynman recovered all 100. The system exploited properties common in physical laws, including symmetry, dimensional relationships and separability, to reduce complicated problems into simpler ones.
There is an important limitation: those equations were already known. The experiment demonstrated that an algorithm could rediscover physical relationships from data, not that it had uncovered previously unknown fundamental laws.
More recent research has moved closer to that boundary.
In a 2025 study published in Proceedings of the National Academy of Sciences, researchers at Emory University trained a physics-constrained neural network on three-dimensional trajectories of particles inside laboratory “dusty plasma”, an ionized gas containing charged dust particles.
The model inferred the forces acting between particles with an R² greater than 0.99. More importantly, the researchers reported significant departures from commonly used theoretical assumptions about particle charge and the distance over which interactions are screened. The predictions were checked against laboratory experiments rather than accepted solely because the machine produced them.
The work was later awarded a 2025 PNAS Cozzarelli Prize in the physical and mathematical sciences. The National Academy of Sciences said the research showed how machine learning could help develop physical laws for complicated many-body systems.
That is much closer to AI-assisted discovery: the system was not merely reproducing an equation placed in its training benchmark. It extracted previously unknown interactions from real experimental observations.
AI Is Beginning to Search for Discoveries
The role of AI is also expanding beyond interpreting existing measurements.
A 2026 Nature review examined systems capable of designing physics experiments by searching enormous numbers of possible arrangements of lasers, detectors and other components. The researchers reported that AI-designed configurations can challenge conventional experimental designs while matching or exceeding the performance of human-designed setups in some settings.
This matters because new physics depends not only on analyzing data, but on deciding what measurements should be made in the first place.
Another development arrived in theoretical physics this year. Michael Brenner, Vincent Cohen-Addad and David Woodruff reported a neuro-symbolic system combining Google’s Gemini Deep Think with tree search and numerical feedback. According to their preprint, the system derived exact analytical solutions to an open mathematical problem involving gravitational radiation from cosmic strings and produced several different solution methods. The work is currently a preprint, so its status should be distinguished from peer-reviewed experimental findings.
Other researchers are testing whether AI can search much larger theoretical spaces. A 2026 preprint describing a system called Albert reported that it was given historical particle-physics data that contained no direct observation of the top quark. The system reconstructed the Standard Model structure and inferred the need for a top quark with a mass broadly consistent with later measurements. Because the actual physics is already known today, this remains a reconstruction experiment rather than the discovery of a new particle.
Together, these projects point toward a possible scientific pipeline in which AI proposes mathematical relationships, searches theoretical possibilities, designs experiments and interprets the resulting measurements.
That would be substantially different from using machine learning simply to classify data.
The Physics Problem AI Has Not Solved
The ultimate test is not whether AI can find patterns. It is whether it can recognize a principle that scientists have never conceived.
Einstein did not derive general relativity by fitting increasingly complicated curves to existing measurements. The theory required new conceptual ideas about gravity, geometry, space and time.
Researchers are now explicitly testing whether AI can reproduce that kind of intellectual leap.
One proposal, discussed this year by Google DeepMind chief executive Demis Hassabis, is to train a model only on scientific knowledge available before a historical cutoff and determine whether it can independently reconstruct later breakthroughs such as general relativity. Nature reported in September that early experiments with these “historical” models have exposed substantial limitations in present-day systems.
A July 2026 analysis by researchers including Marin Soljačić of MIT and Ido Kaminer of Technion argues that modern AI has become exceptionally capable at prediction but may be moving away from the type of principle-driven reasoning responsible for physics’ deepest theories. The authors identify abilities such as formulating the right scientific questions, inventing new principles and designing falsifiable tests as important missing pieces. The work is a preprint and presents an interpretation of the field rather than an established scientific finding.
There is another fundamental problem: correlation can masquerade as a law.
A sufficiently flexible algorithm may find an equation that fits every measurement scientists currently possess and still fail when conditions change. Physics therefore demands more than predictive accuracy. Candidate laws must survive independent experiments, make successful predictions beyond the data used to construct them and remain consistent with other established observations, or provide compelling evidence for why established theory needs revision.
Human scientists remain central to that process. Even the Emory plasma system incorporated physical constraints chosen by researchers, while scientists interpreted and experimentally validated what the neural network found.
The most realistic near-term future may therefore be neither AI replacing physicists nor physicists merely using AI as a faster calculator.
It could be a partnership in which machines explore mathematical and experimental possibilities on a scale humans cannot, while scientists determine which patterns represent meaningful physics and subject them to experimental tests.
AI has already demonstrated that it can rediscover known laws and expose unexpected relationships in real physical systems. Whether a machine can eventually originate something comparable to relativity, quantum mechanics or an experimentally verified theory beyond the Standard Model remains unanswered.
If it does, the defining moment will not be when an AI generates an impressive equation. It will be when nature agrees.
Sources:
Science Advances — AI Feynman: A Physics-Inspired Method for Symbolic Regression
https://www.science.org/doi/10.1126/sciadv.aay2631
MIT DSpace — AI Feynman: A Physics-Inspired Method for Symbolic Regression
https://dspace.mit.edu/entities/publication/ad578bc8-57eb-4b7d-84fe-0da95829c1e3
Proceedings of the National Academy of Sciences — Physics-Tailored Machine Learning Reveals Unexpected Physics in Dusty Plasmas
https://www.pnas.org/doi/10.1073/pnas.2505725122
Emory University — AI Reveals Unexpected New Physics in Dusty Plasma
https://news.emory.edu/features/2025/07/ai-reveals-unexpected-new-physics-dusty-plasma
National Academy of Sciences — PNAS Announces Six 2025 Cozzarelli Prize Recipients
https://www.nasonline.org/news/pnas-2025-cozzarelli-recipients/
Nature — Designing Physics Experiments With Artificial Intelligence
https://www.nature.com/articles/s41586-026-10898-6
TU Wien — Artificial Intelligence Suggests New Physics Experiments
https://www.tuwien.at/en/all-news/news/artificial-intelligence-suggests-new-physics-experiments
Google Research — Solving an Open Problem in Theoretical Physics Using AI-Assisted Discovery
https://research.google/pubs/solving-an-open-problem-in-theoretical-physics-using-ai-assisted-discovery/
arXiv — Solving an Open Problem in Theoretical Physics Using AI-Assisted Discovery
https://arxiv.org/abs/2603.04735
arXiv — Autonomous Discovery of Particle Physics Theories From Experimental Data
https://arxiv.org/abs/2603.28935
Nature — The Einstein Test: What Happens When AI Tries to Rediscover Relativity?
https://www.nature.com/articles/d41586-026-02804-x
arXiv — Can AI Follow in Einstein’s Footsteps?
https://arxiv.org/abs/2607.27794
About the Author
Brad Socha is the founder of The Universal Record, focused on sourced, factual global reporting. Coverage includes international news, geopolitics, technology, and major developments.