AI Isn't Just Answering Physics Questions Anymore — It's Running the Experiments

Two new superconductors and a neural network that models colliding neutron stars: this summer's most interesting AI news didn't come from a product launch. It came from two physics papers, published quietly a few weeks apart, that show machine learning quietly changing jobs — from research assistant to embedded lab instrument.

While the AI headlines this year have mostly been about agents, coding assistants, and enterprise deployment, a smaller and arguably more consequential story has been unfolding in condensed matter physics and nuclear astrophysics. In both fields, researchers have started building neural networks directly into the machinery of discovery itself — not to write up the results, but to make previously impossible searches computationally tractable.

The Superconductor Hunt

In early July, an international team working within the SuperC consortium — led by Professor Paivi Torma at Aalto University, with sample synthesis handled by Professor Emilia Morosan's group at Rice University — announced two newly confirmed superconducting compounds: YRu3B2 and LuRu3B2. Both draw their superconductivity from an unusual source: electrons forming "flat bands" within a kagome lattice, a hexagonal atomic arrangement named after a Japanese basket-weaving pattern that geometrically frustrates electron movement in ways that can favor superconductivity.

The compounds themselves are notable, but Torma has been clear that the real story is the method. Of the more than 7,000 known superconductors, only around 20 were ever theoretically predicted in advance — historically, superconductor discovery has been a slow, one-material-at-a-time slog of synthesis and testing. The SuperC team flipped that process: a machine-learning model pre-screens enormous swaths of possible element combinations for the electronic signatures associated with flat-band superconductivity, and only the most promising candidates get pushed through to expensive, targeted quantum-mechanical calculations and physical synthesis.

"Our method uses machine-learning-based pre-screening followed by targeted calculations on promising candidates," Torma said of the approach. "This approach will greatly speed up superconductor discovery in the future." The team estimates the method could eventually let them process material combinations numbering in the billions, rather than the hundreds or thousands a human-directed search could realistically attempt.

It's worth being precise about what was and wasn't achieved here. YRu3B2 and LuRu3B2 superconduct at temperatures far too cold for any practical application — nowhere near room temperature. The SuperC consortium, formed in 2023, has a stated goal of finding a room-temperature superconductor by 2033, a material Torma says would "forever change the way we consume energy" by cutting waste in everything from power grids to the increasingly electricity-hungry data centers running today's AI models. These two compounds aren't that breakthrough. They're proof that the search method works — a validated first catch from a net that can now be cast far wider than before.

Modeling the Universe's Heavy-Element Factory

A few weeks earlier, a separate team at Germany's GSI/FAIR nuclear physics facility published a different kind of AI-for-physics tool, this one aimed at one of astrophysics' most computationally punishing problems: modeling how colliding neutron stars forge the universe's heaviest elements.

When two neutron stars merge, the collision briefly creates conditions extreme enough to trigger the r-process — rapid neutron capture — which is responsible for producing roughly half of all elements heavier than iron, including gold, platinum, and uranium. Simulating that process in detail requires tracking the energy released by thousands of simultaneous nuclear reactions inside a churning, expanding cloud of merger debris, which is so computationally expensive that most hydrodynamic simulations have historically used crude approximations of the heating involved.

The GSI/FAIR team, led by Dr. Oliver Just, built a deep learning model called RHINE (r-process heating implementation in hydrodynamic simulations with neural networks) to close that gap. Rather than calculating every nuclear reaction network from scratch during a simulation, RHINE was trained on an extensive library of full reference calculations and now estimates r-process heating rates on the fly, plugging directly into hydrodynamic merger simulations as they run.

"Our new model, RHINE, which uses artificial intelligence, offers an efficient alternative" to the oversimplified heating models simulations have relied on until now, Just explained. Because r-process heating directly shapes how fast merger debris expands, how it's distributed, and what electromagnetic signal it produces as a kilonova, a more accurate — and dramatically cheaper — heating model means astrophysicists can run more, and more detailed, simulations of events that telescopes like LIGO and Virgo are watching for in real time. The team published the work in Physical Review D and released the RHINE source code publicly, inviting other groups to build on it directly.

The Pattern Underneath Both Papers

Put side by side, these are two very different papers in two unrelated fields, but they share a structural idea that's easy to miss if you only follow AI news from the product side. In neither case is the neural network doing the physics. In the superconductor search, the underlying quantum mechanical theory is exactly the same theory physicists have used for decades — the model just decides where to point that theory next. In RHINE, the underlying nuclear reaction physics doesn't change either — the network is a fast, learned approximation standing in for a calculation that used to be too expensive to run in full, every time, everywhere.

That's a meaningfully different role for AI than the one dominating this year's headlines. It's not an agent taking actions, not a chatbot answering questions, not even a research assistant drafting a paper — a role this publication covered back on July 7 when we looked at Anthropic's agentic research workbench for reproducible science. It's closer to a new kind of instrument: a component wired directly into the experimental or computational pipeline, whose entire job is to make an intractable search space tractable enough that human scientists and their existing theories can do the rest.

If that pattern holds, it suggests where the AI-for-science bottleneck moves next. It won't be raw compute, and increasingly it won't be theoretical understanding either — both of these teams built on physics nobody had to reinvent. The bottleneck becomes the quality and coverage of the training data used to build the approximator: the reference calculations RHINE learned from, the labeled examples that taught SuperC's pre-screening model what a promising flat-band candidate looks like. Get that curation wrong, and the model searches confidently in the wrong direction across a space too large for anyone to double check by hand.

Neither of these papers made much noise outside their fields. No product launch, no funding round, no keynote. But if you're trying to figure out where AI actually earns its keep in 2026, two teams of physicists quietly building neural networks into their own instruments might be a better signal than anything on a conference stage.

Related reading: our July 7 piece on Anthropic's agentic research workbench and the industry's bet on reproducible AI-assisted science covers a different, complementary angle on AI's role in research.

Sources

Aalto University, researchers identify new superconductors: https://www.aalto.fi/en/news/researchers-identify-new-superconductors-unlocking-process-that-could-yield-thousands-more

ScienceDaily coverage of the superconductor discovery: https://www.sciencedaily.com/releases/2026/07/260701205006.htm

Physical Review Research publication for superconductivity in YRu3B2 and LuRu3B2: https://journals.aps.org/prresearch/abstract/10.1103/lpqj-7hyg

GSI/FAIR, understanding neutron star mergers with artificial intelligence: https://www.gsi.de/en/start/news/details?tx_news_pi1%5Bnews%5D=6265

Phys.org coverage of the RHINE neutron-star model: https://phys.org/news/2026-06-neutron-star-merger-simulations-gain.html

ScienceDaily coverage of the RHINE neutron-star model: https://www.sciencedaily.com/releases/2026/06/260626030426.htm

Physical Review D publication for RHINE: https://link.aps.org/doi/10.1103/gl2l-7f3g