Sitting 1,700 miles from the reactor it controlled, an AI loop trimmed power on a 64-year-old Indiana teaching reactor in real time, the first autonomous adjustment ever recorded in the US

The reactor core sat in its pool in West Lafayette, Indiana, emitting the faint blue glow it has cast for six decades.

Nobody in the building was touching the controls.

The adjustment came from Idaho, routed through a cloud environment in Virginia, guided by an AI model that had learned the reactor’s physics by simulating them thousands of times.

DeepLynx functions as the nervous system of the arrangement, translating the AI’s output into a physical command the reactor’s instruments could execute.

When the power level nudged back into line, the correction was logged, confirmed, and quietly became the first autonomous remote power adjustment ever recorded at a US nuclear reactor.

What makes this moment matter is not its scale but its shape, and that shape points directly at the generation of reactors now being licensed.

How an AI taught itself to steer a reactor it had never touched

The team first demonstrated remote power adjustments through an automatic adjustment system called a digital control loop, then made the loop more autonomous using a reinforcement learning model running in software that simulates how physical forces interact within the reactor. Reinforcement learning works the way a chess engine learns openings: the model plays millions of trial runs inside a virtual copy of the reactor, collecting penalties for straying too far from the target power level until it finds sequences of control rod movements that keep the core stable.

The virtual copy is the critical piece. On the reactor’s digital backbone, the team built what Purdue describes as the first live digital twin of a US reactor. That twin runs in real time, fed by actual sensor readings, so the AI is always working from current conditions rather than a static textbook model of how a reactor should behave.

The system analyzes conditions, predicts outcomes and adjusts autonomously, acting as a smart bridge that links virtual models to physical systems without overriding essential safety protections. The reactor’s own licensed safety controls sit above everything the AI can reach, a hard ceiling the software cannot cross.

The Indiana reactor that became a long-distance test bench

The demonstration linked PUR-1, Purdue’s 64-year-old research reactor, with Idaho National Laboratory and a cloud system in a three-site loop spanning Indiana and Idaho. Specifically, the demonstration linked three sites operating together in real time: high-performance computing systems in Idaho, the PUR-1 reactor in Indiana, and a cloud environment in Virginia.

During the demonstration, the team calculated and delivered instructions for the movement of an auxiliary control rod through INL’s DeepLynx, an advanced data and control platform, using cloud-based connectivity with the PUR-1 digital twin and INL high-performance computing systems. DeepLynx functions as the nervous system of the arrangement, translating the AI’s output into a physical command the reactor’s instruments could execute.

PUR-1 had been converted in 2019 to the first fully digital instrumentation and control system licensed by the US Nuclear Regulatory Commission, a DOE-funded conversion that gave researchers a platform for testing control architectures that future small modular reactors and microreactors are expected to use. That earlier conversion is what made July’s demonstration physically possible: without a fully digital control backbone, there would have been no interface for the remote system to reach.

What was actually measured and when it happened

On July 13, 2026, researchers from Idaho National Laboratory, the Grainger College of Engineering at the University of Illinois Urbana-Champaign, and Purdue University used a geographically distributed control system to remotely and automatically adjust the power of a research reactor in real time, with the reactor’s own safety systems retaining full control throughout.

From Idaho Falls, the team fine-tuned reactor power to minimize small power fluctuations and maintained steady reactor operation. The power involved is tiny by commercial standards: PUR-1 is a research reactor rated well below one megawatt, operated as a teaching tool rather than a grid-connected generating unit. But the significance lies in the loop itself, not in the kilowatts it touched.

This milestone builds on digital twin technology originally demonstrated in 2023 and reactor secure communications demonstrated the following year, a deliberate three-step architecture: secure the link, build the twin, then let the AI act. This approach aligns with Nuclear Regulatory Commission requirements by operating alongside, not in place of, the reactor’s safety control systems.

Where the demonstration runs into harder ground

The lead researcher is careful about what the demonstration actually proved. According to Stylianos Chatzidakis, the Purdue assistant professor and associate PUR-1 director who led the work, “We showed that you can achieve remote monitoring and remote control of a reactor,” but he qualifies it clearly: “This is an experiment. We didn’t actually remotely control the reactor.” The distinction matters: the AI issued instructions, but licensed human operators and hard-wired safety circuits remained the final authority over the physical machine at every moment.

Scaling this architecture to a commercial reactor would demand a far more demanding regulatory path. Every sensor link, every cloud handshake, and every AI inference step would need to satisfy NRC cybersecurity rules written for a world where the biggest threat to a control system was a disgruntled insider, not a distributed software agent operating from another time zone. None of that work has been done yet, and the researchers say so.

There is also the question of what happens when a cloud connection drops mid-adjustment. PUR-1’s low power level means a communication failure carries almost no consequence. A 1,200 MW pressurized water reactor operating near full load is a different problem entirely, and the Idaho team has not claimed otherwise.

Why the shape of this experiment matters for small modular reactors

Digital control and operation would allow advanced reactors to be operated remotely, aligning with the planned remote locations of some small modular reactors and microreactors, and would also allow for AI monitoring and real-time measurements of reactor performance. Many of the microreactor designs now before the NRC are intended for deployment at military bases, remote mines, or Arctic communities where a resident licensed operator would be impractical to maintain full time.

That is precisely the gap this experiment was designed to address. If a 10 MW microreactor sitting above the Arctic Circle can be monitored, diagnosed, and adjusted from a control room in the continental United States, the economics of remote deployment change completely. The reactor becomes, in effect, a remotely operated industrial asset rather than a facility that must carry its own full crew.

“This advancement greatly expands the kinds of experiments and control system research we can perform at PUR-1,” said Stylianos Chatzidakis, assistant professor and associate reactor director at Purdue. That expansion is already visible in the research pipeline: the same digital twin and DeepLynx platform are being positioned as the test bed for future SMR control studies, and the July demonstration gives those follow-on projects a verified baseline rather than a theoretical one. For a curious look at how a different kind of energy innovation crossed a physics threshold through patient incremental steps, the ceramic superconductivity record broken in a Texas lab follows a structurally similar arc. The open question for the reactor work is how quickly regulators can develop a cybersecurity framework that treats a cloud-connected AI not as a threat to be excluded, but as a tool to be licensed, much as early digital I&C itself once had to be, a process that took nearly a decade at PUR-1 alone. Consider also the parallel challenge of training a new workforce for these architectures, an effort already underway through programs like the Aberdeen Power Academy preparing the next generation of grid engineers. The six-decade-old teaching reactor in Indiana has, improbably, become the country’s most advanced test bench for the nuclear control rooms of the next half-century.