A new AI model reconstructs the internal physics of nuclear reactor components in real time - without a single sensor inside them.
Researchers have published a framework called MIMONet that learns to map sparse boundary measurements to full internal field states inside physically sealed or hostile environments. The target is nuclear-grade thermal-fluid systems: pressurized water reactor subchannels and heat exchangers where stuffing in physical sensors is either structurally impossible or would compromise the system. Tested across problems of escalating complexity, MIMONet stays below 5% relative error and runs inference in 0.35 milliseconds on an NVIDIA H200, consuming 46 millijoules per heat-exchanger pass. It holds those numbers even when sensor readings are corrupted by up to 50% noise.
The architectural choice that makes this possible is framing the problem as operator learning: instead of fitting a model to a fixed grid or a specific reactor geometry, MIMONet learns the solution operator itself, letting it generalize across configurations without per-instance retraining. That matters because existing methods typically require either precise governing equations written out by hand or retraining from scratch whenever the geometry changes - neither is practical at deployment scale. With nuclear energy back in serious conversation as a power source for data-center buildouts and grid decarbonization, the ability to monitor reactor internals without redesigning instrumentation for every new plant design is an actual engineering problem, not a theoretical one.
The paper is careful to call its results "simulation-based feasibility" - all validation used simulated data, not live reactor measurements. Experimental validation and cross-solver testing remain future work. That gap, between a compelling research result and something a nuclear operator would stake a formal safety case on, is still real.