A deep learning model can now approximate days-long nuclear reactor accident simulations in under 60 seconds.
Researchers trained a surrogate model on output from ASTEC, the standard code used to study severe accidents in nuclear plants. The model pairs an AutoEncoder — which compresses the simulation's high-dimensional data by a factor of over 300 — with a Neural Ordinary Differential Equation that steps through time. Together they predict roughly 80 physical variables simultaneously, covering thermal-hydraulics, core degradation, and fission product release, and hold stable for up to 50,000 time steps. The training data covers two accident classes: station blackouts and loss-of-coolant accidents.
The practical payoff is significant for operator training. ASTEC runs can take days, which rules out any interactive use. A surrogate that finishes a 40-hour simulated scenario in under a minute on ordinary CPU hardware changes what is possible in a simulator room. That matters because operator error in the early hours of a severe accident is exactly what training programs exist to prevent.
The authors frame this as a first look at what deep learning can and cannot do with ASTEC's notoriously non-linear physics — meaning the hard validation work is still ahead, and no one is proposing to replace the full code with a neural network anytime soon.