Building a Nuclear-Inspired Digital Twin with V-JEPA

Inside the simulator, frozen video models, physics residuals, and experiments that showed why good anomaly scores can still produce bad alarms.

The simulated loop across five scenarios. Tank color represents temperature and moving pipe particles represent flow. Labels and onset markers are report annotations, not video-model inputs. Figure: nuclear-vjepa-twin project.

A pump slows down, circulation falls, and temperatures begin to change. That is the kind of event I want a monitoring system to notice. Now imagine that the process stays healthy, but the colors in its camera view change. A useful system should distinguish those situations. My nuclear-inspired V-JEPA twin shows how easily an anomaly detector can confuse them.

Disclosure: This article was drafted with AI assistance and checked against the project’s code and saved experiment artifacts. The figures come from the project’s simulator and recorded evaluations.

I built nuclear-vjepa-twin to investigate a practical question: what can pretrained video representations contribute when a monitoring system already has sensor measurements and a simple physics model?

The result is a research prototype with synthetic episodes, three pretrained V-JEPA encoders, anomaly detectors, state estimation, public-data benchmarks, and a local dashboard. Its most useful finding is also its least promotional: adding a large video model did not automatically improve…