A pretrained AI model that had never seen a jet engine got noticeably better at predicting engine failure once researchers told it how the engine's parts are physically wired together.
A new study benchmarked five pretrained time-series foundation models (TSFMs) - general-purpose AI models trained to forecast sequences of data - on C-MAPSS, the standard simulated dataset used to predict a jet engine's remaining useful life (RUL), meaning how many more cycles it can run before failing. Models that tracked multiple sensor readings at once beat single-sensor versions by a wide margin, especially when engines operated under changing conditions. The researchers then fed the models structural information from a digital twin, a virtual map of how an engine's components connect, and used it to restrict which sensors the model could cross-reference, rather than letting it freely compare every sensor to every other one. An ablation study across C-MAPSS subsets of differing complexity showed that pretraining, task-specific tuning, and this topology constraint each added value on their own, and stacking all three beat any single approach.
The sales pitch for foundation models has always been: drop a general-purpose model into a new domain and skip the expensive custom engineering. This result complicates that pitch a little - the frozen model only became competitive once it was handed real structural knowledge about the hardware, not just raw sensor streams. That's a useful signal for anyone building predictive maintenance tools: digital-twin topology isn't just documentation, it's apparently a usable input for better predictions.
The accuracy bump from adding topology was real but small, a useful reminder that claims of a foundation model understanding your factory still come with some assembly required.