FAQ

Common questions.

It is a neural network trained on two things at once: the measured data, and the laws of physics governing the system. Predictions that contradict the physics are penalized during training, so the model learns solutions that both fit the data and make physical sense.
A regular network learns only the statistical patterns in its training data, so it needs many examples and can produce impossible answers outside them. A PINN also has the governing equations built into its training objective, which lets it learn from far less data and keeps its predictions physically consistent even in conditions it has never observed.
Yes, and that is one of their main selling points. The physics acts as a substitute for missing data: the equations constrain what the model can predict everywhere, including regions with no measurements. The original Raissi, Perdikaris and Karniadakis work demonstrated solutions to nonlinear differential equations from small data sets.
Because the hardest problems in maintenance modeling are scarce failure data, poor generalization to new conditions, and unexplainable alerts. Encoding failure physics lets a model reason about failure modes it has few or no examples of, transfer across similar assets, and justify its warnings in engineering terms.
The governing physics must be known well enough to encode, training can be hard for stiff or multi-scale problems, and the physics terms make training more expensive. In practice most deployments are hybrids that use physics constraints where the equations are reliable and learned components elsewhere.
To the top
READY WHEN YOU ARE

See physics-informed reasoning on your own asset.

Bring one critical machine. We'll show you how Vermilion models its physics, what it predicts, and the explanation behind the prediction. Thirty minutes, with a reliability engineer.