Watch a surrogate learn a CFD case
A surrogate model is a fast stand-in for a slow calculation. Take computational fluid dynamics (CFD), which predicts how air or liquid flows around a shape: one run can take hours. A surrogate learns from a handful of those runs, then predicts a new case in seconds. Here is the whole method on that one example.
What goes in
A set of CFD runs you already have. Each pairs the inputs you can change, geometry and flow conditions, with the output you want: the full velocity field. Often tens of runs, not thousands.
What the model learns
STOCHOS fits the relationship between those inputs and the field, and learns how sure it should be in each region from where the runs actually landed.
What comes out
Give it a new case and it returns the predicted field in seconds, with confidence attached. In the clip, the prediction matches the solver and the error stays near zero.
Where it is trusted
Inside the range it learned from, the prediction holds and its confidence says so. Outside that range, confidence drops. The solver still validates the design you commit to.
How STOCHOS builds a surrogate
STOCHOS uses DIM-GP, a model family built for small, noisy datasets, changing geometries, and multi-fidelity data. Every prediction carries quantified uncertainty, so a reliable interpolation is easy to tell from a risky guess. It handles fields, geometries, and meshes, not just tidy tables.
How the DIM-GP model works →See where a surrogate stays reliable with uncertainty quantification, and cut its data cost with multi-fidelity modeling.
- Predict a full field, not a single numberKeep the spatial answer across the geometry.
- Which of your parameters actually matterRank the inputs before the next campaign.
- AI for EngineeringWhere this fits in CFD, FEM, and CAE work.
- High fidelity digital twins using AIA multi-scale composite case study.
- The science behind STOCHOSThe research the method is built on.
A surrogate answers in seconds, which is why the demo can score every attempt the moment you make it. Run the demo
Common questions
01Is a surrogate model the same as the solver?
No. A solver computes the physics from scratch at high cost. A surrogate learns that input-output behavior from existing runs and predicts new cases in seconds. It cuts how often you run the solver, which still validates the results that matter.
02How much data does a surrogate model need?
There is no fixed number. STOCHOS is built for small, expensive datasets, often tens of runs rather than thousands. It handles noisy and multi-fidelity data, and flags where confidence is low so you can add runs where they help most.
03Can a surrogate replace CFD or FEM?
It replaces many repeated runs inside the range it was trained on, but not the solver itself: you still need that for training data, new regimes, and final verification. A surrogate is an acceleration layer, not new physics.
04What is a surrogate model?
A surrogate model is a fast, data-driven stand-in for an expensive simulation or experiment. Trained on runs you already have, it predicts new cases in seconds, and in STOCHOS it reports the confidence of each prediction, so you know where the shortcut can be trusted.
