Curlscape logo
Solution 05 · Surrogate Modeling

Surrogate models that predict in minutes

Surrogate models compress the design cycle by orders of magnitude — simulation results in minutes instead of hours or weeks — and finally put your legacy simulation archives to work as training data.

Where it fits

Built for recurring products and large problems

A surrogate pays back when the same kind of simulation runs again and again, or when a design space is too large to simulate every candidate.

01 / 03
Recurring product families
Train once on past runs, then predict each new variant in minutes.
02 / 03
Large design spaces
Search thousands of candidates with the surrogate, and verify the shortlist with full simulation.
03 / 03
Complex physics
Dedicated models for problems too slow to solve at every iteration.

Curlscape Sift — Surrogate Modeling & Model Order Reduction

Surrogate models compress the design cycle by orders of magnitude — simulation results in minutes instead of hours or weeks — and finally put your legacy simulation archives to work as training data.

Sift design explorer screen recording: dragging channel parameters on a cold plate and switching its topology, while the surrogate model redraws the temperature, pressure and velocity fields and updates the performance figures.

Design explorer: change a parameter, the surrogate predicts the flow

Sift Copilot screen recording: an engineer asks to minimise peak cell temperature and pressure drop for a cold plate. Copilot searches 32 designs, plots the trade-off, loads the best design in 3D and writes a report.

Copilot: state the goal, get the trade-off and a recommendation
01
Neural operator surrogates
Import simulation data, visualize, configure, train, and monitor in one interface. Leading architectures from NVIDIA and Emmi AI — DoMINO, AB-UPT, MeshGraphNet — in a single unified platform.
02
ML-based model order reduction
For edge deployments and cases where full 3D fields aren't needed: lightweight ML surrogates, plus digital twin models like prognostics and RUL estimation using Bayesian techniques.
03
Multi-fidelity modeling
Fuse many fast low-fidelity runs (e.g., BEMT for propeller design) with a few high-fidelity CFD runs via co-kriging, so the design space can be explored at low-fidelity cost.

AI Adoption Framework

The other paths

FAQ

Clearly answered

01How much simulation data do you need to train a surrogate model?

It depends on the physics and the size of the design space, and multi-fidelity methods cut the expensive part. In Curlscape's published studies, 10 fine-mesh CFD runs corrected a model built on cheap coarse-mesh runs, and 12 CFD runs corrected 2,560 blade-element evaluations. Existing simulation archives can also serve as training data.

Read the studies
02How accurate is a surrogate model compared with CFD?

Accurate inside the range of designs it was trained on, and Curlscape reports how accurate. In a published 1U-server thermal study, fusing fine and coarse CFD cut cross-validated error by 62–67% against the coarse mesh alone; in a propeller study, error fell 95% for thrust and 91% for torque. Key designs are still verified with a full simulation.

See the error tables
03Which surrogate architectures does Curlscape Sift support?

Curlscape Sift supports neural operator surrogates from NVIDIA and Emmi AI (DoMINO, AB-UPT and MeshGraphNet) in one platform, lightweight ML model order reduction for edge deployments, and multi-fidelity co-Kriging that fuses low- and high-fidelity runs.

How Sift works
04When is a surrogate model not worth building?

When a simulation runs only once or twice. A surrogate pays back when the same kind of simulation runs again and again, as with a recurring product family, or when the design space is too large to simulate every candidate.

How we run this ourselves

Schedule a demo