Surrogate models and DoE
A surrogate model learns from a manageable number of simulations or tests how the system responds. After that you can assess further variants without computing them again.
- How many sample points does the model need?
- Which parameters have the largest influence?
- Where in the parameter space is the optimum?
From test plan to model
Design of experiments gives structured coverage of the parameter space. Instead of evaluating each point on its own, we train a model on these data that predicts values in between, shows sensitivities and narrows down optima.
Knowing the limits matters: a surrogate is only reliable inside the range it was trained on. That is why we show quality metrics such as R² and cross-validation and mark where additional data points are needed.
Where it pays off
Wherever a single run takes hours and many variants need assessing: crash, thermal, flow, injection molding, material design.
Related projects
Predicting residual wall thickness in water-assisted injection molding
A neural network predicts the residual wall thickness of a hollow plastic part from its process parameters. Run in reverse, it returns the process setting needed for a target wall thickness. A project with Plastic Innovation.
→ View projectDesigning EV charging cables with phase change material
Fast charging drives high currents through cables and battery systems, and they heat up. An ML model built on FEM and prototype data sizes the conductor for a given maximum temperature, using phase change material as a passive heat buffer.
→ View projectDoes this match your question?
Send us a short description. We will tell you whether and how we can help.