When time is critical, running computationally expensive simulations (e.g. 3D CFD or detailed chemistry analysis) may not be practical. MoDS offers several fast-response surrogates that capture the behaviour of the underlying model, but which can be evaluated in a tiny fraction of the time. The following surrogate forms are available:
- Polynomial
- Kriging
- High Dimensional Model Representation (HDMR)
- Deep Kernel Learning (with optional auto-tuning of hyperparameters)
- Thin-plate spline interpolation
- Inverse distance weighting interpolation
HDMR is a technique well suited to models with many inputs/outputs.
- Outputs are approximated by combinations of ortho-normal basis functions.
- Surrogate construction is significantly faster than for a polynomial of the same order.
- Global sensitivities of each output to each input follow naturally from the expansion coefficients.
Surrogate models can optionally be evaluated via APIs written in the following languages:
- Java
- MATLAB/Octave
- gPROMS
