CFD models often contain several imprecisely specified parameters that must be calibrated or inferred from measurements. Unfortunately, methods for such inference struggle when the number of parameters is large. The benefits of dimension reduction cannot be overstated. If one is able to approximate a model with hundreds of inputs by a comparable interface with a handful of derived inputs, then several otherwise intractable techniques (e.g., optimization, inverse calibration, response surface construction) become possible.
I will discuss our research efforts and progress on active subspaces for dimension reduction, including tests that discover if such dimension reduction is possible and strategies to exploit it when present.