Towards autonomous discovery of stiff structures in CFD Design of Experiment space, from random-walk automaton steered by the “spiral” criterion

• We study parameter-wise stiffness (canyons) for high-Reynolds configurations.
• A DoE refinement criterion is based on the local stiffness of the CFD problem.
• We perform random walks across the DoE space, directed by the local stiffness.
• On a model canyon, this stochastic automaton scores far better than uniformity.

We seek to compute the response surface – applicable e.g. to Uncertainty Quantification (UQ) and control laws for flight – of aerodynamic observables in CFD solutions for high Reynolds flows, in cases displaying stiff evolutions in the Design of Experiment space (DoE). Combined with another relevant parameter (e.g. the angle of attack or a parameter of the turbulence model) in a 2D DoE, this stiffness in parameter space may then sometimes be visualised as a ridge- or canyon-like structure of the response surface. For adequate precision in the computation of the response surface, we thus need to first search the DoE for such structures. We have chosen to base the search algorithm not on an a priori hypothesis but on our original “spiral” criterion, which we deem to be a measure of the local stiffness of the problem. An semi-autonomous discovery of “stiff” structures is then accomplished through an automaton steered by this criterion, using smaller steps and increased displacement variance in potential Regions of Interest (RoIs). This is tested on a synthetic “canyon”, using several Factors of Merit for comparison of the automaton performance relative to a uniform grid of points.