Turbomachinery design has become a simulation-driven process, permanently confronted to the dual need to reduce the cycle time and to further integrate complexity and multiple physics. This duality pushes towards high-dimensional design spaces and favours a multisimulation environment that assesses different operating points, different disciplines, or even different fidelity levels. To manage CPU cost, surrogate-based optimisation (SBO) has become an established approach. One of the key enablers for efficient SBO is being able to avoid the regions in the design space where simulation failures occur, and this is of a particular interest in a multisimulation environment since the failure of a single computation does not necessarily imply the systematic failing of the other computations.
The current work presents innovative auto-adaptive surrogates, exploiting a blend of interpolation/regression and classification, that have been implemented in the integrated optimisation platform Minamo. A multisimulation demonstrator, based on NASA Rotor 37, has been set up to perform an aero-mechanical multi-point optimisation. With one of the objectives of the optimisation to improve the stall margin, the aerodynamic simulations are forced to flirt with the numerical stability limits. It is shown that the introduction of different success models has a beneficial impact on the course of the multisimulation optimisation. The improved model quality, obtained by using partial information, has allowed for a more efficient search process, and at the same time a better global success rate has been obtained.