A data-driven probabilistic learning approach for the prediction of controllable pitch propellers performance

Abstract

The multi-fidelity machine learning framework proposed in this paper leverages a probabilistic approach based on Gaussian Process modeling for the formulation of stochas- tic response surfaces capable of describing propeller performance for different mission profiles. The proposed multi-fidelity techniques will help coping with the scarcity of high-fidelity measure- ments by using lower-fidelity numerical predictions. The existing correlation of the multi-fidelity data sets is used to infer high-fidelity measurements from lower fidelity numerical predictions. The probabilistic formulations embedded in Gaussian Process regressions gives the unique op- portunity to learn the target functions describing propeller performance at different operating conditions, while quantifying the uncertainty associated to that specific prediction. While the multi-fidelity autoregressive scheme allows to construct high accurate response surfaces using only few experimental data, Uncertainty Quantification (UQ) provides an important metric to asses the quality of the learning process. We demonstrate the capability of the proposed frame- work to predict the performance of a controllable pitch propeller using few experimental data coming from towing tank experiments and many medium-fidelity predictions obtained using an in-house developed BEM, validated and verified in many previous studies.