Multi-stage multi-fidelity information correction for artificial neural network based meta-modelling

Multi-fidelity meta-modelling has become a popular means of efficiently distributing computational resource across various levels of simulation fidelity to obtain numerically accurate predictions of an expensive function. Such techniques have significant potential within an engineering design paradigm incorporating either many-query analyses or outer-loop applications. This paper presents a hybrid parametric/non-parametric information correction method incorporating the sequential application of several distinct stages within an artificial neural network based surrogate framework. The proposed methodology may be used to correct any domain encompassing set of low-fidelity input/output correspondence using a small subset of high-fidelity samples. A global surrogate can then be generated via a double loop ANN hyper-parameter selection and training procedure. To demonstrate the effectiveness of the proposed metamodelling approach, the aerodynamic response prediction of a parametrized waverider-based re-entry vehicle is examined. Results suggest that the incorporation of multiple corrective stages leveraging low-fidelity data can offer significant improvements in computational efficiency when modelling the expensive high fidelity function compared with single stage correction. The costs to achieve global accuracy are examined and compared across single/multi-stage variants, with consideration given to surrogate construction and evaluation. Results are compared both with the low-fidelity approximation and a surrogate of the true response built using only the high-fidelity samples available to the corrective method.