Modern aircraft design requires efficient prediction of complex, time-dependent flow phenomena. While high-fidelity datasets offer accuracy, they come with high computational costs. To address this, NATO’s research group AVT-351 is investigating reduced-order models (ROMs). Prope
...
Modern aircraft design requires efficient prediction of complex, time-dependent flow phenomena. While high-fidelity datasets offer accuracy, they come with high computational costs. To address this, NATO’s research group AVT-351 is investigating reduced-order models (ROMs). Proper Orthogonal Decomposition (POD) is widely used in ROMs, but truncated POD bases struggle with sharp gradients and shocks. The enriched POD-LSTM neural network ROM
(ePOD-LSTM-ROM) has demonstrated improved accuracy for airfoil pressure distributions
and was extended to full wing surfaces. This approach was effective for sharp discontinuities
moving significantly in time. For realistic cases, such as the ONERA M6 wing with multiple shocks, domain decomposition allowed for multiple enrichments in separate subdomains,
but at the cost of rapidly increasing degrees of freedom. To improve LSTM-NN accuracy,
enrichments were optimized using mutual information, filtering, and error trade-offs, reducing the total ePOD-LSTM-ROM error by 23.8%. Finally, a goal-oriented reduced-order model
(GOROM) was developed to alter POD modes optimized for a specific goal function, improving
shock accuracy by 5.0% with minor losses elsewhere.