{"title":"A multi-fidelity model management framework for multi-objective aerospace design optimisation","authors":"Ben Parsonage, C. Maddock","doi":"10.3389/fpace.2023.1046177","DOIUrl":null,"url":null,"abstract":"This paper presents a multi-fidelity meta-modelling and model management framework designed to efficiently incorporate increased levels of simulation fidelity from multiple, competing sources into early-stage multidisciplinary design optimisation scenarios. Phase specific/invariant low-fidelity physics-based subsystem models are adaptively corrected via iterative sampling of high(er)-fidelity simulators. The correction process is decomposed into several distinct parametric/non-parametric stages, each leveraging alternate aspects of the available model responses. Globally approximating surrogates are constructed at each degree of fidelity (low, mid, and high) via an automated hyper-parameter selection and training procedure. The resulting hierarchy drives the optimisation process, with local refinement managed according to a confidence-based multi-response adaptive sampling procedure, with bias given to global parameter sensitivities. An application of this approach is demonstrated via the aerodynamic response prediction of a parametrized re-entry vehicle, subjected to a static/dynamic parameter optimisation for three separate single-objective problems. It is found that the proposed data correction process facilitates increased efficiency in attaining a desired approximation accuracy relative to a single-fidelity equivalent model. When applied within the proposed multi-fidelity management framework, clear convergence to the objective optimum is observed for each examined design optimisation scenario, outperforming an equivalent single-fidelity approach in terms of computational efficiency and solution variability.","PeriodicalId":365813,"journal":{"name":"Frontiers in Aerospace Engineering","volume":"25 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2023-02-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"1","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Frontiers in Aerospace Engineering","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.3389/fpace.2023.1046177","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 1
Abstract
This paper presents a multi-fidelity meta-modelling and model management framework designed to efficiently incorporate increased levels of simulation fidelity from multiple, competing sources into early-stage multidisciplinary design optimisation scenarios. Phase specific/invariant low-fidelity physics-based subsystem models are adaptively corrected via iterative sampling of high(er)-fidelity simulators. The correction process is decomposed into several distinct parametric/non-parametric stages, each leveraging alternate aspects of the available model responses. Globally approximating surrogates are constructed at each degree of fidelity (low, mid, and high) via an automated hyper-parameter selection and training procedure. The resulting hierarchy drives the optimisation process, with local refinement managed according to a confidence-based multi-response adaptive sampling procedure, with bias given to global parameter sensitivities. An application of this approach is demonstrated via the aerodynamic response prediction of a parametrized re-entry vehicle, subjected to a static/dynamic parameter optimisation for three separate single-objective problems. It is found that the proposed data correction process facilitates increased efficiency in attaining a desired approximation accuracy relative to a single-fidelity equivalent model. When applied within the proposed multi-fidelity management framework, clear convergence to the objective optimum is observed for each examined design optimisation scenario, outperforming an equivalent single-fidelity approach in terms of computational efficiency and solution variability.