Multifidelity cross-entropy estimation of conditional value-at-risk for risk-averse design optimization

Anirban Chaudhuri, Benjamin Peherstorfer, Karen Willcox

Research output: Chapter in Book/Report/Conference proceedingConference contribution


There is an increasing demand for designs in aerospace engineering that guarantee baseline performance even in limit states, i.e., outside of nominal operating conditions of vehicles. Typically, the numerical optimization for such risk-averse designs is computationally challenging because in each iteration of the optimization loop the performances of the designs are estimated for the rare events corresponding to the limit states. This work proposes a multifidelity approach to make tractable the optimization of large-scale risk-averse designs that are based on the conditional value-at-risk (CVaR) as the risk measure. The multifidelity method leverages low-cost, low-fidelity models to speed up the CVaR estimation in each iteration of the risk-averse optimization to reduce the runtime compared to traditional Monte Carlo estimators that rely on the high-fidelity models alone. At the same time, the proposed approach makes occasional recourse to the expensive high-fidelity model to guarantee convergence to design points that satisfy the high-fidelity optimality conditions. In numerical experiments with an aerostructural design problem, the multifidelity approach achieves speedups of almost one order of magnitude compared to a traditional single-fidelity method.

Original languageEnglish (US)
Title of host publicationAIAA Scitech 2020 Forum
PublisherAmerican Institute of Aeronautics and Astronautics Inc, AIAA
ISBN (Print)9781624105951
StatePublished - 2020
EventAIAA Scitech Forum, 2020 - Orlando, United States
Duration: Jan 6 2020Jan 10 2020

Publication series

NameAIAA Scitech 2020 Forum
Volume1 PartF


ConferenceAIAA Scitech Forum, 2020
Country/TerritoryUnited States

ASJC Scopus subject areas

  • Aerospace Engineering


Dive into the research topics of 'Multifidelity cross-entropy estimation of conditional value-at-risk for risk-averse design optimization'. Together they form a unique fingerprint.

Cite this