TY - GEN
T1 - Gaussian processes and support vector regression for uncertainty quantification in aerodynamics
AU - Palar, Pramudita Satria
AU - Zakaria, Kemas
AU - Zuhal, Lavi Rizki
AU - Shimoyama, Koji
AU - Liem, Rhea Patricia
N1 - Funding Information:
Pramudita Satria Palar and Lavi Rizki Zuhal were funded in part through the Penelitian Dasar administered by Direktorat Riset dan Pengabdian Masyarakat-Direktor Jenderal Penguatan Riset dan Pengembangan-Kementerian Riset dan Teknologi / Badan Riset dan Inovasi nasional, Republik Indonesia. Part of the work was also carried out under the Collaborative Research Project 2020 of the Institute of Fluid Science, Tohoku University.
Publisher Copyright:
© 2021, American Institute of Aeronautics and Astronautics Inc, AIAA. All rights reserved.
PY - 2021
Y1 - 2021
N2 - This paper investigates the performance of two kernel-based surrogate models, namely Gaussian Process regression (GPR) and support vector regression (SVR), for solving uncertainty quantification (UQ) problems in aerodynamics. This research aims to shed light on both surrogate models’ approximation performance to get better insight for practical purposes. To that end, experiments using various kernel functions were performed to study their impact on GPR and SVR accuracy. Besides, the use of a composite kernel learning technique is also studied. Computational experiments show that GPR with Matern-5/2 is the most robust technique when an individual kernel is used. However, SVR with the Matern-5/2 kernel also performs better than GPR in some problems. The results suggest that there is no single best performing method when averaged over all sets of problems. Finally, we also demonstrated that using composite kernel learning, provided sufficient data samples, can further reduce the approximation error for both GPR and SVR.
AB - This paper investigates the performance of two kernel-based surrogate models, namely Gaussian Process regression (GPR) and support vector regression (SVR), for solving uncertainty quantification (UQ) problems in aerodynamics. This research aims to shed light on both surrogate models’ approximation performance to get better insight for practical purposes. To that end, experiments using various kernel functions were performed to study their impact on GPR and SVR accuracy. Besides, the use of a composite kernel learning technique is also studied. Computational experiments show that GPR with Matern-5/2 is the most robust technique when an individual kernel is used. However, SVR with the Matern-5/2 kernel also performs better than GPR in some problems. The results suggest that there is no single best performing method when averaged over all sets of problems. Finally, we also demonstrated that using composite kernel learning, provided sufficient data samples, can further reduce the approximation error for both GPR and SVR.
UR - http://www.scopus.com/inward/record.url?scp=85100388138&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85100388138&partnerID=8YFLogxK
M3 - Conference contribution
AN - SCOPUS:85100388138
SN - 9781624106095
T3 - AIAA Scitech 2021 Forum
SP - 1
EP - 12
BT - AIAA Scitech 2021 Forum
PB - American Institute of Aeronautics and Astronautics Inc, AIAA
T2 - AIAA Science and Technology Forum and Exposition, AIAA SciTech Forum 2021
Y2 - 11 January 2021 through 15 January 2021
ER -