TY - GEN
T1 - Neural Network-Based Sequential Global Sensitivity Analysis Algorithm
AU - Liu, Yen Chen
AU - Leifsson, Leifur
AU - Koziel, Slawomir
AU - Pietrenko-Dabrowska, Anna
N1 - Funding Information: Acknowledgements. This material is based upon work supported by the United States National Science Foundation under grant no. 1739551. Publisher Copyright: © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2022
Y1 - 2022
N2 - Performing global sensitivity analysis (GSA) can be challenging due to the combined effect of the high computational cost, but it is also essential for engineering decision making. To reduce this cost, surrogate modeling such as neural networks (NNs) are used to replace the expensive simulation model in the GSA process, which introduces the additional challenge of finding the minimum number of training data samples required to train the NNs accurately. In this work, a recently proposed NN-based GSA algorithm to accurately quantify the sensitivities is improved. The algorithm iterates over the number of samples required to train the NNs and terminates using an outer-loop sensitivity convergence criteria. The iterative surrogate-based GSA yields converged values for the Sobol’ indices and, at the same time, alleviates the specification of arbitrary accuracy metrics for the NN-based approximation model. In this paper, the algorithm is improved by enhanced NN modeling, which lead to an overall acceleration of the GSA process. The improved algorithm is tested numerically on problems involving an analytical function with three input parameters, and a simulation-based nondestructive evaluation problem with three input parameters.
AB - Performing global sensitivity analysis (GSA) can be challenging due to the combined effect of the high computational cost, but it is also essential for engineering decision making. To reduce this cost, surrogate modeling such as neural networks (NNs) are used to replace the expensive simulation model in the GSA process, which introduces the additional challenge of finding the minimum number of training data samples required to train the NNs accurately. In this work, a recently proposed NN-based GSA algorithm to accurately quantify the sensitivities is improved. The algorithm iterates over the number of samples required to train the NNs and terminates using an outer-loop sensitivity convergence criteria. The iterative surrogate-based GSA yields converged values for the Sobol’ indices and, at the same time, alleviates the specification of arbitrary accuracy metrics for the NN-based approximation model. In this paper, the algorithm is improved by enhanced NN modeling, which lead to an overall acceleration of the GSA process. The improved algorithm is tested numerically on problems involving an analytical function with three input parameters, and a simulation-based nondestructive evaluation problem with three input parameters.
KW - Global sensitivity analysis
KW - Neural networks
KW - Sobol’ indices
KW - Surrogate modeling
KW - Termination criteria
UR - https://www.scopus.com/pages/publications/85134342953
U2 - 10.1007/978-3-031-08757-8_37
DO - 10.1007/978-3-031-08757-8_37
M3 - Conference contribution
SN - 9783031087561
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 445
EP - 454
BT - Computational Science - ICCS 2022, 22nd International Conference, Proceedings
A2 - Groen, Derek
A2 - de Mulatier, Clélia
A2 - Krzhizhanovskaya, Valeria V.
A2 - Sloot, Peter M.A.
A2 - Paszynski, Maciej
A2 - Dongarra, Jack J.
PB - Springer Science and Business Media Deutschland GmbH
T2 - 22nd Annual International Conference on Computational Science, ICCS 2022
Y2 - 21 June 2022 through 23 June 2022
ER -