Advanced Statistical Calibration via Stochastic Variational Gaussian Process Model
International Journal on Advanced Science, Engineering and Information Technology · 14.10.2026 · 16 (5) · 1565-1574
DOI: 10.18517/ijaseit.16.5.14566 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal on Advanced Science, Engineering and Information Technology, online erschienen am 14.10.2026. Verfasst von Yonggwan Shin und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Statistical calibration determines tuning parameters that align simulator outputs with physical observations. The conventional approximate nonlinear least squares estimation (ANLSE) framework employs a Gaussian process (GP) emulator as a surrogate model; however, the computational complexity of arising from GP covariance matrix inversion can become impractical as the dataset size grows. This study proposes SVGP-ANLSE, a calibration method that replaces the standard GP emulator with a stochastic variational Gaussian process (SVGP) model. By introducing inducing variables and optimizing variational evidence lower bound (ELBO) through stochastic gradient descent, SVGP-ANLSE reduces computational complexity to , where denotes the number of inducing points , which becomes increasingly advantageous as datasets grow beyond the regime where standard GP is tractable. The Kullback-Leibler divergence term within the ELBO regularizes the variational posterior toward the prior, reducing overfitting and improving numerical stability. The method was evaluated on five synthetic test functions involving two to six tuning parameters and up to twelve total dimensions (input variables plus tuning parameters), with 50 replicated datasets per function, and validated on empirical data from the Baldur tokamak fusion simulation model. Across all test functions, SVGP-ANLSE yielded the smallest average Euclidean distance and RMSPE among the three methods compared, with smaller variability across replications, consistently achieving improvements of 20-60% (approximately 41% on average) relative to GP-ANLSE and Kriging-ANLSE. On the empirical tokamak dataset, SVGP-ANLSE achieved prediction accuracy competitive with established calibration methods, including SMLE and Min-max. These results indicate that SVGP-ANLSE is a viable calibration method whose scalability advantage is expected to increase with dataset size.
Zitieren
APA
Shin, Y., Bae, T. H., Heo, Y., & Seo, Y. A. (2026). Advanced Statistical Calibration via Stochastic Variational Gaussian Process Model. International Journal on Advanced Science, Engineering and Information Technology, 16(5), 1565-1574. https://doi.org/10.18517/ijaseit.16.5.14566
Vancouver
Shin Y, Bae TH, Heo Y, Seo YA. Advanced Statistical Calibration via Stochastic Variational Gaussian Process Model. International Journal on Advanced Science, Engineering and Information Technology. 2026;16(5):1565-1574. doi:10.18517/ijaseit.16.5.14566
BibTeX
@article{Shin2026,
author = {Shin, Yonggwan and Bae, Tae Hwan and Heo, Yujun and Seo, Yun Am},
title = {Advanced Statistical Calibration via Stochastic Variational Gaussian Process Model},
journal = {International Journal on Advanced Science, Engineering and Information Technology},
year = {2026},
volume = {16},
number = {5},
pages = {1565--1574},
doi = {10.18517/ijaseit.16.5.14566}
}
EUVIMED