An improved conjugate gradients method for quasi-linear Bayesian inverse problems, tested on an example from hydrogeology

• Verfasst von: Klein, Ole [VerfasserIn]
• Titel: An improved conjugate gradients method for quasi-linear Bayesian inverse problems, tested on an example from hydrogeology
• Verf.angabe: Ole Klein
• Jahr: 2021
• Umfang: 29 S.
• Teil: year:2021
• Teil: pages:357-385
• Teil: extent:29
• Fussnoten: First online: 02 December 2020 ; Gesehen am 29.06.2021
• Titel Quelle: Enthalten in: International Conference on High Performance Scientific Computing (7. : 2018 : Hanoi)Modeling, simulation and optimization of complex processes HPSC 2018
• Ausgabe Quelle: 1st ed. 2021
• Ort Quelle: Cham : Springer International Publishing, 2021
• Jahr Quelle: 2021
• Band/Heft Quelle: (2021), Seite 357-385
• ISBN Quelle: 978-3-030-55240-4
• DOI: doi:10.1007/978-3-030-55240-4_17
• Datenträger: Online-Ressource
• Sprache: eng
• K10plus-PPN: 1761490435

Abstract

We present a framework for high-performance quasi-linear Bayesian inverse modelling and its application in hydrogeology; extensions to other domains of application are straightforward due to generic programming and modular design choices. The central component of the framework is a collection of specialized preconditioned methods for nonlinear least squares: the classical three-term recurrence relation of Conjugate Gradients and related methods is replaced by a specific choice of six-term recurrence relation, which is used to reformulate the resulting optimization problem and eliminate several costly matrix-vector products. We demonstrate that this reformulation leads to improved performance, robustness, and accuracy for a synthetic example application from hydrogeology. The proposed prior-preconditioned caching CG scheme is the only one among the considered CG methods that scales perfectly in the number of estimated parameters. In the highly relevant case of sparse measurements, the proposed method is up to two orders of magnitude faster than the classical CG scheme, and at least six times faster than a prior-preconditioned, non-caching version. It is therefore particularly suited for the large-scale inversion of sparse observations.