An Accelerated Dual Proximal Gradient Method for Applications in Viscoplasticity

• Published in: arXiv.org 2016-09
• Main Authors: Treskatis, Timm, Moyers-Gonzalez, Miguel A, Price, Chris J
• Format: Article
• Language: English
• Subjects: Algorithms; Computational fluid dynamics; Constitutive models; Free boundaries; Iterative methods; Mathematical models; Nonlinear equations; Nonlinear systems; Optimization; Regularization; Two dimensional flow; Viscoplasticity
• ISSN: 2331-8422
• DOI: 10.48550/arxiv.1509.05084

We present a very simple and fast algorithm for the numerical solution of viscoplastic flow problems without prior regularisation. Compared to the widespread alternating direction method of multipliers (ADMM / ALG2), the new method features three key advantages: firstly, it accelerates the worst-case convergence rate from \(O(1/\sqrt{k})\) to \(O(1/k)\), where \(k\) is the iteration counter. Secondly, even for nonlinear constitutive models like those of Casson or Herschel-Bulkley, no nonlinear systems of equations have to be solved in the subproblems of the algorithm. Thirdly, there is no need to augment the Lagrangian, which eliminates the difficulty of choosing a penalty parameter heuristically. In this paper, we transform the usual velocity-based formulation of viscoplastic flow problems to a dual formulation in terms of the stress. For the numerical solution of this dual problem we apply FISTA, an accelerated first-order optimisation algorithm from the class of so-called proximal gradient methods. Finally, we conduct a series of numerical experiments, focussing on stationary flow in two-dimensional square cavities. Our results confirm that Algorithm FISTA*, the new dual-based FISTA, outperforms state-of-the-art algorithms such as ADMM / ALG2 by several orders of magnitude. We demonstrate how this speedup can be exploited to identify the free boundary between yielded and unyielded regions with previously unknown accuracy. Since the accelerated algorithm relies solely on Stokes-type subproblems and nonlinear function evaluations, existing code based on augmented Lagrangians would require only few minor adaptations to obtain an implementation of FISTA*.