An Incremental Singular Value Decomposition Approach for Large-Scale Spatially Parallel & Distributed but Temporally Serial Data -- Applied to Technical Flows

TL;DR

This paper introduces an incremental SVD method for efficiently analyzing large-scale, time-evolving 3D CFD data, enabling significant data compression with minimal loss of accuracy in complex flow simulations.

Contribution

It proposes a novel incremental SVD framework with a bunch matrix for efficient, scalable data reduction in large-scale, parallel CFD simulations, applicable to various computational science problems.

Findings

Achieved approximately 95% data compression with 1% error in force and wave elevation.

Validated the approach on flow around a cylinder, hydrofoil, and a container ship in waves.

Demonstrated low computational overhead (~10%) for large snapshot matrices.

Abstract

The paper presents a strategy to construct an incremental Singular Value Decomposition (SVD) for time-evolving, spatially 3D discrete data sets. A low memory access procedure for reducing and deploying the snapshot data is presented. Considered examples refer to Computational Fluid Dynamic (CFD) results extracted from unsteady flow simulations, which are computed spatially parallel using domain decomposition strategies. The framework addresses state of the art PDE-solvers dedicated to practical applications. Although the approach is applied to technical flows, it is applicable in similar applications under the umbrella of Computational Science and Engineering (CSE). To this end, we introduce a bunch matrix that allows the aggregation of multiple time steps and SVD updates, and significantly increases the computational efficiency. The incremental SVD strategy is initially verified and validated by simulating the 2D laminar single-phase flow around a circular cylinder. Subsequent studies analyze the proposed strategy for a 2D submerged hydrofoil located in turbulent two-phase flows. Attention is directed to the accuracy of the SVD-based reconstruction based on local and global flow quantities, their physical realizability, the independence of the domain partitioning, and related implementation aspects. Moreover, the influence of lower and (adaptive) upper construction rank thresholds on both the effort and the accuracy are assessed. The incremental SVD process is applied to analyze and compress the predicted flow field around a Kriso container ship in harmonic head waves at Fn = 0.26 and ReL = 1.4E+07. With a numerical overhead of O(10%), the snapshot matrix of size O(R10E+08 x 10E+04) computed on approximately 3000 processors can be incrementally compressed by O(95%). The storage reduction is accompanied by errors in integral force and local wave elevation quantities of O(1E-02%).