A spectral-subspace-augmented POD-Galerkin method for parametrized PDEs with limited snapshot data

Jul 16, 2026

Parametrized partial differential equations (PDEs) appear in simulation, optimization, control, and uncertainty quantification, yet repeated full-order solves limit the number of high-fidelity snapshots available, especially in computational energy science for multiscale porous-media flow and transport. Proper orthogonal decomposition (POD) builds compact reduced bases but can show weak out-of-sample prediction when snapshots poorly cover the solution manifold. The SS-POD method enriches POD-Galerkin using a problem-adapted spectral approximation space, orthogonal subspace partitioning, and local POD with energy-balanced splitting. Coupled with DEIM for nonlinear cases, SS-POD improves out-of-sample accuracy while retaining compact bases, exemplified by a Laplace–Beltrami test on the unit sphere using five snapshots.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages29

Reading Time1 h 13 min

Introduction

Challenges of reduced-order modeling with limited snapshots

POD and reduced basis methods

Structured POD variants (SPOD, mPOD)

Proposed spectral-subspace augmentation approach

arXiv :2607 .01534v1 [math .NA] 1 Jul 2026 Noname manuscript No. (will be inserted by the editor) A spectral-subspace-augmented POD-Galerkin method for parametrized PDEs with limited snapshot data Tianhao Hu · Zecheng Gan Received: date / Accepted: date Abstract Parametrized partial differential equations (PDEs) arise in manyquery simulation, optimization, control, and uncertainty quantification, where repeated full-order solves restrict the number of high-fidelity snapshots that can be generated. This limitation is particularly pronounced in computational energy science, where multiscale models of porous-media flow, transport, and energy materials often make large snapshot datasets impractical. Proper orthogonal decomposition (POD) constructs compact reduced bases from solution snapshots, but it may exhibit limited out-of-sample predictive capability when the snapshots insufficiently sample the solution manifold. To address this limitation, we propose a spectral-subspace-augmented POD-Galerkin method (SS-POD) tailored to limited-data regimes. SS-POD starts from a problemadapted spectral approximation space, partitions it into orthogonal subspaces, and performs POD locally on the projected snapshot matrices. An energybalancing rule determines the spectral partition so that the resulting local POD problems are assigned comparable amounts of snapshot energy. For nonlinear parametrized PDEs, SS-POD is coupled with the discrete empirical interpolation method (DEIM) . Numerical experiments show that SS-POD improves out-of-sample accuracy over standard POD-Galerkin while retaining compact reduced bases in limited-snapshot regimes. In particular, for a Laplace–Beltrami problem on the unit sphere with only 5 snapshots, SS-POD achieves a relative error of 3 .9 × 10 −8 using 91 basis functions, whereas the standard POD error saturates at 7 .8 × 10 −4 and the spectral-Galerkin method requires 256 basis functions for comparable accuracy. These results indicate that Hu, T. Hong Kong University of Science and Technology (Guangzhou) E-mail: [thu176@connect.hkust-gz.edu.cn](thu176@connect.hkust-gz.edu.cn) [Gan](Gan), [Z](Z). Hong Kong University of Science and Technology (Guangzhou) E-mail: [zechenggan@hkust-gz.edu.cn](zechenggan@hkust-gz.edu.cn) SS-POD provides an effective strategy for high-fidelity reduced-order modeling from limited snapshot data. Keywords POD-Galerkin method · Reduced-order modeling · Parametrized PDEs · Spectral-subspace augmentation · Limited snapshot data Mathematics Subject Classification (2020) 65N35 · 65M60 · 65F55 · 41A46 1 Introduction Parametrized partial differential equations (PDEs) arise in optimization, control, inverse problems, and uncertainty quantification, where the same governing model must be solved for many parameter values [30] . High-fidelity finite element, finite difference, and spectral discretizations can make such manyquery studies computationally expensive [17,21,48] . Similar bottlenecks arise in computational energy science, including subsurface flow, transport, and data-driven energy-materials modeling, where simulations and high-fidelity data generation often span multiple scales [69]. Reduced-order models (ROMs) reduce this cost by separating an offline basis-construction stage from a lowdimensional online solve. This strategy is effective when the solution manifold admits a compact approximation and the offline data sufficiently identify the relevant solution components [3,30,51] . When only a few high-fidelity snapshots are available, however, the reduced space may fail to capture components that are essential for out-of-sample prediction. Reduced basis methods (RBMs) and proper orthogonal decomposition (POD) are two widely used approaches for constructing reduced spaces. RBMs select parameter samples adaptively, often through greedy algorithms guided by a posteriori error estimators [4,50,53,54,56,67] . In contrast, POD starts from a snapshot ensemble

FAQs

Why does standard POD-Galerkin often underperform with limited snapshot data?

Standard POD bases depend on the empirical snapshot distribution. When snapshots are sparse or uneven, the reduced space can reproduce sampled states but miss important components in unsampled regions, leading to large out-of-sample errors.

How does SS-POD improve POD-Galerkin in limited-data regimes?

SS-POD starts from a problem-adapted spectral approximation space, partitions it into orthogonal subspaces, and performs POD locally on projected snapshot matrices. An energy-balancing rule chooses the spectral partition so local POD problems receive comparable snapshot energy.

What role does DEIM play in SS-POD for nonlinear parametrized PDEs?

For nonlinear parametrized PDEs, SS-POD couples with the discrete empirical interpolation method (DEIM) to enable efficient handling of nonlinear terms during reduced-order computation.

Jul 16, 2026

Upload by Lucas Martin