Meshless data-driven decompositions with RBF-based inner products Journal: Computers and Fluids
M. Ratz, A. Parente, M. A. Mendez · 2025
Read Meshless data-driven decompositions with RBF-based inner products Journal: Computers and Fluids by M. Ratz, A. Parente, M. A. Mendez (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-03264.txt.
Introduction
Manuel Ratz, Alessandro Parente, and Miguel Alfonso Mendez free two classic data-splitting tools from the need for neat grids. Methods that pull out main patterns from flow data usually assume samples sit evenly in space and time. Particle simulations and particle tracking in labs break that rule, scattering points unevenly as they move.
Their fix fits the scattered data with smooth radial basis shapes, then uses those smooth forms to do the math the pattern tools require. Holding the basis fixed in time cuts cost sharply. Tests show the mesh-free route can recover the leading patterns where sample counting fails. The approach opens pattern study to a wide class of particle-based data that grids cannot serve.
Catalog record
| CREATOR | M. Ratz, A. Parente, M. A. Mendez |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 79 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-03264
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Meshless data-driven decompositions with RBF-based inner products Journal: Computers and Fluids M. Ratz, A. Parente, M. A. Mendez Abstract Data-driven modal decompositions are useful tools for compressing data or identifying dominant structures. Popular ones like the dynamic mode decomposition (DMD) and the proper orthogonal decomposition (POD) are defined with continuous inner products. These are usually approximated with samples of data uniform in space and time. However, not every dataset fulfills this requirement. Numerical simulations with smoothed particle hydrodynamics or experiments with Lagrangian particle tracking velocimetry produce scattered data varying in time and space, rendering sample-based inner products impossible. In this work, we extend a previous approach that computes the modal decompositions with meshfree radial basis functions (RBFs). We regress the data and use the continuous representation of the RBFs to compute the required inner products. We choose our basis to be constant in time, greatly reducing the computational cost since the inner product of the data reduces to the inner product of the basis functions. We use this approach in the most popular decompositions, namely the POD, DMD, multi-scale POD, and the two versions of the spectral POD. For all decompositions, the RBFs give a mesh-free representation of the spatial structures. Two test cases are considered: particle image velocimetry measurements of an impinging jet and large eddy simulations of the flow past a transitional airfoil. In both cases, the RBF-based approach outperforms classical binning and better recovers relevant structures across all data densities. 1 Introduction Data-driven modal decompositions such as proper orthogonal decomposition (POD) and dynamic mode decomposition (DMD) are widely used to extract dominant structures and dynamics from complex systems. These techniques have been extensively applied to obtain compact representations of high-dimensional data across diverse fields, including fluid mechanics 2 , 53 , 68 , image processing and filtering 40 , face recognition 27 , reacting flows 48 , structural fault detection 13 and vibration 15 , biomechanics 74 , neuroscience 44 , epidemiology 59 to mention but a few. Despite their wide applicability, the algorithms used to compute data-driven modal decompositions are traditionally formulated under the assumption that data are available on a spatial grid—as in image processing, experimental techniques such as particle image velocimetry (PIV, 49 ), or standard numerical frameworks like finite difference, finite volume, and finite element methods 29 . This grid may be structured or unstructured, dense or sparse, but is typically assumed to be fixed in time. Such an assumption simplifies computing the inner products inherent to any linear decomposition 38 , from classical methods like POD [ 33 , 32 , 61 ] and DMD [ 53 , 55 ] to more recent variants such as spectral POD (SPOD) 60 , 73 and multi-scale POD (mPOD) 36 , 37 . The spatial structures on this discrete grid also allow to link dominant features to specific physical locations. A variant within this context is the so-called gappy scenario, where an underlying grid or spatial structure is still assumed, but the available data have values missing in space at scattered locations in time. A common strategy in this setting is to resort to interpolation 67 , radial basis functions 12 , or low-rank matrix completion 5 to reconstruct the missing data either prior to or during the computation of a modal decomposition. Examples of such approaches include gappy POD 45 , 11 , 21 , the recently proposed gappy Spectral POD 43 , and interpolative variants of DMD designed to handle non-uniform data in space 41 or time 30 , 62 . Alternatively, when interpolation or regression methods become ineffective due to excessive gappiness, compressed sensing techniques 10 , 4 can be employed to exploit sparsity and directly reconstruct the dominant flow structures from limited measurements, enabling full-field recovery from observations restricted to a sparse subset of grid points. These ideas have opened the path toward compressed sensing formulations of POD 35 , 80 and DMD 4 , 14 . A more significant departure from the gappy framework arises when the data are not simply incomplete, but intrinsically scattered , with no underlying spatial grid or fixed set of sampling locations. This situation arises in Lagrangian computational methods, such as smoothed particle hydrodynamics 42 or vortex methods 28 , where the computational particles themselves move with the flow. Similarly, in experimental techniques like Lagrangian particle tracking velocimetry 46 , 54 , 69 , 57 , measurements are obtained at the positions of tracer particles advected by the fluid, leading to data sampled at locations varying in both space and time. To the authors’ knowledge, there is no established framework for applying modal decompositions directly to data in their scattered, Lagrangian form. The most common strategy is to circumvent the irregularity by mapping the data onto a fixed spatial grid—typically through interpolating, projecting, or binning techniques—and then proceeding with traditional grid-based decomposition methods 1 , 56 , 66 . While this preprocessing step enables to use established tools like POD (e.g. 79 ) and DMD (e.g. 34 ), it introduces interpolation errors and may obscure important features of the original scattered data, particularly in regions with sparse sampling or rapid deformation. More advanced binning strategies iterate between mapping the data to the grid and computing the POD and progressively improve the spatial modes 7 , 20 . Within the image velocimetry community, significant effort has recently been devoted to methods for mapping scattered data onto fixed grids 18 , 26 , 58 . More recently, however, a meshless paradigm based on radial basis function (RBF) has begun to emerge 64 , 63 , 65 , 31 , 51 , 52 , aiming to remove grids entirely. These approaches are considered meshless since they operate directly on scattered data without requiring a structured or unstructured mesh, and do not assume a fixed spatial discretization for computing derivatives, integrals, or more complex tasks such as solving partial differential equations 12 , 82 , 6 , 17 . Recently, the meshless RBF framework was extended to compute a POD directly from scattered data 70 . The approach is conceptually related to functional principal component analysis (FPCA) [ 50 , 77 , 22 ] , which generalizes traditional principal component analysis (PCA) to the case where each observation is treated as a continuous function rather than a finite-dimensional vector. The key difference, among others discussed in 70 , is that the FPCA uses interpolation methods to approximate an infinite-dimensional eigenvalue problem, whereas the meshless POD retains the classical POD structure but replaces the grid-based inner product with RBF-based quadrature. This work extends the framework introduced in 70 in several important respects. First, we replace the interpolative RBF formulation with a regression-based approach and substitute the Gauss–Legendre quadrature used to assemble the temporal correlation matrix with a fully RBF-based quadrature scheme. This modification not only reduces computational cost by enabling the reuse of precomputed weight matrices but also improves flexibility in handling complex geometries. Second, we introduce an RBF-based projection strategy operating directly on the continuous functional form obtained from the RBF approximation of the snapshots, enabling efficient evaluation of the spatial structure of each mode. Finally, the proposed framework is generalized beyond classical POD to encompass more advanced modal