par Ratz, Manuel ;Parente, Alessandro ;Mendez, M.A.
Référence Computers & fluids, 315, 107133
Publication Publié, 2026-07-01
Article révisé par les pairs
Résumé : Data-driven modal decompositions are powerful tools for compressing data and identifying dominant patterns. Standard methods like proper orthogonal decomposition (POD) and dynamic mode decomposition (DMD) assume data on a fixed spatial grid, which allows consistent inner products and thus correlation and projection across snapshots. However, modern experimental techniques such as Lagrangian particle tracking velocimetry and computational methods like smoothed particle hydrodynamics (SPH) generate scattered data that vary in space and time, making standard formulations hard to apply. This work proposes a meshfree framework based on radial basis function (RBF) representations to compute the inner products underlying modal decompositions. Building on an RBF regression of the data, an efficient quadrature formulation is derived in which inner products reduce to weighted products of the RBF coefficients, with precomputable integration operators. By selecting a basis that is constant in time, all snapshot correlations can be assembled through matrix operations on the RBF weights, resulting in a significant reduction in computational cost. Meshfree formulations of several widely used decompositions are developed, including POD, DMD, multi-scale POD, and variants of spectral POD. The mesh free formulations yield continuous and differentiable representations of the spatial structures. Two test cases are considered for benchmarking: particle image velocimetry measurements of an impinging jet and large eddy simulations of flow past a transitional airfoil. In both cases, the RBF-based approach outperforms classical binning and more accurately recovers relevant structures across a range of data densities.