Calibration refinement for light field particle image velocimetry through iterative polynomial model tweaking

Volumetric calibration for light field particle image velocimetry (LF-PIV) establishes the voxel-to-pixel mapping and is essential for reliable particle field reconstruction. Conventional target-based calibration often degrades due to feature-localization errors and traverse uncertainties. We introduce an iterative polynomial model tweaking (IPT) strategy that improves accuracy, robustness, and generalization. A baseline polynomial model is built from calibrated-target features, and sparse flow tracers act as geometric anchors to refine the model under a regularized Levenberg–Marquardt scheme. View bias from LF sub-aperture images is explicitly modeled and compensated to decouple view-dependent offsets from geometric terms. Model stability is improved by expanding a high-confidence core to the full volume to capture higher-order aberrations. Comparative evaluations are performed against the baseline and a representative learning-based model (BP-NN) using synthetic data, and swirl flame flow experiments are conducted on a centrally staged combustor. Results indicate that IPT yields lower re-projection error than the baseline and BP-NN models and produces fewer lost and ghost particles in particle reconstructions. These gains become significant under perturbations to calibrated-target features and traverse distance. In swirl flow measurements, LF-PIV with IPT calibration achieves better agreement with benchmark planar PIV than the baseline calibration, characterized by a reduction in inlet-velocity overestimation and a smaller outward transverse bias.