A deep learning method for non-uniform flow field based on KAN and MLP neural networks

Abstract Many flow fields exhibit pronounced non-uniformity, especially when initial conditions vary over time. Although physics-informed neural networks have advanced flow-field reconstruction, precision in centeracterizing spatiotemporal non-uniformity remains inadequate. To address this, we propose a confidence-weighted, locally reinforced learning (LRL) strategy that strengthens the extraction of local features. Validations on a fluid–structure interaction (FSI) benchmark and a triangular artificial-reef flow show that LRL better resolves local details and delivers higher-accuracy reconstructions. Building on this, we further examine the Kolmogorov–Arnold Network (KAN) for flow reconstruction and find that its accuracy is governed primarily by structure-dependent configurations rather than network size; in its vanilla form, KAN underperforms, motivating a hybrid KAN + MLP model that improves accuracy at the expense of increased computational cost. This study foregrounds the proposed LRL methodology, systematically evaluates the potential and limitations of KAN for complex-flow modeling, and provides guidance for neural-architecture design in fluid-mechanics applications.