Hybrid neural interpolation of a sequence of wind flows

Rapid and accurate wind field prediction is essential for modeling particle transport in emergency scenarios. Traditional Computational Fluid Dynamics (CFD) approaches are too slow for real-time applications, necessitating surrogate models. We develop a hybrid neural interpolation method for constructing surrogate models that interpolate steady-state Reynolds-Averaged Navier–Stokes (RANS) solutions across varying inlet wind angles on a canonical test geometry. Our approach combines Tucker tensor decomposition with neural networks to interpolate RANS solutions across varying inlet wind angles. The method decomposes high-dimensional velocity, pressure, and eddy viscosity field datasets into a core tensor and factor matrices, then uses Fourier interpolation for angular modes and k-nearest neighbors convolution for spatial interpolation. A neural network correction mitigates interpolation artifacts introduced by the Fourier stage. We validate the approach as a controlled proof-of-concept on a cylinder-sphere configuration that exhibits nonlinear wake interactions. Relative to a strong pure neural network (NN) benchmark, the hybrid model achieves comparable accuracy (R2>0.99) with significantly reduced training time. The pure NN remains a feasible reference model; the hybrid provides an accelerated approximate alternative that maintains wake dynamics, with computational efficiency that motivates future extension to larger urban domains.