PresCast: Physics-Constrained Fourier Kolmogorov-Arnold Networks for Bluff Bodies Spatiotemporal Wall Pressure Forecast

Junle Liu, Yanyu Ke, Haoyan Li, Wenliang CHEN, Tianle Niu, Tim K.T. Tse, Gang Hu

Keywords: Spatiotemporal forecasting, Bluff body aerodynamics, Physics-constrained deep learning, Fourier neural networks, Urban infrastructure resilience

Abstract: Understanding and forecasting wall pressure on bluff bodies is crucial for assessing urban structural safety and infrastructure resilience. While recent studies have explored deep learning for predicting mean and fluctuating pressure coefficients, most existing frameworks are restricted to single-snapshot predictions and lack physical interpretability. This work introduces PresCast, a physics-constrained Fourier Kolmogorov-Arnold neural network (FKAN) designed for spatiotemporal wall pressure forecasting. FKAN integrates a Fourier neural operator within the Kolmogorov-Arnold architecture, and incorporates a physics-constrained frequency loss to guide the training process and enhance generalization to high-frequency dynamics. Wind tunnel experiments were conducted on a classical bluff body, a rectangular cylinder of side ratio 1.5 under zero angle of attack, using high-frequency pressure scanning, building a comprehensive dataset for training and validation. Results demonstrate that PresCast achieves accurate spatiotemporal forecasting of wall pressure, capturing both statistical properties and physical signatures, including temporal fluctuations, power spectral density, and spatial distribution. This study highlights the potential of physics-constrained deep learning frameworks for advancing urban aerodynamics and resilient infrastructure design.

Submission Number: 73