A physics-oriented review of neural-network-based modeling for bluff-body flows

Abstract

Bluff-body flows combine early separation, coherent wake organization, vortex shedding, and strong sensitivity to geometry, Reynolds number, inflow condition, and three-dimensional transition. These features make them both canonical fluid-mechanics benchmarks and demanding targets for neural-network-based modeling. This review provides a representative, physics-oriented synthesis of neural-network approaches for bluff-body flows. Instead of organizing the field by architecture or prediction task, the review asks which wake mechanisms current models represent and which validation evidence makes those claims trustworthy. It examines how current models represent separation, wake topology, pressure-wake coupling, coherent shedding, turbulence statistics, closure behavior, and flow-control mechanisms. We discuss data sources, flow representations, model classes, validation practices, and practical uses in surrogate prediction, reconstruction, monitoring, control, and digital-twin-type workflows. The review emphasizes that low numerical error by itself does not establish physical learning, and argues for validation based on topology, spectra, force response, uncertainty, generalization, and interpretable mechanism evidence.