Design and control of an open-source fan array wind tunnel for adaptive flow profiling using deep neural networks

Fan array wind tunnels (FAWTs) enable programmable, spatially structured inflows for aerodynamic and civil-engineering studies, but reproducible research is hindered by facility-specific implementations and the lack of open benchmarks for learning-based control. This work presents an open-source FAWT platform and benchmarking suite motivated by inverse flow control but focused on forward-surrogate learning on a modular 6 × 4 fan array with multi-distance velocity measurements. We release the hardware designs, control firmware and software interfaces, and a large-scale experimental dataset, together with a unified training/test protocol for systematic model comparison. Using this benchmark, we evaluate three representative neural architectures (MLP, CNN2D, CNN2D+Attention) that map fan commands to downstream velocity fields, and we report performance across seven downstream planes (18–78 cm). We further introduce simple forward baselines for downstream mixing (identity/affine, ridge regression, and a shared nonnegative local mixing kernel) to contextualize spatial coupling and provide interpretable reference points. Beyond aggregate accuracy metrics, we report multi-seed validation at the reference distance (78 cm) with uncertainty intervals and paired significance tests between architectures. We also analyze a salient failure mode of unconstrained regression on this dataset: systematic bias in zero-velocity regimes, which motivates physics-guided constraints for reliable closed-loop deployment. Overall, the paper contributes a reproducible dataset-and-code benchmark and an empirical characterization of how spatial inductive bias and minimal physical structure affect performance on small FAWT arrays.