Machine-learning surrogates are increasingly used to accelerate computational fluid dynamics, yet progress is limited by the lack of benchmarks capturing realistic, time-dependent turbulent flows. This thesis introduces a 13 TB dataset of high-fidelity implicit large-eddy simulations of three-dimensional turbulent wakes behind super-elliptical cylinders. Unlike existing datasets, it captures three-dimensional turbulence with an active energy cascade driven by vortex stretching, combining (i) 380 long-horizon trajectories of 400 time steps each with 3–9 million points per frame on irregular meshes, (ii) systematic variation across geometry, Reynolds number, and angle of attack, and (iii) a temporal resolution that preserves the full inertial subrange of the turbulent energy spectrum. Building on this dataset, state-of-the-art neural operators are evaluated across three prediction tasks of increasing complexity: mean-field prediction from governing parameters, the inverse pressure problem, and long-horizon autoregressive spatio-temporal forecasting. Neural operators accurately recover mean flow fields, yet fail progressively as the target fields gain high-frequency content: fine-scale spatial structure is systematically suppressed in instantaneous flow fields, and all evaluated architectures collapse for temporal predictions. These failure modes are attributed to several concurrent mechanisms: memory constraints forcing sparse point-cloud subsampling, latent-space compression discarding high-frequency spatial content, and a spectral bias of the mean-squared-error objective that de-prioritises the high-frequency residuals dominant in turbulent flows. Together, these results provide a new standard benchmark for turbulent CFD surrogates and show that progress requires methodological advances beyond architecture scaling, including physically motivated loss functions and latent representations capable of retaining fine-grained spatial content. ... Read more