The generation of high-quality anisotropic meshes remains a critical challenge in large-scale computational mechanics and computational fluid dynamics (CFD), particularly for industrial applications involving complex geometries and wide parametric spaces. While adaptive meshing techniques based on error estimation can produce near-optimal meshes, their computational cost and reliance on repeated solution cycles limit their practicality for design exploration and rapid simulation of unseen configurations. This contribution presents a data-driven framework for the prediction of near-optimal anisotropic mesh spacing functions for CFD simulations involving previously unseen operating conditions and geometric configurations. The proposed approach exploits the extensive availability of high-fidelity industrial simulation data to learn anisotropic spacing information directly from prior adapted solutions. Target spacing is represented using metric tensors, which provide a mathematically rigorous description of local mesh resolution and directional anisotropy. An artificial neural network is trained to map problem parameters, such as geometric descriptors and flow conditions, to the corresponding anisotropic metric fields. The trained model outputs metric tensors at the nodes of a coarse background mesh, providing a common reference domain across different geometries. A robust transfer strategy is introduced to apply the predicted anisotropic spacing to unseen configurations. This includes background mesh morphing and metric interpolation, enabling the automatic generation of high-quality meshes without further adaptive refinement. The influence of network hyperparameters and training dataset composition on prediction accuracy is systematically investigated. Accuracy is assessed through quantitative comparisons with reference near-optimal metrics, as well as through the quality and suitability of the generated meshes for subsequent CFD simulations. The methodology is demonstrated on a challenging industrial-scale application involving a fully parameterised aircraft configuration with up to 11 geometric parameters. Results show that the predicted meshes retain the essential anisotropic features required for accurate flow resolution, while significantly reducing the meshing and computational cost associated with traditional adaptive workflows.