This study improved the capture efficiency and robustness of an industrial Local Exhaust Ventilation (LEV) arrangement. Using CFD augmented by an AI-enabled surrogate model, we explored contaminant transport from source to hood, the effect of cross-drafts and thermal plumes, and the balance between capture velocity and energy use. A fast-running metamodel (trained on a targeted set of high-fidelity CFD runs and refined via active learning) enabled Bayesian, multi-objective optimisation across hood shapes, slot distributions, face-velocity targets and duct transitions, while accounting for the influence of upstream obstructions and operator positioning. We defined decision metrics such as capture efficiency at the source, containment margin at the hood, and system pressure loss to align model outputs with engineering choices. The optimisation loop produced uncertainty-aware Pareto trade-offs and flagged high-value design variants for CFD confirmation; mesh and solver settings for these confirmations were selected to capture near-field jets and entrainment reliably, with sensitivity checks increasing confidence in relative option performance. Results were communicated through intuitive contour plots, streamlines, surrogate-predicted response surfaces and short narrative summaries suitable for internal review and risk assessment. Recommendations prioritised changes with high benefit-to-effort (such as local geometry tweaks and balanced extraction across hoods) supported by surrogate-screened pressure-loss estimates and CFD-validated fan selection. The outcome is an AI-optimised, more resilient LEV design that protects operators and product quality while controlling energy consumption.