Long-term chloride-ingress forecasts are needed for durability assessment, but repeated high-fidelity transport simulations are costly. This study tests whether classical dynamic mode decomposition (DMD) remains reliable when boundary configuration and spatial data representation change. A common rank-eight model uses concentration snapshots from 20 year to 40 year and predicts through 100 year for one-surface, two-surface, and internal-source exposure. Across both representations, DMD reconstructs the training fields and extrapolates all three concentration fields for 60 year without further Finite Element time stepping. Maximum prediction errors range from 19.51% to 20.80% for surface ingress and from 3.00% to 6.61% for internal-source ingress. Attainable accuracy therefore depends on boundary configuration, although DMD remains applicable to every tested case. Switching from unstructured to structured data changes the surface-case maxima by -0.62 and +1.29 percentage points, so regular sampling provides no consistent benefit. For Case III, interpolating the structured snapshots onto unstructured nodes before retraining gives 2.97%, nearly identical to 3.00% from the native structured data. The learned DMD evolution is therefore largely insensitive to the tested spatial representation, allowing available structured or unstructured snapshots to follow the simulation or measurement workflow.