Physics-Guided Multi-Fidelity Surrogate and Regime-Aware Evolutionary Optimization for Climate-Adaptive Eco-Hydrological Wetland Restoration

Overview Restoring degraded wetlands under non-stationary climate forcing requires balancing habitat biodiversity, carbon sequestration, flood attenuation, and lifecycle engineering cost across coupled two-dimensional (2D) shallow-water hydrodynamics and multi-decadal vegetation succession. Evaluating a single restoration plan with a high-fidelity coupled hydrologic-hydraulic (H&H) and biogeochemical solver is computationally prohibitive for evolutionary search, while data-driven or soft-penalty physics-informed surrogates violate interfacial mass and momentum conservation across adjustable weirs, tidal sluices, and dredged channels. Under compound climate extremes the resulting flux errors distort hydroperiod duration and porewater salinity, which pushes evolutionary search toward candidates that look optimal but cross salinity-toxicity or anoxic-collapse thresholds. We present Physics-Guided Regime-Aware Multi-objective Evolutionary Optimization (PG-RAMEO), which integrates three modules and treats conservation, boundary awareness, and repair as one loop: (1) a Saint-Venant Flux-Constrained Multi-Fidelity Graph Neural Operator (SV-MFGNO) places an antisymmetric Riemann flux layer over coupled coarse-grid and fine-grid control volumes, so internal fluxes cancel exactly; (2) a Climate-Regime-Aware Reflection and Feasibility-Guided Evolutionary Harness (CR-FEH) combines ensemble epistemic spread, conservation residual, and ecological tipping-margin proximity into one query-allocation score, then repairs infeasible offspring along the constraint gradient before they reach the solver; and (3) a Non-Stationary Regime-Curriculum Reference Vector Adaptation (NS-RCA) rescales reference vectors as the forcing regime shifts. Across 240 restoration units spanning three geomorphological archetypes and three climate regimes (Historical, CMIP6 SSP2-4.5, CMIP6 SSP5-8.5), PG-RAMEO reaches a held-out test Hypervolume of 0.842 ± 0.011 against 0.798 ± 0.013 for the strongest baseline Bi-Level-SAEA (Holm-Bonferroni p = 1.91 × 10^-30), raises ecological regime feasibility from 83.5% to 94.6%, and delivers an 18.4× computational speedup. Contents This record contains 16 files. The table lists every data table in the archive with its row count. FileRowsSize (KB)data/derived/ablation_table.csv81data/derived/dataset_summary_table.csv40data/derived/failure_mode_breakdown.csv81data/derived/main_table.csv162data/derived/regime_archetype_table.csv484data/derived/sensitivity_table.csv928data/derived/surrogate_consistency_table.csv70data/processed/ablation_basin_evaluations.csv1,440248data/processed/method_basin_regime_evaluations.csv5,760890data/processed/pareto_front_solutions.csv20011data/processed/regime_shift_trajectories.csv1117data/processed/seed_level_test_runs.csv72048data/processed/sensitivity_sweep_grid.csv928data/processed/surrogate_operator_grid.csv635data/processed/wetland_basins_metadata.csv24028 Reuse Released under cc-by-4.0. If you use these data, please cite this record together with the associated article.