Adaptive-Horizon Graph Neural Operator Cuts Morphodynamic Bed-Update Times by 150-360x
AH-GNO replaces the sediment-bed update step in morphodynamic codes with a learned graph operator that adapts forecast horizon and history length on the fly. On three standard test cases it reproduces key morphological patterns at 150-360x lower computational cost than the reference GAIA solver while preserving TELEMAC-2D hydrodynamics. The approach enables rapid scenario analysis for river management and climate-impact studies.
The framework treats bed evolution as a graph-operator regression task that selects adaptive history length K and update interval B according to local flow and bed conditions. Trained on TELEMAC-2D/GAIA solutions, it reproduces bend-scale erosion, deposition, and bar migration on the official Bump and Yen-bend tests plus the Waiho River field case. Post-training inference runs in 0.08-3.34 s versus 12-1204 s for the reference solver, corresponding to 150-360x gains on 4-core CPU and 105-186x on 20-core CPU deployments. The method keeps full hydrodynamic coupling intact while replacing only the sediment-bed module. Because the operator is mesh-native and horizon-adaptive, it scales to long-duration, multi-scenario runs without the stability constraints of explicit bed-updating schemes. Remaining hydrodynamic solve time now dominates total runtime, shifting the bottleneck from morphodynamics to flow computation. Future work could couple the same operator to 3-D or two-phase models once sufficient training data exist.
Cao et al.: within 18 months at least two operational flood-risk agencies will report production use of AH-GNO-style bed-update surrogates reducing full morphodynamic ensemble runtime by >100x.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2610.22432)
- [2]Supporting Source(https://www.telemac-system.com)