Global River Forecasting with a Topology-Informed AI Foundation Model

TL;DR

GraphRiverCast (GRC), a topology-informed AI foundation model designed to simulate multivariate river hydrodynamics in global river systems, enables rapid and cross-scale adaptive simulation, establishing a collaborative paradigm bridging global hydrodynamic knowledge with local hydrological reality.

Abstract

River systems operate as inherently interconnected continuous networks, meaning river hydrodynamic simulation ought to be a systemic process. However, widespread hydrology data scarcity often restricts data-driven forecasting to isolated predictions. To achieve systemic simulation and reduce reliance on river observations, we present GraphRiverCast (GRC), a topology-informed AI foundation model designed to simulate multivariate river hydrodynamics in global river systems. GRC is capable of operating in a "ColdStart" mode, generating predictions without relying on historical river states for initialization. In 7-day global pseudo-hindcasts, GRC-ColdStart functions as a robust standalone simulator, achieving a Nash-Sutcliffe Efficiency (NSE) of approximately 0.82 without exhibiting the significant error accumulation typical of autoregressive paradigms. Ablation studies reveal that topological encoding serves as indispensable structural information in the absence of historical states, explicitly guiding hydraulic connectivity and network-scale mass redistribution to reconstruct flow dynamics. Furthermore, when adapted locally via a pre-training and fine-tuning strategy, GRC consistently outperforms physics-based and locally-trained AI baselines. Crucially, this superiority extends from gauged reaches to full river networks, underscoring the necessity of topology encoding and physics-based pre-training. Built on a physics-aligned neural operator architecture, GRC enables rapid and cross-scale adaptive simulation, establishing a collaborative paradigm bridging global hydrodynamic knowledge with local hydrological reality.

Figures (8)

• Figure 1: Schematic illustration of the GraphRiverCast framework.a, Global river system pre-training: GRC is pre-trained on physically based simulations of global river hydrodynamics, learning large-scale multivariate patterns that form the Global River Foundation Model. b, Multi-source inputs: The framework integrates static geomorphic features, hydrometeorological forcing, dynamic states, and river network topology, providing a complete description of river systems. c, Model architecture: A topology-informed neural operator framework fuses feature, temporal, and topological encoders to represent nonlinear, non-Euclidean river hydrodynamics with efficient spatiotemporal learning. d, Regional fine-tuning: Sparse local hydrological observations refine the pre-trained model into the regional river refined model. This pre-training and fine-tuning strategy enhances basin-specific predictive skill and enables accurate full-network forecasts across both gauged and ungauged reaches.
• Figure 2: Global river hydrodynamic pseudo-hindcasts with GraphRiverCast (GRC). Global spatial distributions of predictive skill relative to CaMa-Flood outputs for river discharge, depth, and storage over the future 7-day horizon, evaluated by the NSE and FHV. Results are presented for GRC-ColdStart (top three rows) and GRC-HotStart (bottom three rows). Boxplots summarize the distribution of performance metrics across major climate zones (Tropical, Temperate, Polar, Cold, and Arid). Inset scatter plots illustrate the relationship between prediction accuracy and river scale based on cumulative upstream area.
• Figure 3: Module-wise ablation analysis of GraphRiverCast (GRC).a, b, Cumulative probability distributions of NSE and FHV for the full GRC model and its ablated variants in (a) GRC-ColdStart and (b) GRC-HotStart modes. c, d, Venn diagrams illustrating the specific contributions of each encoding component (static features, temporal information, and topological structure) to predictive skill relative to the MLP baseline. Values represent the marginal NSE and FHV gains for each component and their intersections in (c) ColdStart and (d) HotStart scenarios. e, f, Temporal evolution of performance metrics. e, Evolution of predictive skill in the ColdStart mode, spanning the warm-up period starting from day 1. f, Decay of predictive skill over the forecast horizon in the HotStart mode. Shaded areas indicate the 95% confidence intervals.
• Figure 4: Evaluation of fine-tuning performance and physical consistency against in-situ discharge observations in the Amazon River Basin.a, Spatial distribution of supervised (fine-tuning) and unsupervised (withheld) GRDC stations. b, Predictive skill (NSE) relative to observed discharge for CaMa-Flood, GRC-pre-trained, GRC-fine-tuned, the non-topological variants (GRC-pre-trained_nonTopo, GRC-fine-tuned_nonTopo), and GRC trained from scratch (randomly initialized). These tunable models were trained exclusively on supervised gauges to assess spatial generalization to ungauged reaches. c, Standardized cross-correlation matrices illustrating external alignment with observed discharge ($Q_\mathrm{obs}$) and internal physical coupling among simulated states, specifically discharge ($Q_\mathrm{sim}$), water depth ($H_\mathrm{sim}$), and storage ($S_\mathrm{sim}$).
• Figure 5: Fine-tuning evaluation in the Upper Danube Basin demonstrating cross-scale transferability of the GRC.a, Map of the study domain, showing the river network and the locations of 325 reference gauges (red dots) from the LamaH-CE dataset matched to the 6-arc-minute CaMa-Flood network. b, c, Performance comparison (Median NSE) on unsupervised gauges (b) and all 325 gauges (c) under varying supervision ratios. The plots compare the GRC Foundation Model (purple line), the Scratch Model (red line), and the physical baseline (CaMa-Flood, black dashed line). Annotations (N) in b indicate the number of unsupervised gauges available for evaluation at each ratio. Shaded areas represent the 95% confidence interval.