Mechanistically informed multi-scale spatiotemporal autoregressive graph learning reveals cascading impacts of extreme precipitation

Economic losses caused by extreme climate are not confined to the locations where events occur, but can propagate across regions through physical and economic linkages. Yet existing climate-impact assessment methods remain poorly suited to tracing how shocks spread across space and reshape the geography of economic loss. Here we develop a mechanistically informed multi-scale spatiotemporal autoregressive graph neural network model to quantify spatially cascading climate impacts. The model couples scale-specific, physically structured spatiotemporal autoregressive processes through an adaptive gating mechanism, allowing heterogeneous cross-scale interactions to be learned from data. Model estimation is achieved through tailored graph convolutional neural networks that are mathematically equivalent to spatiotemporal autoregressive models, enabling scalability while preserving transparent parameter interpretation. Monte Carlo simulation experiments show that the model accurately recovers true parameters and distinguishes between scale-dependent processes. Applying the framework to extreme precipitation, we find that large-scale upwind-to-downwind cascades driven by atmospheric circulations dominate aggregated economic losses. A one-standard-deviation increase in log extreme precipitation is associated with a 0.19 percentage-point decline in economic growth rate at the large scale, with 62.3% of the loss arising from spatial cascades. These findings highlight the need for transboundary risk governance that incorporates spatial cascading into climate-extremes monitoring and early-warning. Description of the uploaded file Monte Carlo simulation codedata_generator_factors.py: Data generation script for multi-scale Monte Carlo simulation experiments.sarnn_model.py:Implementation of the proposed MS-STARGNNs model architecture definition.train.py:Training pipeline script for the MS-STARGNNs model. Data and spatial weights matrices for empirical analysisglobal_panel_1deg_std.parquet:Standardized Large-scale (1°) datase;global_panel_2km_std.parquet:Standardized Small-scale (2 km) dataset.W_global_2km_knn8.pt:Small-scale spatial weights matrix based on 8-nearest neighbors (KNN8).W_Large-scale:A large-scale spatial weights matrix derived from moisture transport pathways (2005–2021)mapping_1deg_to_2km.parquet: Correspondence file mapping large-scale (1°) grid cells to small-scale (2km) grid cells.ipcc_region_mapping_coarse_8regions.parquet: Mapping file linking Large-scale grid cells to the 8 IPCC AR6 reference regions.ipcc_region_mapping_fine_8regions.parquet: Mapping file linking Small-scale grid cells to the 8 IPCC AR6 reference regions. Codemodel.py: Core architecture definitions for empirical analysis. Provides the base classes and computational layers engineered to handle real-world geospatial complexities. All subsequent training scripts import modules from this file.STARGNNs.py: Implementation of the single-scale baseline. Serves as a reference point for evaluating the efficacy of cross-scale feature fusion.MS-STARGNNs_fixed.py: Configuration script for the MS-STARGNNs model utilizing fixed autoregressive coefficients.MS-STARGNNs.py: Configuration script for the MS-STARGNNs model utilizing annually varying autoregressive coefficients.MS-STARGNNs_8 regions.py: Executes the MS-STARGNNs model with decoupled regional parameters, loading unique autoregressive weights and βvectors for each IPCC region. Implements null value handling for regions lacking observational data (e.g., Antarctica).