Sub-seasonal Atmospheric River & Coastal Flood Prediction using GraphCast ML Integration
El Niño weather shifts intensify atmospheric river corridors that deposit massive precipitation along coastal mountain ranges, causing flash flooding and widespread infrastructure damage. Physics-based global numerical weather prediction models require hours on supercomputers. Deep learning weather models like GraphCast provide sub-seasonal forecasts at fraction-of-a-second inference speeds. In this challenge, you will build an inference pipeline integrating deep learning weather predictions (GraphCast or ECMWF Open Data weather weights) with local hydrological runoff models. Your solution will process 3D atmospheric temperature/humidity pressure grids, predict integrated vapor transport (IVT), and forecast river gauge flood height triggers 10 days in advance.
What you are building
The core problem, expected build, and operating context for this challenge.
Integrate GraphCast neural weather forecasting outputs with hydrological models to predict sub-seasonal atmospheric river flooding.
How work is evaluated
Evaluates river stage height forecast RMSE against synthetic stream gauge observations.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
physical_bounds_check
Predicted stage height stays within positive physically realistic domain
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
stage_height_rmse
Root Mean Squared Error of peak river stage height forecast • target: 0.35 • range: 0-3
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Incorporate deep learning weather forecast checkpoints into Python ML pipelines.
Calculate Integrated Vapor Transport (IVT) metrics from atmospheric pressure level tensors.
Couple global neural weather outputs with local catchment hydrological runoff models.
Evaluate flood forecast skill score (Brier Score) against physical river gauge observations.
Reference links and supporting material
Synthetic NetCDF weather grids containing atmospheric pressure levels, specific humidity, wind vectors, and corresponding river gauge height records.
Challenge setup
Evaluates river stage height forecast RMSE against synthetic stream gauge observations.
Challenge input
JSON with catchment_id, mean_ivt_kg_m_s, 72h_predicted_precip_mm
GraphCast
DeepMind neural weather prediction architecture.
Evaluated output
JSON with peak_stage_height_m, flood_stage_exceeded (bool), lead_time_hours
- Predicted stage height stays within positive physically realistic domain
- Root Mean Squared Error of peak river stage height forecast • target: 0.35 • range: 0-3
- Benchmark: NeuralHydrologyBench
- Stage Height Rmse target: 0.35
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Amazon Bedrock · Required
- Google Jules · Optional
- Mistral Large 2 · Optional
- Google Jules · Optional
Evaluation contract
- physical_bounds_check · Weight 1
- stage_height_rmse · Weight 1
Metrics
- stage_height_rmse · Target 0.35
Benchmarks
- NeuralHydrologyBench
- energy
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON with catchment_id, mean_ivt_kg_m_s, 72h_predicted_precip_mm
Output: JSON with peak_stage_height_m, flood_stage_exceeded (bool), lead_time_hours
Execution harness: Python
Evaluation tools
- GraphCast · Neural Weather Model
Recipe freeze marker
No recipe freeze time is recorded.