China AI Weather Forecasting vs Physics NWP Simulator

Target Scenario: Typhoon Doksuri Landfall

Grid Resolution: 25 km

Lat/Lon Bounds: 10°N-45°N, 100°E-140°E

AI Forecast Trajectory (FengWu-V2)

Physics NWP Path (ECMWF-IFS-HRES)

Observed Extreme Wind Vector / Convective Depth

Extreme Weather Preset

Model Configurations

AI Surrogate Model

Physics NWP Solver

Initial Conditions

Central Sea-Level Pressure 940 mb

Max Sustained Wind Speed 195 km/h

Sea Surface Temperature (SST) 29.5 °C

10-Day Forecast Timeline (6h Step) T + 0h

Real-Time Telemetry & Compute

24h Track Drift Error
38.2 km
AI vs Physics Spatial Lead

72h Track Drift Error
112.5 km
Ensemble Uncertainty Lead

Compute Latency 0.84 s vs 4.20 hrs

Supercomputer Energy Footprint 0.12 kWh vs 1450 kWh

Forecast Evidence Summary COMPUTED

AI Inference Latency 0.84 sec

NWP Compute Time 4.20 hours

AI GPU Energy 0.12 kWh

NWP Cluster Energy 1450 kWh

24h Track Error 38.2 km

72h Track Error 112.5 km