09:41
ESGF CLIMATE AI
●●
Active model: shermansiu/dm_graphcast · DownscalingHead v2
R²=0.99
⚠
Heat Stress Alert — Great Plains
Projected TAS anomaly +3.2°C above 1951–1980 baseline. Urban heat risk elevated. Agriculture impact: HIGH.
◎
Drought Risk — Southwest Region
30-day precipitation deficit. Soil moisture below 40th percentile. Water planning review recommended.
Mean TAS
14.2°C
▲ +1.8°C anomaly
Model R²
0.991
Holdout validated
RMSE
0.009
▼ scaled
Within 1°C
97.4%
Holdout set
Train 40%
R²=0.99
Val 15%
R²=0.99
Test 15%
R²=0.99
Holdout 30%
R²=0.97
◈
Run Downscaling Projection
×4 spatial enhancement · NA-CORDEX grid
→
⊞
Export Regional Report
PDF · CSV · GeoJSON
→
△
Impact Assessment Wizard
Urban planning · Agriculture
→
Pacific North America (NA-CORDEX) Atlantic ×4 downscaled · GraphCast + CanRCM4 + CESM2
TAS ANOMALY
+0–1°C
+1–2°C
+2–3°C
+3–4°C
Tap any grid cell for regional details
Forecast Horizon 2065
Downscale Resolution ×4
Confidence Interval 95%
YearΔT (°C)ScenarioRisk
2030+1.2SSP1-2.6LOW
2050+1.8SSP2-4.5MED
2075+3.1SSP3-7.0HIGH
2100+4.2SSP5-8.5CRIT
Urban Planning
Heat island effect · Infrastructure stress
MODERATE
Heat Stress
72%
Flood Risk
45%
Energy Demand
68%
Agriculture
Crop yield · Growing season · Water stress
HIGH RISK
Crop Yield
-38%
Water Stress
HIGH
Frost Days
-15d
Water Resources
Streamflow · Snowpack · Groundwater
ELEVATED
Snowpack
-52%
Streamflow
-28%
Groundwater
STRESS
2025–2030
Implement urban heat mitigation. Update zoning codes for +1.5°C baseline. Install early-warning flood sensors.
IMMEDIATE
2030–2050
Transition irrigation infrastructure. Climate-resilient crop varieties. Revise building energy codes for 2°C scenario.
MEDIUM-TERM
2050–2100
Major infrastructure redesign. Managed retreat from high-risk coastal/flood zones. Carbon removal at scale.
LONG-TERM
DownscalingHead v2
Input (16 features)
→ BatchNorm1d
→ Linear(16→256) + LayerNorm + GELU + Drop(0.35)
→ ResidualBlock(256) × 2
→ SpatialAttention(256)
→ Linear(256→128) + LayerNorm + GELU
→ ResidualBlock(128)
→ Linear(128→1) [TAS output]
GraphCast warm-init SVD projection Mixup augment AMP training
Best Val RMSE
0.009
scaled space
CV Mean RMSE
0.025
σ=0.0006
Pearson r
0.997
Holdout
Ensemble RMSE
0.007
GBR+RF+ETR
cell_mean_tas
0.94
tas_anomaly
0.72
sin_lat
0.55
cos_lat
0.48
sin_t
0.38
lat_lon
0.28
◉
NA-CORDEX · CanRCM4
1951–1970 daily TAS · 5 files · 1.3 GB
FINE
◈
CMIP6 · CESM2
1850–2014 monthly TAS · 243 MB
COARSE
⬡
GraphCast Weights
shermansiu/dm_graphcast · ERA5 0.25°
NPZ
✓ IndexError – 2-D lat/lon meshgrid (NA-CORDEX rotated-pole)
✓ ReduceLROnPlateau verbose kwarg (PyTorch ≥ 2.2)
✓ mean_squared_error squared=False (sklearn ≥ 1.4)
✓ torch.amp API for PyTorch ≥ 2.0 compatibility
✓ GraphCast .npz SVD warm-init encoder
✓ All 5 NA-CORDEX files now loading successfully