research/weather-meta-learning/resources/Model 5.218 test regions · 15 training regions
resources/
Every result from the weather study in one place: where the model learned, where it was tested, and how it did, grouped by Köppen climate.
map/
Where it learned, where it was tested.
Hover or tap a square. Filled squares are test regions coloured by error; dashed outlines are the regions used for meta-training. Filter by climate group to compare like with like.
test region, colour = error (green is lower)training regionlow errorhigh
Each square is the real 5° × 5° area the model sees. Background dots are 2° cells coloured by their main Köppen group. Error is normalised MSE over 688 forecasts in Jan–Mar 2025, where 1.0 equals climatology.trained-on/
15 training regions across 5 climate groups.
A · Tropical
IndiaAw 77% · BSh 19% · Cwa 4%
ThailandAm 25% · Aw 5%
Thailand/CambodiaAw 81% · Am 15% · Af 2%
Amazon BasinAm 63% · Af 29% · Aw 8%
B · Arid
Libya/EgyptBWh 100%
SpainBSk 57% · Csa 14% · BWk 1%
Australia OutbackBWh 53% · BSh 47%
Andes (Peru/Bolivia)BSk 24% · Cwb 22% · ET 21%
C · Temperate
E. JapanCfa 15% · Dfb 9% · Dfa 7%
TexasCfa 69% · BSh 19% · BSk 12%
D · Continental
Russia (Moscow)Dfb 100%
N. Italy/AlpsDfb 29% · Cfb 20% · Dfc 19%
NE ChinaDwb 66% · Dwa 34%
AlbertaDfb 69% · Dfc 31%
E · Polar
GreenlandEF 73% · ET 9%
Köppen mixes are measured on each region's 0.25° grid from the project's climate raster. Percentages are of the whole box, so coastal regions add up to less than 100%.
results/a-tropical/
Tropical hot and humid all year · 6 regions
India
Aw
81%less error than climatology
1.58 Ktemp RMSE
0.192norm. MSE
inputtruepredicted
area also used in meta-training
Thailand
Am
68%less error than climatology
0.9 Ktemp RMSE
0.321norm. MSE
inputtruepredicted
area also used in meta-training
Indonesia
Af
65%less error than climatology
0.99 Ktemp RMSE
0.353norm. MSE
inputtruepredicted
Debundscha, Cameroon
Am
65%less error than climatology
0.81 Ktemp RMSE
0.353norm. MSE
inputtruepredicted
Queensland, Australia
Aw
61%less error than climatology
0.83 Ktemp RMSE
0.388norm. MSE
inputtruepredicted
Amazon, Brazil
Af
43%less error than climatology
1.32 Ktemp RMSE
0.572norm. MSE
inputtruepredicted
results/b-arid/
Arid deserts and steppes · 4 regions
Sudan
BWh
86%less error than climatology
1.85 Ktemp RMSE
0.138norm. MSE
inputtruepredicted
Sahara, Mali
BWh
81%less error than climatology
2.4 Ktemp RMSE
0.191norm. MSE
inputtruepredicted
Afar, Ethiopia
BWh
77%less error than climatology
1.69 Ktemp RMSE
0.230norm. MSE
inputtruepredicted
Argentina
BSh
51%less error than climatology
2.77 Ktemp RMSE
0.489norm. MSE
inputtruepredicted
results/d-continental/
Continental cold winters, warm summers · 5 regions
Oymyakon, Russia
Dwc
79%less error than climatology
4.38 Ktemp RMSE
0.207norm. MSE
inputtruepredicted
Verkhoyansk, Russia
Dwc
75%less error than climatology
4.38 Ktemp RMSE
0.251norm. MSE
inputtruepredicted
Moscow
Dfb
58%less error than climatology
3.33 Ktemp RMSE
0.416norm. MSE
inputtruepredicted
area also used in meta-training
Lytton, Canada
Dfc
57%less error than climatology
2.82 Ktemp RMSE
0.434norm. MSE
inputtruepredicted
New York
Dfb
46%less error than climatology
3.65 Ktemp RMSE
0.544norm. MSE
inputtruepredicted
results/e-polar/
Polar tundra and ice · 3 regions
Afghanistan
ET
69%less error than climatology
2.97 Ktemp RMSE
0.312norm. MSE
inputtruepredicted
North Siberia
ET
60%less error than climatology
4.24 Ktemp RMSE
0.404norm. MSE
inputtruepredicted
Svalbard, Norway
ET
42%less error than climatology
2.6 Ktemp RMSE
0.579norm. MSE
inputtruepredicted
notes/
Scores are over all 688 forecast windows from January to March 2025 per region. The charts show only the first window, which is what the original plots show.
The redrawn temperature lines were read back from the validation plots, accurate to about ±0.2 K. The original images are linked under each chart.
“Less error than climatology” is 1 − normalised MSE: the share of the baseline's squared error that the model removes.