Every grid point inside the dashed frame is scored, sea included.
Results against ERA5
Show
Version
Models
Reference
One day ahead, AIFS's error against ERA5 was 35% below IFS's for temperature and 26% below for wind, over 171 days. At the 30 German stations the same comparison gave 10% below and 1% below, over 174 days. The two views draw closer by day 7. AIFS was trained mostly on ERA5, which probably favours it here, so take the size of any short-range lead from the station results.
One day ahead, AIFS's error against ERA5 was 34% below IFS's for temperature and 27% below for wind, over 241 days. At the 30 German stations the same comparison gave 18% below and 0% above, over 248 days. The two views draw closer by day 7. AIFS was trained mostly on ERA5, which probably favours it here, so take the size of any short-range lead from the station results.
500 hPa geopotential height: skill against climatology
171 days, 23 May 2026 to 9 November 2026, with no winter or spring. About 41 independent days. Provisional.
500 hPa geopotential height: skill against persistence
171 days, 23 May 2026 to 9 November 2026, with no winter or spring. About 41 independent days. Provisional.
500 hPa geopotential height, AIFS 2: how AIFS's error compared, over 171 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−2%not judged, no verdict
▼−8%, AIFS lower error, tested
▼−14%, AIFS lower error, tested
−7%, not separated
0%, not separated
GFS
−34%not judged, no verdict
▼−32%, AIFS lower error, tested
−29%, not separated
▼−19%, AIFS lower error, tested
−11%, not separated
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Over this sample only.
Day 1 is not judged: at day 1 the models start from different analyses, and ERA5 is an ECMWF product.
2 m temperature, gridded: skill against climatology
171 days, 23 May 2026 to 9 November 2026, with no winter or spring. About 6 independent days. Provisional.
2 m temperature, gridded: skill against persistence
171 days, 23 May 2026 to 9 November 2026, with no winter or spring. About 6 independent days. Provisional.
2 m temperature, gridded, AIFS 2: how AIFS's error compared, over 171 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−35%, no verdict
−29%, no verdict
−21%, no verdict
−15%, no verdict
−7%, no verdict
GFS
−46%, no verdict
−41%, no verdict
−33%, no verdict
−25%, no verdict
−14%, no verdict
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Sign only, not tested. No verdict is given: each model's temperature refers to its own terrain height, IFS starts from its own analysis, and AIFS was trained on ERA5.
10 m wind speed, gridded: skill against climatology
171 days, 23 May 2026 to 9 November 2026, with no winter or spring. About 65 independent days. Provisional.
10 m wind speed, gridded: skill against persistence
171 days, 23 May 2026 to 9 November 2026, with no winter or spring. About 65 independent days. Provisional.
10 m wind speed, gridded, AIFS 2: how AIFS's error compared, over 171 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−26%, no verdict
−18%, no verdict
−15%, no verdict
−8%, no verdict
−2%, no verdict
GFS
−44%, no verdict
−35%, no verdict
−27%, no verdict
−17%, no verdict
−11%, no verdict
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Sign only, not tested. No verdict is given: IFS starts from its own analysis, and AIFS was trained on ERA5.
500 hPa geopotential height: skill against climatology
166 days, 23 May 2026 to 4 November 2026, with no winter or spring. About 40 independent days. Provisional.
500 hPa geopotential height: skill against persistence
166 days, 23 May 2026 to 4 November 2026, with no winter or spring. About 40 independent days. Provisional.
500 hPa geopotential height, AIFS 2: how AIFS's error compared, over 166 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−3%not judged, no verdict
▼−8%, AIFS lower error, tested
▼−14%, AIFS lower error, tested
−7%, not separated
0%, not separated
GFS
−34%not judged, no verdict
▼−32%, AIFS lower error, tested
▼−30%, AIFS lower error, tested
▼−19%, AIFS lower error, tested
−10%, not separated
AIGFS
−39%not judged, no verdict
▼−29%, AIFS lower error, tested
▼−23%, AIFS lower error, tested
−11%, not separated
−6%, not separated
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Over this sample only.
Day 1 is not judged: at day 1 the models start from different analyses, and ERA5 is an ECMWF product.
2 m temperature, gridded: skill against climatology
166 days, 23 May 2026 to 4 November 2026, with no winter or spring. About 7 independent days. Provisional.
2 m temperature, gridded: skill against persistence
166 days, 23 May 2026 to 4 November 2026, with no winter or spring. About 7 independent days. Provisional.
2 m temperature, gridded, AIFS 2: how AIFS's error compared, over 166 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−35%, no verdict
−29%, no verdict
−21%, no verdict
−15%, no verdict
−7%, no verdict
GFS
−46%, no verdict
−41%, no verdict
−33%, no verdict
−25%, no verdict
−14%, no verdict
AIGFS
−38%, no verdict
−31%, no verdict
−23%, no verdict
−15%, no verdict
−10%, no verdict
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Sign only, not tested. No verdict is given: each model's temperature refers to its own terrain height, IFS starts from its own analysis, and AIFS was trained on ERA5.
10 m wind speed, gridded: skill against climatology
166 days, 23 May 2026 to 4 November 2026, with no winter or spring. About 49 independent days. Provisional.
10 m wind speed, gridded: skill against persistence
166 days, 23 May 2026 to 4 November 2026, with no winter or spring. About 49 independent days. Provisional.
10 m wind speed, gridded, AIFS 2: how AIFS's error compared, over 166 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−26%, no verdict
−18%, no verdict
−15%, no verdict
−8%, no verdict
−2%, no verdict
GFS
−43%, no verdict
−34%, no verdict
−26%, no verdict
−17%, no verdict
−11%, no verdict
AIGFS
−33%, no verdict
−24%, no verdict
−18%, no verdict
−11%, no verdict
−7%, no verdict
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Sign only, not tested. No verdict is given: IFS starts from its own analysis, and AIFS was trained on ERA5.
500 hPa geopotential height: skill against climatology
108 days, 20 July 2026 to 4 November 2026, with no winter or spring. About 18 independent days. Provisional.
500 hPa geopotential height: skill against persistence
108 days, 20 July 2026 to 4 November 2026, with no winter or spring. About 18 independent days. Provisional.
500 hPa geopotential height, AIFS 2: how AIFS's error compared, over 108 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
+2%not judged, no verdict
−5%, not separated
−16%, not separated
−9%, not separated
−2%, not separated
GFS
−34%not judged, no verdict
−31%, not separated
−32%, not separated
−22%, not separated
−12%, not separated
AIGFS
−37%not judged, no verdict
−29%, not separated
−26%, not separated
−15%, not separated
−8%, not separated
GraphCast
−21%not judged, no verdict
−11%, not separated
−11%, not separated
−8%, not separated
−3%, not separated
FourCastNet v2
−35%not judged, no verdict
−33%, not separated
−27%, not separated
−22%, not separated
−11%, not separated
Pangu
−22%not judged, no verdict
−21%, not separated
−20%, not separated
−19%, not separated
−9%, not separated
Aurora
−7%not judged, no verdict
−12%, not separated
−13%, not separated
−3%, not separated
+17%, not separated
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Over this sample only.
Day 1 is not judged: at day 1 the models start from different analyses, and ERA5 is an ECMWF product.
2 m temperature, gridded: skill against climatology
108 days, 20 July 2026 to 4 November 2026, with no winter or spring. About 6 independent days. Provisional.
2 m temperature, gridded: skill against persistence
108 days, 20 July 2026 to 4 November 2026, with no winter or spring. About 6 independent days. Provisional.
2 m temperature, gridded, AIFS 2: how AIFS's error compared, over 108 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−34%, no verdict
−28%, no verdict
−21%, no verdict
−16%, no verdict
−9%, no verdict
GFS
−45%, no verdict
−40%, no verdict
−33%, no verdict
−26%, no verdict
−16%, no verdict
AIGFS
−35%, no verdict
−29%, no verdict
−21%, no verdict
−16%, no verdict
−10%, no verdict
GraphCast
−29%, no verdict
−25%, no verdict
−18%, no verdict
−13%, no verdict
−7%, no verdict
FourCastNet v2
−41%, no verdict
−39%, no verdict
−29%, no verdict
−20%, no verdict
−10%, no verdict
Pangu
−18%, no verdict
−16%, no verdict
−16%, no verdict
−15%, no verdict
−9%, no verdict
Aurora
−25%, no verdict
−21%, no verdict
−14%, no verdict
−5%, no verdict
+10%, no verdict
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Sign only, not tested. No verdict is given: each model's temperature refers to its own terrain height, IFS starts from its own analysis, and AIFS was trained on ERA5.
10 m wind speed, gridded: skill against climatology
108 days, 20 July 2026 to 4 November 2026, with no winter or spring. About 29 independent days. Provisional.
10 m wind speed, gridded: skill against persistence
108 days, 20 July 2026 to 4 November 2026, with no winter or spring. About 29 independent days. Provisional.
10 m wind speed, gridded, AIFS 2: how AIFS's error compared, over 108 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−26%, no verdict
−17%, no verdict
−16%, no verdict
−8%, no verdict
−3%, no verdict
GFS
−44%, no verdict
−35%, no verdict
−27%, no verdict
−17%, no verdict
−12%, no verdict
AIGFS
−31%, no verdict
−22%, no verdict
−17%, no verdict
−9%, no verdict
−6%, no verdict
GraphCast
−10%, no verdict
−8%, no verdict
−7%, no verdict
−3%, no verdict
−2%, no verdict
FourCastNet v2
−32%, no verdict
−26%, no verdict
−19%, no verdict
−10%, no verdict
−6%, no verdict
Pangu
−16%, no verdict
−13%, no verdict
−10%, no verdict
−7%, no verdict
−3%, no verdict
Aurora
−4%, no verdict
−3%, no verdict
−13%, no verdict
−16%, no verdict
−14%, no verdict
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Sign only, not tested. No verdict is given: IFS starts from its own analysis, and AIFS was trained on ERA5.
500 hPa geopotential height: skill against climatology
241 days, 7 September 2025 to 11 May 2026, with no summer. About 40 independent days. Provisional.
500 hPa geopotential height: skill against persistence
241 days, 7 September 2025 to 11 May 2026, with no summer. About 40 independent days. Provisional.
500 hPa geopotential height, AIFS 1.1: how AIFS's error compared, over 241 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
+5%not judged, no verdict
−6%, not separated
−11%, not separated
−6%, not separated
−6%, not separated
GFS
−33%not judged, no verdict
▼−32%, AIFS lower error, tested
▼−28%, AIFS lower error, tested
▼−17%, AIFS lower error, tested
▼−12%, AIFS lower error, tested
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Over this sample only.
Day 1 is not judged: at day 1 the models start from different analyses, and ERA5 is an ECMWF product.
2 m temperature, gridded: skill against climatology
241 days, 7 September 2025 to 11 May 2026, with no summer. About 10 independent days. Provisional.
2 m temperature, gridded: skill against persistence
241 days, 7 September 2025 to 11 May 2026, with no summer. About 10 independent days. Provisional.
2 m temperature, gridded, AIFS 1.1: how AIFS's error compared, over 241 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−34%, no verdict
−28%, no verdict
−20%, no verdict
−12%, no verdict
−7%, no verdict
GFS
−47%, no verdict
−40%, no verdict
−30%, no verdict
−20%, no verdict
−11%, no verdict
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Sign only, not tested. No verdict is given: each model's temperature refers to its own terrain height, IFS starts from its own analysis, and AIFS was trained on ERA5.
10 m wind speed, gridded: skill against climatology
241 days, 7 September 2025 to 11 May 2026, with no summer. About 72 independent days. Provisional.
10 m wind speed, gridded: skill against persistence
241 days, 7 September 2025 to 11 May 2026, with no summer. About 72 independent days. Provisional.
10 m wind speed, gridded, AIFS 1.1: how AIFS's error compared, over 241 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−27%, no verdict
−17%, no verdict
−13%, no verdict
−8%, no verdict
−6%, no verdict
GFS
−45%, no verdict
−34%, no verdict
−25%, no verdict
−17%, no verdict
−13%, no verdict
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Sign only, not tested. No verdict is given: IFS starts from its own analysis, and AIFS was trained on ERA5.
AIGFS has no forecasts in the AIFS 1.1 period, so this comparison exists for AIFS 2 only.
500 hPa geopotential height: skill against climatology
241 days, 7 September 2025 to 11 May 2026, with no summer. About 40 independent days. Provisional.
500 hPa geopotential height: skill against persistence
241 days, 7 September 2025 to 11 May 2026, with no summer. About 40 independent days. Provisional.
500 hPa geopotential height, AIFS 1.1: how AIFS's error compared, over 241 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
+5%not judged, no verdict
−6%, not separated
−11%, not separated
−6%, not separated
−6%, not separated
GFS
−33%not judged, no verdict
▼−32%, AIFS lower error, tested
▼−28%, AIFS lower error, tested
▼−17%, AIFS lower error, tested
▼−12%, AIFS lower error, tested
AIGFS
–
–
–
–
–
GraphCast
−4%not judged, no verdict
+1%, not separated
−4%, not separated
0%, not separated
−5%, not separated
FourCastNet v2
−31%not judged, no verdict
▼−30%, AIFS lower error, tested
▼−24%, AIFS lower error, tested
−13%, not separated
−6%, not separated
Pangu
−7%not judged, no verdict
▼−10%, AIFS lower error, tested
▼−11%, AIFS lower error, tested
−6%, not separated
−9%, not separated
Aurora
+10%not judged, no verdict
+3%, not separated
0%, not separated
+4%, not separated
▲+13%, AIFS higher error, tested
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Over this sample only.
Day 1 is not judged: at day 1 the models start from different analyses, and ERA5 is an ECMWF product.
2 m temperature, gridded: skill against climatology
241 days, 7 September 2025 to 11 May 2026, with no summer. About 10 independent days. Provisional.
2 m temperature, gridded: skill against persistence
241 days, 7 September 2025 to 11 May 2026, with no summer. About 10 independent days. Provisional.
2 m temperature, gridded, AIFS 1.1: how AIFS's error compared, over 241 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−34%, no verdict
−28%, no verdict
−20%, no verdict
−12%, no verdict
−7%, no verdict
GFS
−47%, no verdict
−40%, no verdict
−30%, no verdict
−20%, no verdict
−11%, no verdict
AIGFS
–
–
–
–
–
GraphCast
−27%, no verdict
−21%, no verdict
−13%, no verdict
−7%, no verdict
−2%, no verdict
FourCastNet v2
−40%, no verdict
−36%, no verdict
−26%, no verdict
−14%, no verdict
−5%, no verdict
Pangu
−17%, no verdict
−14%, no verdict
−13%, no verdict
−8%, no verdict
−7%, no verdict
Aurora
−18%, no verdict
−13%, no verdict
−7%, no verdict
+1%, no verdict
+14%, no verdict
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Sign only, not tested. No verdict is given: each model's temperature refers to its own terrain height, IFS starts from its own analysis, and AIFS was trained on ERA5.
10 m wind speed, gridded: skill against climatology
241 days, 7 September 2025 to 11 May 2026, with no summer. About 72 independent days. Provisional.
10 m wind speed, gridded: skill against persistence
241 days, 7 September 2025 to 11 May 2026, with no summer. About 72 independent days. Provisional.
10 m wind speed, gridded, AIFS 1.1: how AIFS's error compared, over 241 days
Compared with
Day 1
Day 3
Day 5
Day 7
Day 10
IFS
−27%, no verdict
−17%, no verdict
−13%, no verdict
−8%, no verdict
−6%, no verdict
GFS
−45%, no verdict
−34%, no verdict
−25%, no verdict
−17%, no verdict
−13%, no verdict
AIGFS
–
–
–
–
–
GraphCast
−10%, no verdict
−3%, no verdict
−3%, no verdict
−1%, no verdict
−3%, no verdict
FourCastNet v2
−33%, no verdict
−22%, no verdict
−15%, no verdict
−8%, no verdict
−3%, no verdict
Pangu
−15%, no verdict
−8%, no verdict
−6%, no verdict
−3%, no verdict
−3%, no verdict
Aurora
−2%, no verdict
0%, no verdict
−8%, no verdict
−14%, no verdict
−16%, no verdict
Each number is AIFS's RMSE against ERA5 minus the other model's, as a percentage of the other model's: minus is lower (better), plus is higher (worse). Sign only, not tested. No verdict is given: IFS starts from its own analysis, and AIFS was trained on ERA5.
▼AIFS lower error, tested (better)
▲AIFS higher error, tested (worse)
Not separated: no clear difference, too few days, or the verdict depends on how the station is read or on the height correction (hover says which)
Dashed outline: not judged, sign only
Bias and activity: 500 hpa geopotential heightWhether a model runs high or low, and how much it varies
500 hPa geopotential height: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better. For 500 hPa height only, a value below 0.6 is commonly read as a forecast of little use; here that reading is applied to the uncentred form.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9990.999
0.9900.989
0.9480.938
0.8410.804
0.5790.505
IFS
0.9990.999
0.9880.986
0.9300.915
0.8070.761
0.5630.470
GFS
0.9980.998
0.9800.975
0.9060.884
0.7550.697
0.4440.347
500 hPa geopotential height: bias, in gpmForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
+0.95
+1.39
+1.14
−0.18
−0.76
IFS
+1.69
+2.81
+1.68
+3.05
+1.67
GFS
−2.27
+0.16
+1.02
+1.32
−6.10
500 hPa geopotential height: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
1.00
1.00
0.99
1.00
1.03
IFS
1.00
0.98
0.97
0.99
0.99
GFS
1.00
0.98
0.98
0.99
1.02
Bias and activity: 2 m temperature, griddedWhether a model runs high or low, and how much it varies
2 m temperature, gridded: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9630.951
0.9410.921
0.8810.843
0.7700.697
0.5510.428
IFS
0.9150.890
0.8850.851
0.8170.762
0.7040.616
0.5210.397
GFS
0.8790.853
0.8400.803
0.7580.699
0.6330.542
0.4300.323
2 m temperature, gridded: bias, in KForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
−0.07
−0.09
−0.07
−0.03
−0.12
IFS
−0.14
−0.12
−0.09
−0.04
−0.07
GFS
−0.24
−0.21
−0.18
−0.19
−0.34
2 m temperature, gridded: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.98
0.96
0.96
0.95
0.96
IFS
0.96
0.95
0.96
0.97
0.99
GFS
1.02
1.02
1.02
1.02
1.05
Bias and activity: 10 m wind speed, griddedWhether a model runs high or low, and how much it varies
10 m wind speed, gridded: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9310.933
0.8520.856
0.6840.686
0.4270.424
0.1760.171
IFS
0.8810.877
0.7860.781
0.5820.572
0.3400.330
0.1470.139
GFS
0.8250.829
0.7030.702
0.4940.489
0.2730.270
0.0750.071
10 m wind speed, gridded: bias, in m/sForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
−0.19
−0.29
−0.35
−0.38
−0.40
IFS
+0.00
−0.04
−0.07
−0.10
−0.10
GFS
+0.32
+0.26
+0.20
+0.19
+0.22
10 m wind speed, gridded: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.99
0.97
0.96
0.96
0.97
IFS
1.02
1.01
1.00
1.00
1.00
GFS
1.12
1.10
1.09
1.09
1.10
Bias and activity: 500 hpa geopotential heightWhether a model runs high or low, and how much it varies
500 hPa geopotential height: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better. For 500 hPa height only, a value below 0.6 is commonly read as a forecast of little use; here that reading is applied to the uncentred form.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9990.999
0.9900.989
0.9490.939
0.8410.805
0.5770.502
IFS
0.9990.999
0.9880.986
0.9310.917
0.8050.759
0.5620.467
GFS
0.9980.998
0.9800.975
0.9070.884
0.7580.698
0.4450.340
AIGFS
0.9970.998
0.9810.978
0.9190.904
0.7900.743
0.4700.391
500 hPa geopotential height: bias, in gpmForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
+0.91
+1.36
+1.30
+0.14
−0.46
IFS
+1.68
+2.80
+1.73
+3.42
+2.40
GFS
−2.27
+0.26
+1.47
+2.16
−4.28
AIGFS
−3.66
−3.08
−1.97
−0.83
−4.82
500 hPa geopotential height: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
1.00
1.00
0.99
1.00
1.02
IFS
1.00
0.98
0.97
0.99
0.98
GFS
1.00
0.98
0.98
0.98
1.00
AIGFS
1.00
0.99
1.00
1.00
1.00
Bias and activity: 2 m temperature, griddedWhether a model runs high or low, and how much it varies
2 m temperature, gridded: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9630.950
0.9400.920
0.8800.841
0.7670.691
0.5480.422
IFS
0.9140.889
0.8850.850
0.8160.760
0.7020.611
0.5210.393
GFS
0.8770.851
0.8380.800
0.7570.696
0.6300.536
0.4300.318
AIGFS
0.9030.881
0.8720.844
0.7980.752
0.6720.595
0.4430.333
2 m temperature, gridded: bias, in KForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
−0.07
−0.09
−0.06
−0.02
−0.10
IFS
−0.14
−0.11
−0.08
−0.04
−0.06
GFS
−0.26
−0.23
−0.20
−0.19
−0.33
AIGFS
−0.28
−0.33
−0.34
−0.32
−0.41
2 m temperature, gridded: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.98
0.96
0.96
0.95
0.95
IFS
0.96
0.95
0.96
0.96
0.99
GFS
1.03
1.03
1.02
1.03
1.05
AIGFS
0.98
0.97
0.98
0.97
1.00
Bias and activity: 10 m wind speed, griddedWhether a model runs high or low, and how much it varies
10 m wind speed, gridded: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9300.932
0.8490.854
0.6800.682
0.4210.418
0.1690.164
IFS
0.8790.876
0.7830.778
0.5780.568
0.3340.324
0.1410.133
GFS
0.8240.828
0.7010.700
0.4890.484
0.2660.263
0.0720.068
AIGFS
0.8650.866
0.7620.760
0.5720.567
0.3350.331
0.1060.101
10 m wind speed, gridded: bias, in m/sForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
−0.19
−0.29
−0.35
−0.38
−0.40
IFS
+0.00
−0.05
−0.07
−0.10
−0.09
GFS
+0.31
+0.25
+0.20
+0.18
+0.20
AIGFS
+0.19
+0.09
+0.01
−0.03
−0.00
10 m wind speed, gridded: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.99
0.97
0.96
0.96
0.98
IFS
1.02
1.00
1.00
1.00
1.01
GFS
1.12
1.10
1.09
1.09
1.10
AIGFS
1.09
1.06
1.05
1.04
1.06
Bias and activity: 500 hpa geopotential heightWhether a model runs high or low, and how much it varies
500 hPa geopotential height: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better. For 500 hPa height only, a value below 0.6 is commonly read as a forecast of little use; here that reading is applied to the uncentred form.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9990.999
0.9900.988
0.9480.940
0.8440.808
0.5710.504
IFS
0.9990.999
0.9880.986
0.9240.910
0.7910.753
0.5340.451
GFS
0.9980.998
0.9790.975
0.9020.879
0.7420.686
0.4190.336
AIGFS
0.9970.998
0.9810.977
0.9140.900
0.7810.731
0.4440.375
GraphCast
0.9980.999
0.9870.985
0.9370.928
0.8140.772
0.5240.435
FourCastNet v2
0.9980.997
0.9770.973
0.9100.894
0.7400.691
0.4080.328
Pangu
0.9980.998
0.9830.981
0.9250.910
0.7650.712
0.4350.340
Aurora
0.9990.999
0.9870.987
0.9350.927
0.8110.778
0.5870.492
500 hPa geopotential height: bias, in gpmForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
+0.72
+1.22
+1.50
+0.51
+1.85
IFS
+1.28
+2.01
+1.06
+2.75
+2.03
GFS
−2.62
−0.68
+0.65
+0.46
−7.05
AIGFS
−3.42
−2.79
−2.04
−0.81
−3.52
GraphCast
+3.02
+2.12
+2.24
+1.13
+1.02
FourCastNet v2
−0.66
−2.47
−3.11
−5.18
−8.21
Pangu
+1.89
+0.98
−0.66
−2.75
−5.07
Aurora
−1.56
−5.55
−4.68
−2.69
−5.77
500 hPa geopotential height: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
1.00
1.00
0.98
0.99
1.04
IFS
1.00
0.98
0.96
0.97
0.99
GFS
1.00
0.98
0.98
0.98
0.98
AIGFS
1.00
0.99
0.99
0.99
1.00
GraphCast
1.00
1.00
0.99
1.00
1.00
FourCastNet v2
1.00
0.99
0.97
0.96
0.94
Pangu
1.00
1.00
0.99
1.00
1.01
Aurora
1.00
0.99
0.91
0.81
0.66
Bias and activity: 2 m temperature, griddedWhether a model runs high or low, and how much it varies
2 m temperature, gridded: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9620.951
0.9370.919
0.8750.839
0.7640.698
0.5410.427
IFS
0.9120.888
0.8800.848
0.8070.752
0.6910.607
0.5010.383
GFS
0.8800.853
0.8390.802
0.7550.695
0.6330.547
0.4340.330
AIGFS
0.9070.887
0.8730.848
0.7960.754
0.6590.587
0.4400.334
GraphCast
0.9250.904
0.8900.860
0.8140.763
0.6950.614
0.4790.365
FourCastNet v2
0.8890.854
0.8260.780
0.7440.685
0.6040.520
0.3860.263
Pangu
0.9440.927
0.9120.886
0.8300.780
0.6820.589
0.4510.319
Aurora
0.9330.916
0.9000.877
0.8260.784
0.7230.644
0.5650.430
2 m temperature, gridded: bias, in KForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
−0.11
−0.15
−0.13
−0.11
−0.16
IFS
−0.13
−0.10
−0.07
−0.00
−0.06
GFS
−0.16
−0.12
−0.08
−0.09
−0.23
AIGFS
−0.24
−0.33
−0.37
−0.34
−0.35
GraphCast
−0.04
−0.09
−0.10
−0.10
−0.21
FourCastNet v2
−0.06
−0.27
−0.40
−0.48
−0.49
Pangu
+0.09
+0.11
+0.10
+0.08
−0.07
Aurora
−0.12
−0.23
−0.30
−0.29
−0.22
2 m temperature, gridded: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.96
0.94
0.94
0.94
0.97
IFS
0.95
0.94
0.95
0.97
1.02
GFS
1.03
1.04
1.04
1.06
1.10
AIGFS
0.97
0.95
0.95
0.96
0.99
GraphCast
0.94
0.92
0.92
0.93
0.96
FourCastNet v2
0.89
0.85
0.83
0.83
0.86
Pangu
0.94
0.94
0.95
0.94
0.96
Aurora
0.94
0.91
0.84
0.75
0.68
Bias and activity: 10 m wind speed, griddedWhether a model runs high or low, and how much it varies
10 m wind speed, gridded: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9320.934
0.8510.856
0.6790.684
0.4190.420
0.1560.155
IFS
0.8830.881
0.7870.783
0.5650.555
0.3250.317
0.1310.125
GFS
0.8280.833
0.7050.704
0.4830.479
0.2620.259
0.0620.059
AIGFS
0.8790.879
0.7760.772
0.5790.572
0.3420.336
0.1140.107
GraphCast
0.9180.918
0.8260.826
0.6350.633
0.3800.373
0.1180.112
FourCastNet v2
0.8560.860
0.7290.738
0.5200.531
0.2840.282
0.0640.063
Pangu
0.9050.904
0.8060.802
0.6100.601
0.3250.315
0.0930.090
Aurora
0.9270.926
0.8370.838
0.5720.612
0.2350.289
0.0640.074
10 m wind speed, gridded: bias, in m/sForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
−0.18
−0.28
−0.36
−0.41
−0.42
IFS
+0.01
−0.04
−0.09
−0.12
−0.13
GFS
+0.34
+0.28
+0.24
+0.22
+0.24
AIGFS
+0.15
+0.02
−0.07
−0.10
−0.14
GraphCast
−0.14
−0.22
−0.30
−0.33
−0.37
FourCastNet v2
−0.28
−0.43
−0.59
−0.67
−0.79
Pangu
−0.10
−0.12
−0.15
−0.20
−0.27
Aurora
−0.06
−0.23
−0.81
−1.42
−1.90
10 m wind speed, gridded: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.99
0.97
0.95
0.95
0.96
IFS
1.02
1.00
0.99
0.99
1.00
GFS
1.12
1.10
1.08
1.08
1.11
AIGFS
1.09
1.05
1.03
1.02
1.05
GraphCast
1.00
0.97
0.95
0.94
0.95
FourCastNet v2
0.97
0.93
0.90
0.89
0.90
Pangu
0.99
0.98
0.97
0.97
0.97
Aurora
1.00
0.94
0.81
0.67
0.56
Bias and activity: 500 hpa geopotential heightWhether a model runs high or low, and how much it varies
500 hPa geopotential height: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better. For 500 hPa height only, a value below 0.6 is commonly read as a forecast of little use; here that reading is applied to the uncentred form.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9990.999
0.9890.987
0.9450.935
0.8090.779
0.5260.476
IFS
0.9990.999
0.9880.986
0.9310.919
0.7800.750
0.4510.409
GFS
0.9980.998
0.9780.974
0.8960.881
0.7220.684
0.3880.348
500 hPa geopotential height: bias, in gpmForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
+1.56
+0.22
−0.35
−0.93
−0.21
IFS
+1.64
+1.70
+2.22
+2.54
+2.18
GFS
−2.04
−0.89
+0.68
+3.16
+3.86
500 hPa geopotential height: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.99
0.99
0.99
0.98
0.97
IFS
1.00
0.99
0.99
0.98
0.98
GFS
1.00
1.00
0.99
0.98
1.00
Bias and activity: 2 m temperature, griddedWhether a model runs high or low, and how much it varies
2 m temperature, gridded: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9670.959
0.9410.926
0.8730.842
0.7360.681
0.4850.396
IFS
0.9220.905
0.8850.862
0.8030.764
0.6630.605
0.4110.342
GFS
0.8840.861
0.8390.809
0.7420.695
0.5850.514
0.3480.267
2 m temperature, gridded: bias, in KForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
+0.01
−0.03
−0.05
−0.07
−0.13
IFS
−0.19
−0.22
−0.25
−0.27
−0.35
GFS
−0.23
−0.28
−0.30
−0.34
−0.31
2 m temperature, gridded: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.98
0.97
0.97
0.98
0.99
IFS
0.97
0.98
0.99
0.99
1.00
GFS
0.99
0.98
0.99
0.98
0.99
Bias and activity: 10 m wind speed, griddedWhether a model runs high or low, and how much it varies
10 m wind speed, gridded: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9470.947
0.8530.852
0.6730.669
0.4320.423
0.1940.188
IFS
0.9070.903
0.7960.788
0.5960.583
0.3700.353
0.1400.134
GFS
0.8630.862
0.7300.722
0.5150.502
0.2970.285
0.0940.089
10 m wind speed, gridded: bias, in m/sForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
−0.19
−0.31
−0.39
−0.50
−0.51
IFS
+0.04
−0.01
−0.04
−0.10
−0.14
GFS
+0.32
+0.27
+0.23
+0.17
+0.21
10 m wind speed, gridded: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.99
0.97
0.96
0.95
0.95
IFS
1.05
1.03
1.03
1.03
1.02
GFS
1.15
1.14
1.14
1.12
1.13
Bias and activity: 500 hpa geopotential heightWhether a model runs high or low, and how much it varies
500 hPa geopotential height: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better. For 500 hPa height only, a value below 0.6 is commonly read as a forecast of little use; here that reading is applied to the uncentred form.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9990.999
0.9890.987
0.9450.935
0.8090.779
0.5260.476
IFS
0.9990.999
0.9880.986
0.9310.919
0.7800.750
0.4510.409
GFS
0.9980.998
0.9780.974
0.8960.881
0.7220.684
0.3880.348
GraphCast
0.9990.999
0.9900.988
0.9420.932
0.8130.782
0.4950.448
FourCastNet v2
0.9980.998
0.9790.976
0.9050.891
0.7450.712
0.4670.436
Pangu
0.9990.999
0.9870.985
0.9320.921
0.7960.765
0.4700.430
Aurora
0.9990.999
0.9900.988
0.9430.934
0.8020.776
0.5290.475
500 hPa geopotential height: bias, in gpmForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
+1.56
+0.22
−0.35
−0.93
−0.21
IFS
+1.64
+1.70
+2.22
+2.54
+2.18
GFS
−2.04
−0.89
+0.68
+3.16
+3.86
GraphCast
+2.20
+1.68
+0.97
−0.20
+4.93
FourCastNet v2
−1.54
−4.47
−6.47
−9.37
−7.43
Pangu
+0.38
−1.08
−2.56
−3.98
−2.49
Aurora
+0.64
−2.42
+0.21
+7.64
+8.41
500 hPa geopotential height: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.99
0.99
0.99
0.98
0.97
IFS
1.00
0.99
0.99
0.98
0.98
GFS
1.00
1.00
0.99
0.98
1.00
GraphCast
1.00
1.00
1.01
1.00
0.99
FourCastNet v2
1.00
1.00
1.01
1.00
1.01
Pangu
1.00
1.01
1.02
1.03
1.05
Aurora
1.00
1.00
0.93
0.80
0.61
Bias and activity: 2 m temperature, griddedWhether a model runs high or low, and how much it varies
2 m temperature, gridded: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9670.959
0.9410.926
0.8730.842
0.7360.681
0.4850.396
IFS
0.9220.905
0.8850.862
0.8030.764
0.6630.605
0.4110.342
GFS
0.8840.861
0.8390.809
0.7420.695
0.5850.514
0.3480.267
GraphCast
0.9360.921
0.9010.879
0.8310.794
0.6930.634
0.4460.373
FourCastNet v2
0.9040.880
0.8520.819
0.7640.719
0.6270.564
0.4130.345
Pangu
0.9510.938
0.9210.903
0.8390.805
0.6980.643
0.4290.357
Aurora
0.9490.937
0.9190.902
0.8490.815
0.7110.643
0.5110.399
2 m temperature, gridded: bias, in KForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
+0.01
−0.03
−0.05
−0.07
−0.13
IFS
−0.19
−0.22
−0.25
−0.27
−0.35
GFS
−0.23
−0.28
−0.30
−0.34
−0.31
GraphCast
−0.16
−0.18
−0.18
−0.21
−0.27
FourCastNet v2
−0.13
−0.27
−0.40
−0.49
−0.63
Pangu
−0.09
−0.10
−0.16
−0.26
−0.44
Aurora
−0.15
−0.22
−0.27
−0.17
+0.05
2 m temperature, gridded: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.98
0.97
0.97
0.98
0.99
IFS
0.97
0.98
0.99
0.99
1.00
GFS
0.99
0.98
0.99
0.98
0.99
GraphCast
0.95
0.94
0.94
0.95
0.95
FourCastNet v2
0.96
0.94
0.95
0.96
0.98
Pangu
0.98
0.99
1.00
1.02
1.05
Aurora
0.96
0.94
0.90
0.82
0.69
Bias and activity: 10 m wind speed, griddedWhether a model runs high or low, and how much it varies
10 m wind speed, gridded: anomaly correlation, uncentredHow well the forecast pattern matches ERA5, as departures from the 1990 to 2019 climatology. The first value is the uncentred form, as WeatherBench 2 computes it. The smaller value under it is the centred form, with the area-mean anomaly of each field removed first. 1 is perfect and higher is better.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.9470.947
0.8530.852
0.6730.669
0.4320.423
0.1940.188
IFS
0.9070.903
0.7960.788
0.5960.583
0.3700.353
0.1400.134
GFS
0.8630.862
0.7300.722
0.5150.502
0.2970.285
0.0940.089
GraphCast
0.9360.934
0.8460.842
0.6610.653
0.4300.419
0.1670.162
FourCastNet v2
0.8830.881
0.7590.757
0.5580.554
0.3340.325
0.1610.157
Pangu
0.9260.923
0.8270.819
0.6350.624
0.4040.392
0.1580.150
Aurora
0.9450.942
0.8520.849
0.6080.641
0.2880.352
0.0830.101
10 m wind speed, gridded: bias, in m/sForecast minus ERA5, on average. Plus: too high. Minus: too low. Zero is best.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
−0.19
−0.31
−0.39
−0.50
−0.51
IFS
+0.04
−0.01
−0.04
−0.10
−0.14
GFS
+0.32
+0.27
+0.23
+0.17
+0.21
GraphCast
−0.12
−0.21
−0.26
−0.28
−0.32
FourCastNet v2
−0.24
−0.40
−0.55
−0.60
−0.65
Pangu
−0.09
−0.12
−0.17
−0.21
−0.19
Aurora
−0.03
−0.24
−0.91
−1.76
−2.35
10 m wind speed, gridded: activityVariability of the forecast divided by that of the ERA5 field, both as departures from climatology, with each grid cell centred on its own mean at each valid hour over the sample first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
Model
Day 1
Day 3
Day 5
Day 7
Day 10
AIFS
0.99
0.97
0.96
0.95
0.95
IFS
1.05
1.03
1.03
1.03
1.02
GFS
1.15
1.14
1.14
1.12
1.13
GraphCast
1.00
0.98
0.98
0.98
0.97
FourCastNet v2
0.98
0.95
0.93
0.94
0.92
Pangu
0.99
0.98
0.98
0.98
0.98
Aurora
1.00
0.96
0.84
0.69
0.55
Forecasts scored against ERA5 from 35 to 72°N and 12.5°W to 40°E, weighted by latitude. ERA5 is a model product and many AI models were trained on it. Treat this as a second view beside the station measurements. Its numbers cannot be compared with the station scores.
Where the error sat
Lead
Each square averages a 1.5° cell over the days scored. These show patterns only, with no verdict for any square: on the second map the colour gives the sign of the difference and is not a test. They follow the variable and version selected above.