Data-driven weather forecasting requires distinguishing local states and their temporal evolution. However, learned patterns remain implicit in neural forecasters, making their influence difficult to interpret. Failing to distinguish local evolution can lead to similar forecasts for locations that develop differently, increasing prediction error. This motivates explicit references for distinguishing and interpreting local patterns. Prototypes provide explicit pattern references, but sample- or task-level conditioning leaves open how to adapt to distinct locations and variable groups. To deal with this problem, we design ProtoCast, which uses group-specific prototype banks to adapt local features while accounting for cross-variable dependencies. Specifically, (1) we combine each group's historical features and temporal change with forecast context from cross-variable attention for evolution-aware prototype matching. (2) Sparse mixtures of prototype-specific low-rank operators tailor corrections to each group and location. A gated attention block then exchanges the corrected features across groups before prediction. Each group's correction decomposes into explicit prototype contributions before this fusion. (3) We jointly train the forecasting network and operators from scratch and refit each group's prototypes before every training epoch. Slot tracking preserves prototype-operator associations, making each prototype's contribution to local corrections more interpretable. Experiments on GlobalForecast and TyphoonTracking show improved forecasting accuracy, while prototype deletion links local feature contributions to changes in prediction error. This work advances the development of more accurate and interpretable data-driven weather forecasting models. Project page: https://zhixuanli.github.io/projects/li2026protocast
Figures & tables
Figure 1: Motivation for local prototype-conditioned forecasting. (a) A general forecaster represents pattern knowledge implicitly in shared parameters. (b) Prototype matching provides information for sample- or task-level prediction. (c) ProtoCast uses group-specific prototype banks, with one group expanded and the others summarized. Locations A, B, and C form distinct operator mixtures from the same group-specific bank. The operators process local features to produce the orange weighted contributions. At each location, the contributions within each group are summed and added to its features before cross-variable coordination. Curves beside the prototypes depict representative histories within the expanded group. Solid gray curves show historical inputs. Dashed orange and solid blue curves denote ground truth and predictions , respectively.
Figure 2: ProtoCast with four variable groups. (a) Groups use independent encoders, decoders, and prototype banks with shared temporal and spatial modules. Cross-variable attention provides matching context, while gated fusion exchanges corrected features. The last input supplies persistence. Panels (b) and (c) detail one group. (b) History, temporal change, and cross-variable forecast context yield sparse, patch-specific weights shared across forecast leads. (c) Mixed operators map lead-conditioned low-rank features to residuals with explicit prototype contributions before fusion. The bottom strip shows group-wise refitting of matching statistics and keys before each epoch. Previous assignments keep keys paired with their operators across refits.
Method
Venue
T2M
R
UV10
TCC
1st Count
MSE ↓
MAE ↓
RMSE ↓
R2↑
MSE ↓
MAE ↓
RMSE ↓
R2↑
MSE ↓
MAE ↓
RMSE ↓
R2↑
MSE ↓
MAE ↓
RMSE ↓
R2↑
ConvLSTM [ 34 ]
NeurIPS’15
2.8330
1.2108
1.6831
0.9936
22.2595
3.4032
4.7180
0.9193
1.9612
0.9698
1.4004
0.9155
0.0211
0.1050
0.1454
0.7695
0
PredRNN [ 39 ]
NeurIPS’17
2.4768
1.1439
1.5738
0.9944
19.0061
3.0984
4.3596
0.9311
1.6057
0.8767
1.2671
0.9308
0.0200
0.1025
0.1413
0.7824
0
MAU [ 5 ]
NeurIPS’21
4.5610
1.4979
2.1357
0.9897
30.4721
4.0525
5.5202
0.8896
2.5267
1.1423
1.5895
0.8911
0.0234
0.1084
0.1531
0.7444
0
SimVP [ 11 ]
CVPR’22
3.4210
1.3681
1.8496
0.9923
26.7747
3.8266
5.1744
0.9030
2.4634
1.1450
1.5695
0.8938
0.0291
0.1283
0.1707
0.6825
0
ConvNeXt [ 21 ]
CVPR’22
2.7980
1.2215
1.6727
0.9937
23.1956
3.5768
4.8162
0.9159
1.7352
0.9702
1.3173
0.9252
0.0242
0.1151
0.1554
0.7366
0
Table 1: Global multivariate forecasting on GlobalForecast. Models use 12 hourly observations to predict leads of 1-12 hours. Scores are computed after inverse normalization. Lower errors and higher R2 indicate better predictions. The best and second-best values are highlighted. The 1st Count column records the number of best scores for each method.
Figure 3: Qualitative comparison of 12-hour forecasts on the GlobalForecast benchmark. The left panels show ground truth (GT) and predictions from TAU, Met 2 Net, PredFormer, and Ours. The right panels show absolute errors.
Figure 5: Typhoon track prediction for Doksuri (2023), comparing Met 2 Net and Ours. (a) Three-hour predicted centers and ground truth from ERA5 MSL fields. (b) Mean Haversine track error at leads of 1-12 hours (lower is better).
Variant
MSL
U10
V10
T2M
w/o correction
6074.21
0.9993
1.1361
0.5560
Plain adapter
6117.54
0.9957
1.1414
0.5548
Uniform weights
6124.51
1.0138
1.1481
0.5605
w/o pre-attention
6326.96
1.0145
1.1585
0.5758
w/o post-fusion
6258.18
1.0101
1.1713
0.5626
Fixed gate ( γ=1 )
5993.01
1.0006
1.1405
0.5491
Figure 4: Core design ablations on TyphoonTracking. MSE is reported for each variable, with lower values indicating better performance. Best and second-best values are marked.
Method
Venue
MSL
U10
V10
T2M
1st Count
MSE ↓
MAE ↓
RMSE ↓
R2↑
MSE ↓
MAE ↓
RMSE ↓
R2↑
MSE ↓
MAE ↓
RMSE ↓
R2↑
MSE ↓
MAE ↓
RMSE ↓
R2↑
ConvLSTM [ 34 ]
NeurIPS’15
20742.42
103.0576
144.0223
0.9642
2.1290
1.0321
1.4591
0.8879
2.3270
1.0849
1.5254
0.8736
2.3518
1.0257
1.5335
0.9758
0
PredRNN [ 39 ]
NeurIPS’17
9613.80
68.6191
98.0500
0.9834
1.3928
0.8253
1.1802
0.9267
1.5688
0.8810
1.2525
0.9148
1.0492
0.7064
1.0243
0.9892
0
MAU [ 5 ]
NeurIPS’21
10361.82
69.3611
101.7930
0.9821
1.5269
0.8564
1.2357
0.9196
1.7368
0.9131
1.3179
0.9057
1.0331
0.7067
1.0164
0.9894
0
SimVP [ 11 ]
CVPR’22
16821.50
95.4607
129.6977
0.9710
2.1898
1.0740
1.4798
0.8847
2.4468
1.1429
1.5642
0.8671
1.6990
0.9274
1.3035
0.9825
0
ConvNeXt [ 21 ]
CVPR’22
12387.68
82.9830
111.3000
0.9786
1.3465
0.8196
1.1604
0.9291
1.5473
0.8750
1.2439
0.9160
0.9961
0.7120
0.9981
0.9898
0
Table 2: Regional multivariate forecasting on TyphoonTracking. Models use 12 hourly observations to predict leads of 1-12 hours. Metrics for MSL, U10, V10, and T2M are computed after inverse normalization. Lower errors and higher R2 indicate better predictions. The best and second-best values are highlighted. The 1st Count column records the number of best scores for each method.
Figure 5: Interpreting a local T2M prototype contribution. (a) PCA projection of training queries, prototype keys, and the current query. (b) Current and retrieved histories, with original-space prototype matching weights. (c) MAE of T2M w/ and w/o the local weighted Prototype 27 contribution. MAE averages gridwise errors over the 4×4 patch at each lead. Lower MAE is better.
Benchmark
Parameters
s/epoch
50 epochs (h)
GlobalForecast
4,545,645
335.10
4.65
TyphoonTracking
4,146,269
296.48
4.12
Table S1 : Model size and training time of the complete ProtoCast on both benchmarks. Training uses one NVIDIA H200 GPU, batch size 128, and BF16 precision.
Model
Parameters
Peak memory
(GB)
w/o correction
3,738,461
33.60
Plain adapter
4,145,501
34.66
ProtoCast
4,146,269
34.70
Table S2 : Trainable parameters and peak training memory for the complete model and ablated variants on TyphoonTracking.
Method
Architecture configuration
ConvLSTM
Four recurrent layers, 128 hidden channels per layer, kernel size 5, patch size 4.
PredRNN
Four recurrent layers, 128 hidden channels per layer, kernel size 5, patch size 4.
MAU
Four recurrent layers, 64 hidden channels per layer, temporal window 5, spatial reduction factor 4.
SimVP
Four spatial blocks and eight IncepU translator blocks, with hidden widths of 32 and 256.
ConvNeXt
SimVP framework with ConvNeXt translator blocks.
HorNet
SimVP framework with HorNet translator blocks.
Table S3: Baseline architecture configurations shared across GlobalForecast and TyphoonTracking. Input and output dimensions follow the corresponding benchmark.
Variant
MSL
U10
V10
T2M
Uniform weights
6124.51
1.0138
1.1481
0.5605
Softmax
5973.97
0.9903
1.1316
0.5543
Initial refit only
6145.60
0.9972
1.1405
0.5582
Refit for 40 epochs
5954.62
0.9991
1.1537
0.5528
Ours
5888.81
0.9886
1.1292
0.5529
Table S4 : Matching and refitting ablations on TyphoonTracking, measured by MSE. Ours uses entmax and refits before every epoch. Best and second-best values are marked.
Variable
Ours
No history change
No forecast context
MSL (Pa 2 )
5888.8075
5993.1686 ( +1.77% )
6032.4644 ( +2.44% )
U10 (m 2 s -2 )
0.9886
1.0067 ( +1.82% )
0.9842 ( −0.45% )
V10 (m 2 s -2 )
1.1292
1.1745 ( +4.01% )
1.1403 ( +0.98% )
T2M (K 2 )
0.5529
0.5634 ( +1.90% )
0.5483 ( −0.83% )
Mean normalized MSE ( ×10−4 )
1.816433
1.866009 ( +2.73% )
1.821510 ( +0.28% )
Table S5 : Matching-descriptor ablations on TyphoonTracking. Parentheses give percentage changes from Ours. The final row reports mean normalized MSE, scaled by 104 . Lower is better. Best and second-best values are marked.
Variable
Ours
Learned router
No slot tracking
MSL (Pa 2 )
5888.8075
6130.3865 ( +4.10% )
6032.5248 ( +2.44% )
U10 (m 2 s -2 )
0.9886
1.0047 ( +1.62% )
0.9890 ( +0.04% )
V10 (m 2 s -2 )
1.1292
1.1480 ( +1.67% )
1.1435 ( +1.27% )
T2M (K 2 )
0.5529
0.5580 ( +0.93% )
0.5588 ( +1.06% )
Mean normalized MSE ( ×10−4 )
1.816433
1.847644 ( +1.72% )
1.831505 ( +0.83% )
Table S6 : Routing and slot-tracking controls on TyphoonTracking, measured by MSE. Parentheses give percentage changes from Ours. The final row reports mean normalized MSE, scaled by 104 . Lower is better. Best and second-best values are marked.
Variant
MSL
U10
V10
T2M
K=8
6059.33
0.9973
1.1522
0.5545
K=16
6002.10
0.9937
1.1528
0.5538
K=32 (Ours)
5888.81
0.9886
1.1292
0.5529
K=64
5986.45
0.9855
1.1301
0.5467
Table S7 : Prototype-bank size and correction rank on TyphoonTracking. Entries are MSEs. Best and second-best values are marked. Rankings are computed within each panel.
Variant
MSL
U10
V10
T2M
τ=0.1
6021.44
0.9896
1.1468
0.5465
τ=0.2 (Ours)
5888.81
0.9886
1.1292
0.5529
τ=0.4
6029.71
0.9826
1.1370
0.5572
Table S8 : Matching temperature and dimension on TyphoonTracking. Entries are MSEs. Best and second-best values are marked. Rankings are computed within each panel.
Variant
MSL
U10
V10
T2M
1st Count
MSE ↓
MAE ↓
RMSE ↓
R2↑
MSE ↓
MAE ↓
RMSE ↓
R2↑
MSE ↓
MAE ↓
RMSE ↓
R2↑
MSE ↓
MAE ↓
RMSE ↓
R2↑
w/o correction
6074.21
53.7532
77.9372
0.9895
0.9993
0.6842
0.9996
0.9474
1.1361
0.7280
1.0659
0.9383
0.5560
0.4811
0.7457
0.9943
0
Plain adapter
6117.54
53.9834
78.2147
0.9894
0.9957
0.6813
0.9979
0.9476
1.1414
0.7289
1.0684
0.9380
0.5548
0.4814
0.7449
0.9943
0
Uniform weights
6124.51
54.1195
78.2592
0.9894
1.0138
0.6899
1.0069
0.9466
1.1481
0.7316
1.0715
0.9376
0.5605
0.4834
0.7487
0.9942
0
w/o pre-attention
6326.96
54.8428
79.5422
0.9891
1.0145
0.6900
1.0072
0.9466
1.1585
0.7362
1.0763
0.9371
0.5758
0.4897
0.7588
0.9941
0
w/o post-fusion
6258.18
54.3255
79.1087
0.9892
1.0101
0.6870
1.0050
0.9468
1.1713
0.7370
1.0823
0.9364
0.5626
0.4834
0.7500
0.9942
0
Table S9 : Complete ablation results on TyphoonTracking. Lower errors and higher R2 indicate better performance. The best and second-best values are highlighted. The 1st Count column records the number of best scores.
Figure S1 : Interpreting a local temperature correction on GlobalForecast. (a) Three training histories associated with each of two active prototypes, shown relative to the first observation. (b) Matching weights α and mean weighted-residual norms m . (c) Patch-mean temperature forecasts with the full model, the largest- m contribution removed, or a random other active contribution removed, alongside ground truth. Random-deletion forecasts average three draws. Matching weights stay fixed in both removal tests.
Group
N
Largest- m
Random
Increased
Global T2M
96
+0.13608
+0.00723
83/96
Global R
96
+0.44656
+0.00950
80/96
Global UV10
96
+0.10028
+0.00591
87/96
Global TCC
96
+0.00137
-0.00003
61/96
Regional T2M
45
+0.03150
-0.00193
31/45
Table S10 : Local error changes after prototype deletion. Largest- m and random report mean ΔMAE in each variable’s physical units. The last column counts cases with increased error after largest- m deletion. R uses percentage points and TCC a fraction.
Modified group
Removed term
T2M
R
UV10
TCC
D
Δ MAE
D
Δ MAE
D
Δ MAE
D
Δ MAE
T2M
Largest- m
0.448038
+0.136081
0.229989
+0.018922
0.074353
+0.006112
0.007780
+0.000114
T2M
Random
0.067408
+0.007234
0.039725
-0.001520
0.011510
+0.000700
0.001279
-0.000026
R
Largest- m
0.073885
+0.008201
1.296052
+0.446564
0.037506
+0.002874
0.005317
+0.000123
R
Random
0.008442
+0.000006
0.140833
+0.009496
0.005244
+0.000098
0.000716
+0.000023
UV10
Largest- m
0.047069
+0.002947
0.223304
+0.015973
0.328560
+0.100281
0.007632
+0.000694
Table S11 : Cross-variable responses to local prototype deletion on GlobalForecast, averaged over the same 96 cases for each condition. D is the mean absolute forecast response and Δ MAE is the error after deletion minus the full-model error. T2M, R, UV10, and TCC use K, percentage points, m s -1 , and a fraction, respectively.
Modified group
Removed term
MSL
U10
V10
T2M
D
Δ MAE
D
Δ MAE
D
Δ MAE
D
Δ MAE
T2M
Largest- m
2.000565
+0.149463
0.019452
-0.001806
0.027219
+0.002074
0.148729
+0.031502
T2M
Random
0.417193
-0.019567
0.003102
-0.000138
0.003632
+0.000052
0.027290
-0.001931
Table S12 : Cross-variable responses to temperature-prototype deletion on TyphoonTracking over 45 paired cases. Definitions follow Table S11 . MSL uses Pa, U10 and V10 use m s -1 , and T2M uses K.
Figure S2 : Local temperature contributions and forecast responses on TyphoonTracking. (a) Matching weights α and mean weighted-residual norms m for the active prototypes. (b) Patch-mean forecasts from the full model and the two deletion conditions, alongside ground truth. The random-deletion curve averages three draws from the other active prototypes.
Figure S3 : Prototype-bank statistics over 50 training epochs on GlobalForecast (top) and TyphoonTracking (bottom). Columns show agreement of tracked slot assignments, usage perplexity, and the largest pairwise key cosine similarity. Statistics use queries from fixed training reference samples. No empty slots occur in either benchmark. Vertical dotted lines mark the validation-selected checkpoints.
Within RMS
Between RMS
Group
D
Dref
D
Dref
Reduction
T2M
1.2915
1.7536
1.3362
0.6151
26.4%
R
3.6392
5.1764
4.2837
2.1907
29.7%
UV10
0.7228
0.7924
0.3247
0.0021
8.8%
TCC
0.05786
0.05902
0.01170
0.00076
2.0%
Table S13 : Physical-history dispersion on GlobalForecast. D and Dref denote prototype and matched-reference RMS, respectively. Reduction refers to within-prototype RMS. T2M, R, UV10, and TCC use K, percentage points, m s -1 , and a fraction.
Figure S4 : Prototype-associated physical histories on GlobalForecast. Rows show T2M, R, UV10, and TCC. Left panels show the prototype with the greatest total matching-weight mass in each group. Curves are weighted medians, and shaded regions span the pointwise 10th-90th percentiles. Each history is relative to its final observation. UV10 retains U10 and V10 as separate components. Right panels show the within-prototype RMS divided by its matched-reference RMS for all 64 prototypes. Values below one indicate more concentrated histories. Marker area is proportional to matching-weight mass, and black rings identify the examples at left.
Variable
Pairing
MSE ↓
MAE ↓
RMSE ↓
R2↑
GlobalForecast
T2M
Original
0.7850
0.5611
0.8860
0.9982
Permuted
3.6588±0.2986
1.1443±0.0357
1.9113±0.0792
0.9917±0.0007
R
Original
10.5224
2.2646
3.2438
0.9619
Permuted
24.9588±1.1772
3.3368±0.0541
4.9946±0.1172
0.9096±0.0043
UV10
Original
0.7626
0.6013
0.8733
0.9671
Table S14: Prototype-operator pairing test. Original denotes the unchanged pairing, and Permuted reports the mean ± sample standard deviation over ten permutations. Each metric is first computed over the complete test set for one permutation. Lower MSE, MAE, and RMSE and higher R2 indicate better forecasts. The better value within each pair is blue and bold, and the other mean is underlined.
Persistence
ProtoCast
MSE
Variable
MSE ↓
MAE ↓
RMSE ↓
R2↑
MSE ↓
MAE ↓
RMSE ↓
R2↑
reduction
GlobalForecast
T2M
5.9785
1.2859
2.4451
0.9865
0.7850
0.5611
0.8860
0.9982
86.87%
R
65.2149
4.9748
8.0756
0.7637
10.5224
2.2646
3.2438
0.9619
83.87%
UV10
3.9284
1.1971
1.9820
0.8307
0.7626
0.6013
0.8733
0.9671
80.59%
TCC
0.0302
0.1137
0.1739
0.6704
0.0165
0.0892
0.1285
0.8199
45.36%
Table S15: Comparison with persistence on both benchmarks. Metrics aggregate the complete test sets and all forecast leads. Lower MSE, MAE, and RMSE and higher R2 indicate better forecasts. The final column reports relative MSE reduction.
Equal-grid
Area-weighted
Variable
Method
MSE ↓
MSE ↓
MAE ↓
RMSE ↓
R2↑
T2M
Persistence
5.9785
6.8438
1.3327
2.6161
0.9700
TAU
2.3887
1.9577
0.9995
1.3992
0.9914
Met2Net
0.9608
0.8038
0.6102
0.8965
0.9965
PredFormer
1.1814
0.9798
0.6813
0.9899
0.9957
ProtoCast
0.7850
0.6560
0.5046
0.8100
0.9971
Table S16: Area-weighted evaluation on GlobalForecast. Equal-grid MSE uses the same predictions for comparison. Area-weighted metrics aggregate the complete test set and all forecast leads. The best area-weighted value for each variable and metric is blue and bold, and the second best is underlined.
Figure S5 : Forecast error changes after removing one prototype contribution or applying a norm-matched random perturbation. Panels (a)-(d) show GlobalForecast T2M, R, UV10, and TCC, respectively. Panel (e) shows TyphoonTracking T2M. Each curve reports the percentage change in mean patch MAE relative to the full prediction at each lead, using Equation ( S15 ). Positive values indicate worse predictions. Blue denotes contribution deletion, and orange denotes the random control.
Benchmark
Variable
Cases
Full MAE
Deletion MAE
Random MAE
MAE increase
Cases worse
GlobalForecast
T2M (K)
96
0.5831
0.7192
1.0743
23.3%
86.5%
GlobalForecast
R (pp)
96
2.2370
2.6835
3.0101
20.0%
83.3%
GlobalForecast
UV10 (m s -1 )
96
0.6147
0.7150
0.8579
16.3%
90.6%
GlobalForecast
TCC
96
0.0894
0.0908
0.0945
1.5%
63.5%
TyphoonTracking
T2M (K)
48
0.4475
0.4823
0.4758
7.8%
68.8%
Table S17: Effect of removing one prototype contribution. MAE uses the selected patch and all 12 leads. Random denotes ten norm-matched directions averaged within each case. The last two columns compare deletion with the full prediction. R uses percentage points, and TCC uses a fraction.
Selected patch
Full output field
Modified group
Output
Mean di
Median di
Cases di>0
Mean di
Median di
Cases di>0
GlobalForecast, 96 cases per modified group
T2M
T2M (K)
-0.3551
-0.3577
4.2%
-0.0029
-0.0029
5.2%
T2M
R (pp)
-0.0080
-0.0093
42.7%
−7.719×10−5
−8.447×10−5
40.6%
T2M
UV10 (m s -1 )
-0.0020
-0.0012
42.7%
−2.174×10−5
−2.381×10−5
40.6%
T2M
TCC
-0.0013
-0.0009
35.4%
−9.735×10−6
−5.810×10−6
36.5%
Table S18: Complete 12-lead paired random-control comparison. Differences use each output variable’s physical units. Patch and full-field columns use the same cases and feature modifications. The case percentages compare deletion with the case’s mean random control.
Run
Method
MSL (Pa)
U10 (m/s)
V10 (m/s)
T2M (K)
1
Ours
53.1169
0.6793
0.7248
0.4804
w/o correction
53.7532
0.6842
0.7280
0.4811
Plain adapter
53.9834
0.6813
0.7289
0.4814
2
Ours
53.7829
0.6879
0.7343
0.4876
w/o correction
54.7514
0.6973
0.7443
0.5032
Plain adapter
54.4896
0.7104
0.7544
0.5063
Table S19 : MAE for three matched TyphoonTracking training runs. Lower is better. The final block reports arithmetic means.
Figure S15 : Typhoon track prediction for Mawar (2023), comparing TAU and Ours. (a) Three-hour predicted centers and ground-truth (GT) centers extracted from ERA5 target MSL fields. (b) Mean Haversine track error at leads of 1-12 hours. Both panels use 16 forecasts initialized at six-hour intervals.
Weather forecasting predicts future atmospheric conditions, from local storms to global patterns, using physical models, observations, and machine learning.
Most deep learning weather models assign a fixed set of variables and pressure levels to predefined channels, limiting transfer across atmospheric field configurations. This dependence on a fixed field set limits the transferability of trained models across atmospheric field configurations. We propose FlexCast, a field-adaptive weather forecasting model that uses a single set of parameters to produce identity-aligned forecasts for variable-cardinality subsets drawn from a 69-field ERA5 registry. Specifically, a metadata-conditioned adapter the first encodes variable identity, pressure level, and field type and combines them with spatial features. Then, shared rank-16 projec?tions are modulated by metadata-dependent gates to produce field?specific features, while masked set fusion aggregates the available fields into a fixed-width representation. Subsequently, a multiscale U-Transformer processes the fused atmospheric features, while an identity-aware query decoder produces forecasts for the requested fields. Finally, FlexCast learns a standardized six-hour increment and applies it recursively to generate forecasts at longer lead times. Experiments on the 2020 ERA5 test set demonstrate that FlexCast operates across varying field configurations. Compatible cross-field context is associated with lower forecast errors, whereas mismatched context increases them.
Yuang Zhang, Chen Hui, Weisi Lin +4
Nanjing University of Information Science and Technology · Nanyang Technological University · Harbin Institute of Technology +1
Data-driven weather forecasting models are commonly trained on ERA5, yet operational deployment must use timely initial states from physics-based numerical weather prediction (NWP) systems whose distributions differ substantially from ERA5. Direct initialization with these states can degrade autoregressive forecasts, motivating input adaptation that keeps the pretrained forecasting model unchanged. Here, we introduce WeatherBridge, a unified, source-conditioned adapter that maps heterogeneous operational NWP initial states to ERA5 reference states without updating the pretrained forecasting model. WeatherBridge predicts the structured residual corrections with a Swin-UNet and injects source embeddings across multiple spatial scales, allowing the model to capture shared atmospheric structure and source-specific discrepancies. Its training objective combines latitude-weighted ERA5 reconstruction with One-step Rollout Alignment (ORA), which backpropagates one-step forecast error through a frozen FuXi model to optimize the adapted states for both initial state fidelity and downstream forecast compatibility. We evaluate WeatherBridge using daily GFS, HRES, and CMA initializations from 1 January to 22 November 2025 across six variables and lead times up to 240 hours. Relative to raw source initialization of FuXi, WeatherBridge reduces the mean RMSE of FuXi over the 6- 240-hour rollout for 17 of 18 source-variable combinations, with a median reduction of 10.2% and a maximum reduction of 46.3%. The adapter trained with frozen FuXi also improves FengWu forecasts on GFS inputs without retraining either the adapter or FengWu. These results demonstrate that forecast-aware input adaptation can improve forecasts from heterogeneous operational NWP inputs and transfer across pretrained ERA5-based forecasting models.
Mutian Xi, Xinyu Su, Xiaohui Zhong +1
Artificial Intelligence Innovation and Incubation Institute, Fudan University · Shanghai Academy of AI for Science · FuXi Intelligent Computing Technology Co., Ltd.
Recent probabilistic weather forecasters train stochastic predictors with the continuous ranked probability score (CRPS) to generate each ensemble member in a single forward pass. These models learn the predictive distribution from the forecast context alone, which becomes difficult at longer forecast horizons where uncertainty is high. To learn the predictive distribution more effectively, we introduce auxiliary conditional denoising tasks that predict the same future state from the context and its corrupted version, which provides partial future information that can reduce prediction ambiguity. Building on distributional diffusion models, we learn the conditional distributions of these tasks with a single stochastic predictor by minimizing a proper scoring rule across noise levels. At inference, the predictor can still generate each ensemble member in a single forward pass at the fully corrupted endpoint. Standard CRPS training is recovered as the endpoint-only special case of our formulation, so our framework extends existing CRPS-based forecasters with only additional conditioning inputs. Controlled experiments show that the auxiliary tasks improve one-step forecasting across architectures, with larger gains at longer forecast horizons. The gains extend to high-dimensional global weather forecasting under both training from scratch and fine-tuning, along with improved calibration and potential benefits for generalization under distribution shift.