High-resolution weather prediction is important for resolving local atmospheric patterns missed by coarse global forecasts. Large pretrained weather backbones have improved global forecasting, but adapting them to produce future high-resolution fields from recent coarse states remains challenging and costly. We formulate this setting as Predictive Spatiotemporal Weather Downscaling (PSWD), a task that jointly refines spatial and temporal resolution while predicting future high-resolution trajectories. To this end, we propose a framework that decomposes prediction into spatial and temporal stages using a shared pretrained backbone, fixed numerical scaffolds, and lightweight residual heads. For backbone adaptation, we introduce PROLA (PRincipal-Orthogonal Low-rank Adaptation), which allocates a fixed low-rank budget between pretrained weights' principal directions and their orthogonal complement. PROLA further rescales rank-1 optimizer updates using gradient signals, improving adaptation without increasing trainable rank. Experiments in the PSWD setting show that PROLA outperforms representative low-rank adaptation baselines. Further experiments across diverse datasets and tasks support the broader applicability of PROLA, showing that its benefits extend beyond the primary weather downscaling setting.
Figures & tables
Figure 1: Overview of the proposed scaffold-and-correct pipeline for PSWD. Fixed numerical operators Us and Ut and WFM backbone B build spatial and temporal scaffolds, while the residual heads Cs and Ct provide learned corrections.
Figure 2: Overview of PROLA. (a) A fixed rank budget r=rp+ro is split into a principal branch ΔWp=spBpAp , initialized from the top singular components of W0 , and an orthogonal branch ΔWo=soP⊥BoAoQ⊥ , which captures corrections outside the selected principal subspaces. The effective weight combines the frozen residual weight and the two trainable branches as Weff=Wres+ΔWp+ΔWo . (b) Updates from the base optimizer are scaled using gradient-guided branch and component scores. For each factor pair, αkg scales (Δ~bkg,Δ~akg) , yielding (Δbkg,Δakg) .
Method
Trainable Params
Z500
T850
T2M
U10
V10
RMSE
ACC
RMSE
ACC
RMSE
ACC
RMSE
ACC
RMSE
ACC
3h lead
Full FT
1.26B
60.13 [-1pt] ± 1.61
0.998 [-1pt] ± 0.000
0.813 [-1pt] ± 0.002
0.987 [-1pt] ± 0.000
1.064 [-1pt] ± 0.009
0.979 [-1pt] ± 0.000
0.902 [-1pt] ± 0.004
0.975 [-1pt] ± 0.000
0.925 [-1pt] ± 0.009
0.974 [-1pt] ± 0.001
Head only
1.07M
94.93 [-1pt] ± 0.49
0.996 [-1pt] ± 0.000
1.049 [-1pt] ± 0.006
0.978 [-1pt] ± 0.000
2.083 [-1pt] ± 0.009
0.915 [-1pt] ± 0.001
1.648 [-1pt] ± 0.018
0.914 [-1pt] ± 0.001
1.563 [-1pt] ± 0.018
0.924 [-1pt] ± 0.002
LoRA
3.92M
54.72 [-1pt] ± 0.36
0.999 [-1pt] ± 0.000
0.898 [-1pt] ± 0.002
0.984 [-1pt] ± 0.000
1.199 [-1pt] ± 0.007
0.973 [-1pt] ± 0.000
1.219 [-1pt] ± 0.008
0.954 [-1pt] ± 0.000
1.181 [-1pt] ± 0.002
0.957 [-1pt] ± 0.000
DoRA
4.16M
54.32 [-1pt] ± 0.35
0.999 [-1pt] ± 0.000
0.893 [-1pt] ± 0.002
0.984 [-1pt] ± 0.000
1.191 [-1pt] ± 0.006
0.973 [-1pt] ± 0.000
1.209 [-1pt] ± 0.008
0.955 [-1pt] ± 0.000
1.173 [-1pt] ± 0.001
0.958 [-1pt] ± 0.000
Table 1: Quantitative results on WeatherBench 2 [ 16 ] . Results are reported at 3h and 6h lead times. We report latitude-weighted RMSE ( ↓ ) and ACC ( ↑ ). Each cell shows the mean above and ± standard deviation below, over three seeds. Best displayed mean results among PEFT methods are boldfaced.
Figure 3: Spatial spectral fidelity. Prediction-to-target ratios of mean spherical harmonic power for T2M at 6h over test cases. A ratio of one matches target power.
Table 5
Figure 4: Qualitative ERA5-to-CERRA prediction at 6h. T2M predictions from LoRA and PROLA, and the CERRA target, valid at 12:00 UTC on 4 February 2018.
3h lead
6h lead
Method / Variant
Z500
T850
T2M
U10
V10
Z500
T850
T2M
U10
V10
Direct prediction
LoRA
58.91
0.979
1.256
1.481
1.484
129.60
1.214
1.394
1.755
1.781
PiSSA
56.47
0.942
1.191
1.272
1.239
128.00
1.152
1.292
1.392
1.405
OPLoRA
58.15
0.983
1.254
1.463
1.457
131.16
1.219
1.388
1.741
1.735
PROLA
53.13
0.898
1.165
1.211
1.177
125.07
1.095
1.289
1.358
1.344
Table 4: Direct prediction and pipeline ablation. Latitude-weighted RMSE ( ↓ ) at 3h and 6h. Best results in the pipeline ablation are boldfaced.
3h lead
6h lead
Configuration
rp
ro
Scaling
Z500
T850
T2M
U10
V10
Z500
T850
T2M
U10
V10
(i) Branch composition and update scaling
Principal
8
0
–
53.66
0.894
1.193
1.209
1.178
78.55
0.987
1.248
1.294
1.290
Principal
8
0
✓
53.06
0.887
1.184
1.198
1.166
77.71
0.979
1.241
1.289
1.285
Orthogonal
0
8
–
54.54
0.899
1.197
1.216
1.188
79.83
0.994
1.272
1.314
1.314
Orthogonal
0
8
✓
52.81
0.886
1.175
1.189
1.157
76.67
0.977
1.260
1.284
1.278
Table 5: Ablation studies on PROLA design. Latitude-weighted RMSE ( ↓ ) at 3h and 6h. The design ablations use rp+ro=8 . ✓ marks indicate update scaling; ρ and π denote branch score and component score. Best results are boldfaced.
Figure 5: Blockwise update energy. Fractions of total update energy in the principal, cross, and orthogonal blocks.
Appendix figures & tables32 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Layer-wise concentration of rank-1 update scales. Each bar extends from uniform scaling ( Hℓ=1 ) to the mean normalized entropy of one attention block. Blocks are ordered from encoder to decoder.
Method
Trainable parameters
Head only
1,068,684
LoRA
3,919,500
DoRA
4,157,068
AdaLoRA
5,346,060
PiSSA
3,919,500
OPLoRA
3,919,500
Appendix
Table 6: Trainable parameter counts for the main Aurora configuration. Counts include backbone adaptation and correction-head parameters.
Method
Trainable Params
Variable
1h
2h
3h
4h
5h
6h
Z500
48.22 ± 0.67
49.44 ± 1.39
60.13 ± 1.61
72.88 ± 1.82
85.63 ± 2.27
106.68 ± 2.63
T850
0.903 ± 0.011
0.834 ± 0.004
0.813 ± 0.002
0.874 ± 0.002
0.966 ± 0.003
1.114 ± 0.008
T2M
1.472 ± 0.003
1.225 ± 0.004
1.064 ± 0.009
0.955 ± 0.015
1.044 ± 0.019
1.323 ± 0.022
U10
1.239 ± 0.013
1.052 ± 0.008
0.902 ± 0.004
0.873 ± 0.003
0.941 ± 0.007
1.135 ± 0.012
Full FT
1.26B
V10
1.310 ± 0.024
1.094 ± 0.017
0.925 ± 0.009
0.894 ± 0.003
0.984 ± 0.003
1.209 ± 0.010
Head only
1.07M
Z500
68.16 ± 0.72
78.45 ± 0.40
94.93 ± 0.49
112.98 ± 0.68
128.19 ± 0.97
144.03 ± 1.47
Appendix
Table 7: Full RMSE ( ↓ ) results on WeatherBench 2 [ 16 ] . We report latitude-weighted RMSE at all lead times from 1h to 6h as mean ± sample standard deviation over three seeds (0, 1, and 2). Best displayed mean results among PEFT methods are boldfaced, including ties after rounding.
Method
Trainable Params
Variable
1h
2h
3h
4h
5h
6h
Z500
0.999 ± 0.000
0.999 ± 0.000
0.998 ± 0.000
0.997 ± 0.000
0.996 ± 0.000
0.995 ± 0.000
T850
0.984 ± 0.000
0.986 ± 0.000
0.987 ± 0.000
0.985 ± 0.000
0.981 ± 0.000
0.976 ± 0.000
T2M
0.958 ± 0.001
0.972 ± 0.000
0.979 ± 0.000
0.983 ± 0.000
0.980 ± 0.001
0.970 ± 0.001
U10
0.952 ± 0.001
0.966 ± 0.000
0.975 ± 0.000
0.976 ± 0.000
0.973 ± 0.000
0.961 ± 0.001
Full FT
1.26B
V10
0.947 ± 0.002
0.964 ± 0.001
0.974 ± 0.001
0.976 ± 0.000
0.971 ± 0.000
0.956 ± 0.001
Head only
1.07M
Z500
0.998 ± 0.000
0.997 ± 0.000
0.996 ± 0.000
0.994 ± 0.000
0.992 ± 0.000
0.990 ± 0.000
Appendix
Table 8: Full ACC ( ↑ ) results on WeatherBench 2 [ 16 ] . We report latitude-weighted ACC at all lead times from 1h to 6h as mean ± sample standard deviation over three seeds (0, 1, and 2). Best displayed mean results among PEFT methods are boldfaced, including ties after rounding.
ERA5 → CERRA: regional spatiotemporal downscaling
3h lead
6h lead
Method
Z500
T850
T2M
U10
V10
Z500
T850
T2M
U10
V10
LoRA
72.14
1.076
2.150
1.370
1.327
86.33
1.092
2.147
1.373
1.340
PiSSA
71.80
1.072
2.126
1.354
1.311
85.72
1.084
2.114
1.352
1.323
OPLoRA
72.56
1.081
2.132
1.371
1.332
86.13
1.099
2.146
1.375
1.348
PROLA
68.44
1.028
2.021
1.316
1.274
81.41
1.036
2.019
1.319
1.283
Appendix
Table 9: Generalization to regional weather prediction. ERA5-to-CERRA results report latitude-weighted RMSE ( ↓ ) for five variables at 3h and 6h. Regional precipitation results report SEEPS ( ↓ ), RMSE in cm ( ↓ ), and ACC ( ↑ ) at 12h, 24h, and 36h. All methods are trained for 10 epochs and evaluated on the 2018 test split. Best displayed results within each block are boldfaced.
Pretrained
LoRA
PiSSA
OPLoRA
PROLA
EX (%) ↑
45.50
49.54
51.96
49.41
52.87
Appendix
Table 10: Text-to-SQL transfer. Execution accuracy (EX) on BIRD for Qwen3.5-9B.
6h lead
12h lead
Method
Z500
T850
T2M
U10
V10
Z500
T850
T2M
U10
V10
LoRA
138.22
1.241
1.967
2.055
2.079
123.23
1.291
1.757
2.368
3.175
PiSSA
134.41
1.205
1.904
2.556
2.112
121.61
1.247
2.062
2.457
2.272
OPLoRA
139.20
1.237
1.839
2.093
2.097
124.38
1.296
1.760
2.519
3.102
PROLA
133.03
1.192
1.531
2.031
2.077
119.97
1.246
1.614
2.749
2.838
18h lead
24h lead
Appendix
Table 11: Backbone transfer to ArchesWeather. Latitude-weighted RMSE ( ↓ ) on the 2018 test split after one training epoch and total trainable adapter rank 8. PROLA uses rp=ro=4 . Best displayed results within each lead time are boldfaced.
3h lead
6h lead
Method
Z500
T2M
Z500
T2M
LoRA
52.27
1.148
76.14
1.141
PiSSA
54.62
1.171
82.54
1.159
OPLoRA
52.08
1.145
76.25
1.134
PROLA
48.01
1.054
69.58
0.998
Appendix
Table 12: Additional resolution gap. Latitude-weighted RMSE ( ↓ ) on the 2018 test split for 5.625∘ -to- 2.8125∘ downscaling. Best displayed values are boldfaced.
Method
DJF
MAM
JJA
SON
LoRA
82.78
79.28
75.29
79.70
PiSSA
81.70
78.54
74.37
79.25
OPLoRA
83.28
79.88
75.70
79.99
PROLA
75.99
73.22
69.73
73.65
Appendix
Table 13: Seasonal robustness. Latitude-weighted Z500 RMSE ( ↓ ) at 6h on the 2018 test split. Best displayed values are boldfaced.
Method
Upper RMSE ↓
Upper BSS ↑
Lower RMSE ↓
Lower BSS ↑
LoRA
56.87
0.514
129.19
0.780
PiSSA
55.50
0.513
126.28
0.781
OPLoRA
56.63
0.512
127.55
0.776
PROLA
50.80
0.545
116.36
0.795
Appendix
Table 14: Spatial-tail performance. Z500 at 6h on the 2018 test split. Upper and lower tails are defined within each target field as in Appendix D . Best displayed values are boldfaced.
3h lead
6h lead
Method
Z500
T850
T2M
U10
V10
Z500
T850
T2M
U10
V10
LoRA
0.301
0.627
0.272
0.396
0.372
0.077
0.349
0.081
0.286
0.260
PiSSA
0.304
0.619
0.277
0.392
0.369
0.082
0.331
0.079
0.291
0.263
OPLoRA
0.302
0.628
0.277
0.402
0.378
0.077
0.350
0.087
0.290
0.264
PROLA
0.294
0.589
0.229
0.329
0.312
0.072
0.293
0.062
0.248
0.204
Appendix
Table 15: Spherical spectral distance. Mean absolute log-ratio of prediction to target power ( ↓ ), evaluated over degrees 1–59 using all 1,458 test cases in 2018. Best displayed values are boldfaced, including rounded ties.
Figure 7: Spherical power ratios at 3h. Ratios of mean prediction power to mean target power for eight variables, evaluated on all 1,458 test cases in 2018. The dashed horizontal line marks the target ratio of one.
Figure 8: Spherical power ratios at 6h. Ratios of mean prediction power to mean target power for eight variables, evaluated on all 1,458 test cases in 2018. The dashed horizontal line marks the target ratio of one.
Method
SNR ↑
Tendency RMSE ↓
850-hPa wind RMSE ↓
Z500
T2M
Z500
T2M
Divergence
Vorticity
LoRA
30.78
21.46
40.48
0.418
8.70
15.53
PiSSA
30.87
21.62
40.26
0.417
8.61
15.45
OPLoRA
30.73
21.45
40.53
0.418
8.71
15.55
PROLA
31.48
21.87
39.89
0.415
8.46
15.18
Appendix
Table 16: Reconstruction and physical diagnostics at 6h. Results use all 1,458 test cases in 2018. SNR denotes variance-normalized reconstruction SNR (dB); tendency RMSE is in m2s−2 for Z500 and K for T2M; wind-derivative RMSE is in 10−6s−1 . Best displayed values are boldfaced, including rounded ties.
Figure 9: Relative RMSE over long rollouts. Z500 (left) and T2M (right) RMSE normalized by Full FT at each lead time. The dashed horizontal line denotes Full FT at 1.0; lower values indicate lower RMSE.
Method / backbone LR
Z500
T850
T2M
U10
V10
Full FT, 2.5×10−5
116.94
1.116
1.306
1.178
1.235
Full FT, 5×10−5
108.90
1.105
1.304
1.122
1.203
Full FT, 10−4
106.06
1.096
1.324
1.093
1.165
PROLA
73.07
0.958
1.211
1.236
1.234
Appendix
Table 17: Full fine-tuning learning-rate sweep. Latitude-weighted RMSE ( ↓ ) at 6h on the 2018 test split. The main Full FT configuration uses a backbone learning rate of 5×10−5 . Best displayed values are boldfaced, including rounded ties.
Method
Training Time / Epoch
Peak GPU Memory
LoRA
54m 38s
15.56 GiB
DoRA
71m 19s
15.56 GiB
AdaLoRA
58m 45s
15.63 GiB
PiSSA
55m 52s
15.56 GiB
OPLoRA
60m 56s
15.57 GiB
PROLA
67m 18s
15.56 GiB
Appendix
Table 18: Training cost of the PEFT baselines and PROLA. We report the average training time per epoch and the peak per-GPU memory recorded during training.
Figure 10: Qualitative comparison at 2h lead time for multiple variables, showing the target, PROLA, pretrained Aurora, and Full FT predictions.
Figure 11: Qualitative comparison at 4h lead time for multiple variables, showing the target, PROLA, pretrained Aurora, and Full FT predictions.
Figure 12: Qualitative comparison at 6h lead time for multiple variables, showing the target, PROLA, pretrained Aurora, and Full FT predictions.
Figure 13: T2M comparison at 2h lead time between LoRA and PROLA, with absolute-error maps and the pointwise absolute-error difference map.
Figure 14: T2M comparison at 2h lead time between PiSSA and PROLA, with absolute-error maps and the pointwise absolute-error difference map.
Figure 15: T2M comparison at 2h lead time between OPLoRA and PROLA, with absolute-error maps and the pointwise absolute-error difference map.
Figure 16: Z500 comparison at 2h lead time between LoRA and PROLA, with absolute-error maps and the pointwise absolute-error difference map.
Figure 17: Z500 comparison at 2h lead time between PiSSA and PROLA, with absolute-error maps and the pointwise absolute-error difference map.
Figure 18: Z500 comparison at 2h lead time between OPLoRA and PROLA, with absolute-error maps and the pointwise absolute-error difference map.
Figure 19: Qualitative ERA5-to-CERRA wind prediction at 6h. 10 m wind speed from LoRA, PROLA, and the CERRA target, valid at 06:00 UTC on 3 February 2018. All panels share the same color scale and display smoothing.
Method
Variable
RMSE ↓
1h
2h
3h
4h
5h
6h
Z500
101.90
72.91
58.91
72.14
98.30
129.60
T850
1.009
0.973
0.979
1.044
1.118
1.214
T2M
1.419
1.303
1.256
1.230
1.294
1.394
U10
1.426
1.438
1.481
1.567
1.652
1.755
LoRA
V10
1.376
1.417
1.484
1.585
1.677
1.781
Appendix
Table 19: Direct prediction across PEFT methods. We report latitude-weighted RMSE ( ↓ ) at 1–6h. Best displayed values, including ties, are boldfaced.
Method / Variant
Variable
RMSE ↓
ACC ↑
1h
2h
3h
4h
5h
6h
1h
2h
3h
4h
5h
6h
Z500
104.37
72.09
53.13
64.82
92.24
125.07
0.995
0.997
0.999
0.998
0.996
0.992
T850
0.991
0.924
0.898
0.941
1.002
1.095
0.980
0.983
0.984
0.982
0.980
0.976
T2M
1.432
1.245
1.165
1.109
1.174
1.289
0.960
0.970
0.974
0.976
0.974
0.968
U10
1.308
1.247
1.211
1.233
1.271
1.358
0.946
0.951
0.954
0.952
0.949
0.942
Direct (PROLA)
V10
1.245
1.202
1.177
1.203
1.244
1.344
0.953
0.956
0.958
0.956
0.953
0.945
Appendix
Table 20: Full ablation results on scaffold-and-correct pipeline design. We report latitude-weighted RMSE ( ↓ ) and ACC ( ↑ ) at 1–6h. Best displayed values, including ties, are boldfaced.
Configuration
rp
ro
Scaling
Variable
RMSE ↓
ACC ↑
1h
2h
3h
4h
5h
6h
1h
2h
3h
4h
5h
6h
Z500
55.57
50.45
53.66
58.98
63.55
78.55
0.999
0.999
0.999
0.998
0.998
0.997
T850
0.941
0.910
0.894
0.918
0.937
0.987
0.982
0.983
0.984
0.983
0.982
0.980
T2M
1.383
1.258
1.193
1.125
1.152
1.248
0.963
0.969
0.973
0.976
0.975
0.970
U10
1.288
1.240
1.209
1.225
1.241
1.294
0.948
0.952
0.954
0.953
0.952
0.947
Principal
8
0
–
V10
1.227
1.194
1.178
1.204
1.228
1.290
0.954
0.957
0.958
0.956
0.954
0.949
Appendix
Table 21: Full ablation results on PROLA branch composition and update scaling. We report latitude-weighted RMSE ( ↓ ) and ACC ( ↑ ) at 1–6h. Best displayed values, including ties, are boldfaced. All configurations use rp+ro=8 ; checkmarks indicate update scaling.
Configuration
rp
ro
Scaling
Variable
RMSE ↓
ACC ↑
1h
2h
3h
4h
5h
6h
1h
2h
3h
4h
5h
6h
Z500
53.90
48.22
51.46
56.54
60.16
74.19
0.999
0.999
0.999
0.998
0.998
0.997
T850
0.933
0.897
0.876
0.899
0.919
0.970
0.982
0.984
0.985
0.984
0.983
0.981
T2M
1.385
1.229
1.153
1.086
1.127
1.225
0.962
0.971
0.974
0.977
0.976
0.971
U10
1.245
1.193
1.161
1.176
1.193
1.249
0.951
0.955
0.958
0.956
0.955
0.951
Principal-heavy
6
2
✓
V10
1.189
1.152
1.136
1.165
1.189
1.244
0.957
0.960
0.961
0.958
0.957
0.953
Appendix
Table 22: Full ablation results on rank allocation and scaling components. We report latitude-weighted RMSE ( ↓ ) and ACC ( ↑ ) at 1–6h. Best displayed values, including ties, are boldfaced. All configurations use rp+ro=8 ; checkmarks indicate update scaling.
Method
3h lead
6h lead
Z500
T850
T2M
U10
V10
Z500
T850
T2M
U10
V10
Total rank r=8
PiSSA
53.66
0.894
1.193
1.209
1.178
78.55
0.987
1.248
1.294
1.290
OPLoRA
54.54
0.899
1.197
1.216
1.188
79.83
0.994
1.272
1.314
1.314
PROLA (4+4)
50.30
0.866
1.128
1.140
1.119
73.07
0.958
1.211
1.236
1.234
Total rank r=128
Appendix
Table 23: Effect of the total trainable rank. Latitude-weighted RMSE ( ↓ ) at 3h and 6h. PROLA uses equal principal and orthogonal ranks. Best displayed values within each rank budget are boldfaced.
Accurate regional weather prediction requires resolving fine-scale structure while remaining consistent with global dynamics. Traditional limited area models rely on computationally expensive simulations, while many learning-based approaches frame the problem as super-resolution, overlooking statistical and physical mismatches across scales. We propose a foundation-model-driven downscaling framework that learns regional refinements of global forecasts by augmenting a pretrained weather model backbone with lightweight, multi-scale prediction heads operating directly in its latent space. Despite being trained on substantially coarser inputs, the pretrained backbone supports regional adaptation at resolutions corresponding to a two-order-of-magnitude increase in grid-cell resolution, without the need for retraining. The proposed approach uses regional numerical simulations as training targets and is evaluated not only against gridded datasets but also against ground-based weather station observations, enabling analysis of systematic biases between global reanalysis, regional simulations, and in-situ weather station observations. Our experiments show improved accuracy in comparison to NWP on most of the metrics at the fraction of computational cost. Moreover, we observe that building on a latent space of globally pre-trained weather foundation model offers better downscaling capabilities than the standard image-based super-resolution approaches.
Wiktor Kamzela, Jakub Kubiak, Adam Dobosz +5
Pozna´n Supercomputing and Networking Center · Pozna´n University of Technology
High-resolution precipitation information is essential for climate impact assessment, yet global climate models remain too coarse to resolve key small-scale processes. Existing machine learning downscaling methods often require paired low- and high-resolution data for supervised learning, are tied to fixed regions or scale factors during inference, and can be computationally expensive to train and run in physical space. Here we introduce Longwang, a zero-shot latent generative framework for global spatiotemporal precipitation downscaling. Longwang learns a context-conditioned latent generative prior and combines it with a physically informed observation operator through posterior sampling, enabling daily O(10 km) precipitation fields to be generated from monthly O(100 km) inputs. On ERA5 reanalysis, Longwang outperforms standard posterior sampling with an unconditional generative prior in reconstructing fine-scale spatial patterns, preserving temporal coherence, and recovering extreme precipitation intensities. The framework further generalizes to historical climate simulations and future climate projections under substantial distribution shift.
Yue Wang, Daniele Visioni
1*Department of Earth and Atmospheric Sciences, Cornell University, 112 Hollister Drive, Ithaca, 14853, NY, USA.
Deep-learning-based climate downscaling aims to learn relationships from historical low-resolution (LR) and high-resolution (HR) climate data to generate HR climate projections. However, this setting faces a temporal out-of-distribution (OOD) challenge: models trained on historical data are commonly applied to future projections whose distributions may differ substantially from the training period. This study investigates temporal OOD shift for daily temperature downscaling over the Continental United States using paired LR-HR model simulations. We propose a temporal domain-adaptive downscaling framework that combines supervised HR reconstruction on historical data with domain alignment between historical and future climate distributions. Experiments across future validation periods show that the proposed domain-adaptive model consistently outperforms statistical and deep-learning-based bias-correction methods, with the largest gains occurring when the temporal distribution shift is strongest. Spatial analyses indicate stronger improvements over high-elevation and topographically complex regions, along with higher spatiotemporal correlation with the HR target. The extreme analysis shows that domain adaptation also reduces upper-tail temperature bias relative to the non-adaptive model. These results demonstrate that temporal domain adaptation can improve the robustness of HR climate projections under non-stationary climate conditions.
Shuochen Wang, Nishant Yadav, Auroop R. Ganguly
Sustainability and Data Sciences Laboratory, Northeastern University · Microsoft · AI for Climate and Sustainability, The Institute for Experiential AI, Northeastern University