AI-based weather forecasting now rivals traditional physics-based ensembles, but state-of-the-art (SOTA) models rely on specialized architectures and massive computational budgets, creating a high barrier to entry. We demonstrate that such complexity is unnecessary for frontier performance. We introduce \ours, a probabilistic forecaster built on a standard U-Net backbone trained with a simple recipe: deterministic pre-training on Mean Absolute Error followed by short probabilistic fine-tuning on the Continuous Ranked Probability Score (CRPS) using Monte Carlo Dropout for stochasticity. As a result, our model matches or exceeds the probabilistic skill of GenCast and IFS ENS at 1.5∘ resolution while reducing training compute by over 10× compared to leading CRPS-based models and inference latency by over 10× compared to diffusion-based models. U-Cast trains in under 12 H200 GPU-days and generates a 15-day ensemble forecast in 3 seconds. These results suggest that scalable, general-purpose architectures paired with efficient training curricula can match complex domain-specific designs at a fraction of the cost, opening the training of frontier probabilistic weather models to the broader community.
Figures & tables
Figure 1: The Efficiency-Accuracy Pareto Frontier. We visualize forecast skill (y-axis, % improvement over IFS ENS), inference latency (x-axis), and training cost (bubble size). Our model (top-left) achieves state-of-the-art performance while requiring an order of magnitude less compute for training and/or inference compared to leading baselines. See Appendix C.1 for detailed methodology.
Figure 2: 1.5∘ CRPS comparison of U-Cast DeepEns against IFS ENS (left) and GenCast (right). Blue indicates lower (better) CRPS for U-Cast; red indicates baseline superiority. U-Cast broadly outperforms IFS ENS and is competitive with GenCast despite the latter’s finer native resolution ( 0.25∘ ). See text for details.
Figure 3: WeatherBench 2 Comparison ( 1.5∘ resolution). We report the CRPS skill relative to the operational IFS ENS (%, lower is better) as a function of forecast horizon. Baseline scores are sourced from the official leaderboard ( Rasp et al., 2024 ) . Numbers after variable abbreviations denote pressure levels in hPa. Note that the GenCast baseline is the native 0.25∘ model regridded to 1.5∘ (see Section 4.2 for discussion of this resolution asymmetry). U-Cast consistently outperforms baselines on geopotential variables while remaining highly competitive on other fields.
Model
z500
10u
1d
3d
10d
1d
3d
10d
U-Cast
20.3
55.2
256
0.349
0.61
1.55
U-Cast (DE)
19.6
53.5
253
0.345
0.60
1.54
IFS ENS
22.4
58.3
262
0.406
0.69
1.61
GenCast
20.2
54.3
254
0.332
0.60
1.55
ArchesGen
21.2
55.8
254
0.370
0.62
1.55
Table 1: Absolute CRPS results (lower is better) for geopotential at 500hPa and 10m u-component of wind for 1, 3, and 10-day ensemble forecasts. See Figure 3 for more comprehensive results.
Model
Training Cost (in days)
Inference Speed
Coarser Resolution ( ∼1.5∘ )
ArchesWeatherGen
45 (V100)
3.5 min (A100)
NeuralGCM ENS
1280 (TPUv5e)
18.6 sec (TPUv4)
ERDM
20 (H200)
7 min (A100)
Swift
360 (Intel Max1550)
–
MOSAIC
16 (H100)
3 sec (H100)
Table 2: Training compute costs and inference speeds for probabilistic weather forecasting models. Models are grouped by resolution. Inference speed is for one single-member 60-step rollout (15-days for NeuralGCM and IFS ENS, or if a 6h-resolution model). † TPU benchmarks for 1∘ GenCast unavailable, but official documentation indicates a ≈3× speed-up over a H100 GPU at 0.25∘ . ∗ For each of the 4 individual models. 1∘ cost is estimated.
Figure 4: U-Cast ablations. We report CRPS relative to U-Cast (top row) and spread-skill ratio (bottom row; closer to 1 is better).
Figure 5: Curriculum ablation. Validation CRPS ( t850 , 12h lead time) during probabilistic training. The curriculum (orange) fine-tunes from a deterministic checkpoint (dashed blue line) and converges rapidly to a CRPS of 0.218. Training from scratch on CRPS alone (gray) requires >3× more steps to reach comparable performance and plateaus at a worse score (0.225).
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Type
Variable name
Short name
Role
Atmospheric
Geopotential
z
Input/Predicted
Atmospheric
Specific humidity
q
Input/Predicted
Atmospheric
Temperature
t
Input/Predicted
Atmospheric
U component of wind
u
Input/Predicted
Atmospheric
V component of wind
v
Input/Predicted
Atmospheric
Vertical velocity
w
Input/Predicted
Appendix
Table 3: Description of data variables used.
Figure 6: The Efficiency-Accuracy Pareto Frontier. We visualize forecast skill (y-axis, % improvement over IFS ENS in terms of CRPS on the left and RMSE on the right), training cost (x-axis), and inference latency (bubble size). Our model (top-left) achieves state-of-the-art performance while requiring an order of magnitude less compute for training or inference compared to leading baselines.
Figure 7: WeatherBench 2 Comparison ( 1.5∘ resolution): Absolute CRPS. We report the CRPS skill as a function of forecast horizon (lower is better). This figure visualizes the same as Figure 3 but in absolute terms, without normalizing by the performance of IFS ENS.
Figure 8: WeatherBench 2 Comparison ( 1.5∘ resolution): RMSE. We report the 50-member ensemble-mean RMSE skill relative to the operational IFS ENS (%, lower is better) as a function of forecast horizon. Baseline scores are sourced directly from the official leaderboard ( Rasp et al., 2024 ) . Numbers after the variable abbreviations refer to the pressure level in hPa. Note that the GenCast baseline uses native 0.25∘ forecasts regridded to 1.5∘ , which may confer an advantage over models strictly operating at 1.5∘ (like U-Cast, ArchesWeatherGen, and NeuralGCM). Relative to the CRPS results, GenCast performs stronger in terms of RMSE than U-Cast.
Figure 9: WeatherBench 2 Comparison ( 1.5∘ resolution): SSR. We report the Spread-Skill ratio skill as a function of forecast horizon (closer to 1 is better). Baseline scores are sourced directly from the official leaderboard ( Rasp et al., 2024 ) . Numbers after the variable abbreviations refer to the pressure level in hPa. U-Cast generates more overconfident forecasts than the baselines, especially in the 1-to-7-day regime. U-Cast (DE) improves calibration, with SSR rarely falling below 0.85 and achieving better short-range calibration than NeuralGCM ENS for several variables.
Figure 10: Comparison of U-Cast (DE) evaluated on 2022 against IFS ENS (left) and against U-Cast (DE) evaluated on 2020 (right). Blue indicates that U-Cast achieves a lower (better) CRPS, while red favors the baseline. U-Cast consistently outperforms IFS ENS on 91.5% of metrics, with notable exceptions in long-range 2-meter temperature and a few variables at the 12-hour lead time (e.g., a 10.8% deficit in u500 ). On average, U-Cast improves upon IFS ENS by 4.45%, with the strongest gains ( ≈18–22% ) observed at 12-to-36-hour lead times for mslp . These results are comparable to those obtained for the test year 2020 ( Figure 2 ), demonstrating that U-Cast, despite being trained only on 1979–2019 data, generalizes well to more distant years. This robustness is further supported by the right panel, where the maximum degradation of U-Cast (DE; 2022) relative to its 2020 performance remains below 4.5%, concentrated in geopotential variables (especially z500 ), while specific humidity variables even show improved CRPS scores (up to 2%).
Figure 11: Score card comparison of U-Cast (DeepEns) vs. U-Cast. Deep ensembling U-Cast via fine-tuning four different versions of it consistently improves CRPS scores, especially for short-to-mid-range geopotential and stratospheric (by up to 4%; except q50 ) variables.
Figure 12: Curriculum and ensemble-size ablations (full evaluation). Relative CRPS vs. U-Cast across variables and lead times (higher means worse than U-Cast). End-to-end CRPS (orange) trains from scratch without deterministic pre-training; it consistently degrades short-range CRPS by 3–5% and stratospheric variables by 5–15% across all lead times, while recovering or slightly improving long-range scores for select variables (e.g., 2t ). U-Cast ( M=4 ) (blue) doubles the training ensemble size; it yields only marginal improvements despite doubling per-step cost, with modest gains concentrated at long-range stratospheric variables (up to 3% for t50 ).
Figure 13: Spectral density of 10-day forecasts, averaged over mid latitudes ( [25∘,55∘] ). While U-Cast generates realistic spectra for the surface and specific humidity variables, it tends to generate excess power at high frequencies for the other variables.
Figure 14: Example visualizations of U-Cast (second row), the corresponding ground truth (first row), and the bias (last row) for specific humidity at 700 hPa ( q700 ) and forecast lead times 3,7,10, and 14 days.
Figure 15: Example visualizations of U-Cast (second row), the corresponding ground truth (first row), and the bias (last row) for two example variables and forecast lead times 3,7,10, and 14 days.
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.
State-of-the-art medium-range AI weather models can outperform traditional Numerical Weather Prediction (NWP) but require massive training budgets. This restricts usage for under-resourced groups and severely limits fast model iteration. Here we develop Otter Weather, a highly efficient spatiotemporal forecasting model designed to democratise high-performance weather prediction with AI. Evaluated on ERA5 reanalysis data at 1.5° resolution using standard WeatherBench protocols, the Otter family significantly advances the skill-compute Pareto frontier. The deterministic version outperforms the best NWP baseline by 9.6% at a 24-hour lead time while requiring fewer than 3.5 A100-days for training. It provides a 2x efficiency gain over lightweight AI models and a 100-fold reduction in compute compared to resource-intensive frontier architectures. We extend these efficiency gains into probabilistic forecasting by training via the Continuous Ranked Probability Score (CRPS). Scaling to a larger architecture, Otter-XL achieves a 9.7% CRPS improvement over the IFS ENS baseline. This yields an almost two-fold increase in predictive skill over comparable lightweight models at similar compute budgets. Otter-XL also outperforms frontier architectures like GenCast by over 2%, while using an order of magnitude less compute. Finally, Otter is applied out-of-the-box to a complex acoustic scattering PDE task where it outperforms a state-of-the-art foundation modelling approach, suggesting that the advances made here might apply across a range of scientific domains.
Cristiana Diaconu, Jonas Scholz, Aliaksandra Shysheya +4
Generative diffusion models are well-suited for probabilistic precipitation nowcasting, but existing approaches often rely on separately trained compression or deterministic forecasting components and remain costly at inference due to iterative denoising. We introduce Just Weather Scoring (JWS), a single-stage, end-to-end diffusion model which addresses both issues by forecasting directly in radar space and enabling few-step generation. Radar-space modeling greatly simplifies training and inference and eliminates uncertainty arising from lossy compression. JWS combines Masked Asynchronous Diffusion, a timestep-sampling scheme that preserves clean context while adapting diffusion training to high-dimensional spatio-temporal data, with a simple scoring-rule objective that aligns training with probabilistic forecasting and unlocks few-step generation. On the SEVIR and MeteoNet benchmarks, JWS achieves state-of-the-art probabilistic forecasting performance at reduced training and inference cost. Even our smallest model remains competitive using substantially fewer parameters and more than 17x faster inference.
Jannik Wiese, Johannes Schusterbauer, Tommaso Martorella +1