PaperlayerView exact revision on arXiv

Source-linked AI summary

Pushing the Limits of High-Resolution Weather Forecasting through Data Scaling

Yang Zhao, Peisong Niu, Tian Zhou, Ziqing Ma, Guanlong Ma, Rong Jin, Huiling Yuan, Liang Sun

arXiv:2608.14652v1cs.LGcs.AIcs.CV

TL;DR

High-resolution ML weather forecasting is limited by scarce long-term 0.1° data, while existing model transfer from coarse resolutions loses information. BaguanHR instead transfers data using variable-wise super-resolution to synthesize high-resolution training data, achieving superior performance across over 85% of lead times within 72 hours and a 4.0% RMSE reduction versus IFS-HRES.

  • Problem

    High-resolution ML weather forecasting lacks long, consistent global datasets because extensive reanalysis is available only at coarser resolution.

  • Method

    BaguanHR transfers data rather than models by using variable-wise super-resolution to generate extensive 0.1° training data from coarse-resolution reanalysis.

  • Results

    4.0% RMSE reduction compared specifically to IFS-HRES, with superior performance across over 85% of lead times within 72 hours.

  • Takeaways & Limitations

    The findings identify data availability as the primary bottleneck in high-resolution forecasting and support super-resolution data construction as a scalable route for improvement.

  • Takeaways & Limitations

    Model transfer from coarse-resolution forecasting remains limited because coarse representations cannot fully exploit fine-resolution information and adaptation requires learning fine-scale processes and multiscale dynamics.

Abstract

from arXiv · show

The development of 0.1$^{\circ}$ global weather forecasting models based on machine learning (ML) is constrained by the limited availability of high-resolution data, as decades of reanalysis are only available at 0.25$^{\circ}$ resolution. While existing approaches fine-tune 0.25$^{\circ}$ forecast models on limited 0.1$^{\circ}$ samples, we show that this transfer is hindered by the irreversible information loss inherent in coarse-resolution forecasting. Therefore, we propose BaguanHR, a framework that shifts the focus from transferring models to transferring data. We first show that super-resolution (SR) has lower conditional entropy and input amplification than forecasting, making it a more robust vehicle for resolution transfer. By leveraging this advantage through variable-wise SR, we synthesize extensive 0.1$^{\circ}$ data from ERA5. BaguanHR's performance on the synthetic-plus-real dataset exceeds both ML-based methods and IFS-HRES, achieving superior performance across over 85% of the lead times within 72 hours. Furthermore, our findings highlight a power-law scaling effect, as a twofold increase in data reduces RMSE by 4.6% for 72-hour forecasting and 4.9% for 120-hour forecasting. Our results demonstrate that scaling high resolution ML-based forecasting is primarily a data bottleneck, and that variable-wise super-resolution provides a simple yet general solution to unlock long coarse-resolution reanalyses for high-resolution training.

1 Introduction

High-resolution global weather forecasting is bottlenecked by scarce long-term 0.1° data, and transferring data through variable-wise super-resolution offers a more robust alternative to transferring coarse-resolution models. BaguanHR synthesizes training data from ERA5, improves forecasting performance, and exhibits power-law gains as data volume increases.

  • Data bottleneck: ECMWF provides global 0.1° data only after 2016, yielding at most about 10 years of training data up to 2026.This duration is described as far too limited for training large-scale ML forecasts.
  • Limitations of model transfer: Coarse-to-fine transfer is inherently limited because a pretrained 0.25° model cannot fully exploit the information in 0.1° analysis fields.The passage identifies architectural transfer and fine-tuning on small high-resolution datasets as the prevailing strategy.
  • Super-resolution mechanism: Super-resolution has lower conditional entropy than multi-step forecasting, making it easier to learn, more stable under imperfect inputs, and suitable for scalable data transfer.For z500, SR yields RMSE 13.8 versus 23.6 for one-step forecasting, while its lower amplification factor indicates greater stability.
  • BaguanHR framework: BaguanHR trains per-variable SR models on limited paired 0.1°–0.25° samples, synthesizes decades of 0.1° ERA5 data, and trains 0.1° forecast models from scratch.This synthetic-plus-real strategy outperforms coarse-to-fine baselines without specialized architectures or fine-tuning.
  • Forecasting performance: Over 85% of lead times within 72 hours show superior performance, with a 4.0% RMSE reduction compared specifically to IFS-HRES.BaguanHR outperforms various ML-based models and IFS-HRES.
  • Data scaling laws: Expanding the training dataset from 7 to 18 years reduces RMSE by 4.9% for 120-hour forecasts and 4.6% for 72-hour forecasts.The study characterizes this improvement as a power-law scaling effect with increasing data volume.

2 Related Work

Recent 0.1° forecasting methods transfer knowledge from pretrained coarse-resolution systems, whereas this work expands high-resolution data through super-resolution and investigates data scaling under scarce 0.1° analysis data.

  • Model-transfer methods: Aurora uses bilinear upscaling for patch alignment, while Fengwu-GHR uses identity-preserving mappings for high-resolution extrapolation.
  • Model-transfer methods: Existing transfer-learning methods treat high-resolution forecasting primarily as model transfer under limited 0.1° data.
  • Data-centric alternative: This work instead expands the effective 0.1° dataset via super-resolution, offering a data-centric route complementary to architecture-centric neural-operator methods.
  • Scaling laws: Whether performance gains persist when scaling data at 0.1° resolution remains open because high-resolution analysis data are scarce.

3 Rationale for Synthetic-to-Real Transfer

The section compares forecasting with super-resolution as pathways for transferring coarse-resolution information to 0.1° modeling. It finds that variable-wise SR is more robust to input noise and that synthetic-plus-real training substantially outperforms coarse-to-fine transfer, supporting data transfer through physically consistent synthetic data.

  • SR versus forecasting: The forecasting baseline autoregressively maps 0.1° atmospheric states six hours forward, whereas SR reconstructs 0.1° states from 0.25° inputs for spatial data augmentation.Forecasting learns ΦF: Xt → Xt+∆t with ∆t = 6 hours; SR learns ΦSR: Zt → X̂t.
  • SR versus forecasting: SR consistently has lower input-noise amplification than forecasting across t2m, u10, and z500, while maintaining lower RMSE at every tested noise intensity.Sensitivity is measured by (|RMSEσ − RMSE0|/σ).
  • Synthetic-plus-real dataset: The training set combines 0.1° ECMWF 4D-Var analysis data from 2017–2024 with 10 years of 0.1° pseudo-labels generated by applying per-variable SR to 0.25° ERA5 from 2007–2016.Accumulated variables unavailable in the analysis archive come from zero-hour IFS-HRES forecasts.
  • Synthetic-plus-real versus coarse-to-fine: The synthetic-plus-real paradigm significantly outperforms coarse-to-fine transfer across most variables, with RMSE reductions reaching 20%.The comparison contrasts training from scratch on augmented high-resolution data with adapting a pretrained 0.25° model to 0.1° resolution.
  • Theoretical motivation: A theoretical analysis shows that appropriately generated synthetic data can reduce generalization error and yield estimates closer to the oracle than models learned from clean data alone.The result provides theoretical support for using super-resolution models for robust data augmentation.

4 System Architecture and Training Strategies

BaguanHR combines per-variable super-resolution that synthesizes 0.1° pseudo-labels from 0.25° ERA5 with a high-resolution forecasting model. Its hierarchical embedding, replay-buffer strategies, and staged scalable training address memory, error accumulation, and data-throughput constraints.

  • Per-Variable Super-Resolution: BaguanHR uses a per-variable super-resolution generator to synthesize 0.1° pseudo-labels from 0.25° ERA5 data, alongside a high-resolution forecasting model.The SR generator addresses the scarcity of 0.1° EC analysis data.
  • Per-Variable Super-Resolution: Each atmospheric variable is independently downscaled with Swin2SR using paired 0.25°–0.1° fields and an MAE reconstruction objective.Variable-specific training is adopted to improve reconstruction fidelity.
  • High-Resolution Forecasting Model: A two-level hierarchical weather embedding partitions variables into 12 groups, distills group-specific queries, and fuses them into a global summary to reduce GPU memory overhead.The forecasting model also uses variable-specific tokenization, Swin Transformer blocks, and a linear projection head for spatial restoration.
  • Training Strategies: Lead-time-aware loss weighting preserves short-lead gradient influence, while stochastic buffer replacement maintains heterogeneous lead times and improves long-range stability.The strategies are designed to prevent long-lead errors from dominating optimization and to mitigate small-local-batch instability.
  • Scalable Training Framework: The three-stage pipeline uses pretraining, synthetic fine-tuning, and rollout training, with data partitioning and hybrid storage providing a 4× I/O speedup.Training uses 32 NVIDIA A800 GPUs; the full pipeline takes about 20 days and uses about 60 TB for 18 years of synthetic-plus-real data.

5 Experiments

Experiments evaluate BaguanHR using WeatherBench metrics against physics-based and ML-based baselines, demonstrating strong overall, extreme-event, and case-study performance. Additional analyses show data-scaling benefits, per-variable SR advantages, physical consistency, and improved fine-scale structure.

  • Evaluation Setup: Experiments use weighted RMSE and weighted ACC for forecasts initialized at 00/12 UTC during 2025.The evaluation follows standard WeatherBench methodology.
  • Scaling Laws: 4.6% RMSE reduction at 72 hours and 4.9% at 120 hours demonstrate stable power-law-like improvements as training data increases.Extrapolation beyond 18 years suggests marginal gains may gradually diminish, though this is a trend analysis rather than a definitive estimate.
  • Overall Performance: BaguanHR improves over 85% of lead times within 72 hours and reduces RMSE by 4.0% versus IFS-HRES.Average RMSE reductions reach 5.8% at 24 hours and 9.7% at 72 hours relative to IFS-HRES.
  • Extreme Events: BaguanHR exceeds competing methods by over 10% in most categorical moisture-event cases, by 15% short-term and 3% long-term for percentile extremes, and by over 30% versus IFS-HRES for threshold extremes.The evaluations use q850 ≥14 g/kg, local t2m 90th/10th percentiles, and z850 < 13,500 m2/s2, respectively.
  • SR Analysis: 90.7% RMSE improvement for sp and >20% for t2m and msl show per-variable SR benefits, while correlations remain RH ≥0.94, |Vg| ≥0.82, and ∆θ ≥0.93.BaguanHR also preserves higher short-wavelength spectral power, whereas Baguan+swin2sr systematically attenuates energy.
  • Case Studies: BaguanHR achieves comparable tropical-cyclone intensity with superior track accuracy and captures fine spatial detail in a Great Lakes cold-air outbreak.For cyclone CO-MAY, it identifies both landfall sites and outperforms IFS-HRES tracking; it also resolves lake–land temperature contrasts missed by Baguan+swin2sr.

6 Conclusion

BaguanHR is a 0.1° global weather forecasting model built by transferring super-resolution-generated synthetic data. Its results demonstrate power-law scaling with training-data volume and superior performance across nearly all variables, especially during the first 72 hours.

  • Model and framework: BaguanHR is a 0.1° global weather forecasting model based on a data-transfer framework using super-resolution-generated synthetic data.The framework transfers data rather than relying only on model adaptation.
  • Data scaling: Model performance improves according to a power law as training-data volume increases.The paper presents this as a fundamental data-scaling law for high-resolution forecasting.
  • Performance: Using synthetic 0.1° data, BaguanHR outperforms IFS-HRES across nearly all variables, especially in the first 72 hours.The supplied passage states that the model achieves superior performance across nearly all variables, with its strongest emphasis on the first 72 hours.

Supplementary Material · A Theoretical Justification of Synthetic Data Transfer · A.1 Introduction

The introduction frames synthetic-data training against model collapse and the prevailing view that synthetic data worsens generalization. It argues that this conclusion is not universally valid, motivating the theoretical justification developed in the note.

  • A.1 Introduction: Model collapse has been observed when language models are trained on iteratively generated data.The passage cites prior work reporting this phenomenon.
  • A.1 Introduction: Subsequent studies found that model collapse can be avoided when clean data remains included in every training iteration.This condition is presented as preventing collapse during iterative training.
  • A.1 Introduction: The prevailing consensus is that synthetic-data learning produces worse generalization error than learning from clean data.The passage attributes this consensus to multiple cited studies.
  • A.1 Introduction: The note challenges the claim that learning from synthetic data is always inferior to learning from clean data.Its stated purpose is to show that the consensus does not hold in every case.
  • A.1 Introduction: The introduction connects synthetic-data training quality to whether clean data is retained throughout iterative training.The passage contrasts collapse prevention with the broader generalization-error consensus.
  • A.1 Introduction: The section establishes the motivation for a theoretical analysis of synthetic data transfer.The note proceeds from reported observations and prior findings to a claim requiring theoretical justification.

A.2 Analysis · A.3 Proof of Theorem 3 · B Data and Evaluation Details

The analysis shows that unlabeled noisy examples incur an irreducible error tied to variation among Gaussian-specific parameters, whereas recovering their missing indices enables estimation approaching the oracle as N grows. The proof establishes this recovery when σ=o(1), under concentration and separation conditions.

  • A.2 Analysis: Inputs select one of m Gaussian distributions, combine the sampled vector with a one-hot index, and generate outputs through a Gaussian-noisy linear response.The training set Da contains n examples from this procedure.
  • A.2 Analysis: Unlabeled noisy examples omit the Gaussian index, so even infinitely many such samples cannot generally recover the Gaussian-specific regression parameters.The resulting lower bound is σ^2 VAR(w_1, ..., w_m).
  • A.2 Analysis: σ^2 VAR(w_1, ..., w_m) is the lower bound on expected prediction error from using the unlabeled-data regression model.This error arises from variation among the Gaussian-specific vectors w_1, ..., w_m.
  • A.2 Analysis: The index-augmentation approach estimates Gaussian centers, predicts each missing index, augments inputs with the predicted one-hot index, and learns a regression model from the resulting synthetic dataset.The construction is designed to use the additional N noisy examples while restoring the missing structural information.
  • A.2 Analysis: The recovered synthetic dataset yields an estimation error O(sqrt(md/N log(1/δ))), and is significantly closer to the oracle w than the model learned from Da.The bound is stated with probability at least 1 − δ.
  • A.3 Proof of Theorem 3: When σ=o(1), the estimated centers identify the correct Gaussian indices for all N noisy examples with probability 1 − δ.The proof uses concentration of Gaussian sample counts, center estimates, and sample norms, together with mean separation Ω(sqrt(d)).
  • A.3 Proof of Theorem 3: After correct index recovery, linear regression on the augmented dataset has error O(sqrt(md/N log(1/δ))).This follows by applying the same regression analysis to the correctly augmented examples.

B.1 Dataset … C.1 Problem Statement

BaguanHR combines 0.1° real-time analysis with ERA5-derived 0.1° synthetic data for high-resolution forecasting, using chronological splits and WeatherBench-style latitude-weighted RMSE and ACC evaluation. Its problem formulation represents the atmospheric state as a multivariable latitude–longitude grid at 0.1° resolution.

  • B.1 Dataset: The dataset combines 0.1° real-time analysis data with synthetic 0.1° data generated by super-resolution from ERA5.ERA5 spans 1940 to the present at 0.25° resolution and substantially exceeds real-time analysis in volume.
  • B.1 Dataset: The inputs include 13 pressure levels and atmospheric variables such as geopotential, temperature, wind components, specific humidity, and surface fields.Listed surface variables include T2M, D2M, U10, V10, and MSL.
  • Real-time Analysis Data: Real-time 0.1° analysis is produced every 6 hours through 4D-Var assimilation and provides the initial condition for IFS-HRES forecasts.It assimilates satellite retrievals, radiosondes, and surface station reports into a dynamically consistent atmospheric state.
  • Synthetic Data by Super-resolution: The super-resolution model converts 0.25° ERA5 fields into synthetic 0.1° data matching the real-time analysis resolution.This variable-wise conversion is used to exploit ERA5’s larger data volume.
  • Data Partitioning Strategy: Super-resolution training uses paired ERA5 and 0.1° analysis data from 2018–2024, while synthetic data generation covers ERA5 from 2007–2016.The forecasting model combines synthetic data from 2007–2016 with real-time analysis data from 2017–2024; 2025 is reserved for evaluation.
  • B.2 Evaluation Metrics: Evaluation follows WeatherBench using 00/12 UTC forecasts in 2025 and latitude-weighted RMSE and ACC.Both metrics are widely used for evaluating ML-based weather models.
  • Root mean square error (RMSE): RMSE evaluates prediction precision across latitudes using a latitude weighting factor that accounts for differing surface areas on Earth.The weighting factor is based on the cosine of latitude normalized across grid latitudes.
  • Anomaly correlation coefficient (ACC): ACC measures spatial correlation between forecast and ground-truth anomalies relative to historical day-of-year climatology at each forecast time.The time-specific anomaly correlations are averaged across valid test times.

C.2 Super-Resolution Data Generator (Swin2SR) … D.1 Hybrid I/O and Caching Strategies

BaguanHR combines topography-conditioned super-resolution, a residual forecasting architecture, staged training, and self-adaptive rollout optimization to generate and forecast at 0.1° resolution. Its system design addresses the resulting storage bottleneck through hybrid I/O and caching, achieving over 4× training speedup.

  • C.2 Super-Resolution Data Generator (Swin2SR): Swin2SR fuses 0.25° atmospheric variables with 0.1° static topography and reconstructs fine-scale structure through residual learning.A dual-branch stem aligns and combines atmospheric and geographic inputs, while the model predicts a high-frequency residual added to bicubic interpolation.
  • C.3 The BaguanHR Forecasting Model: BaguanHR encodes 86 variables into 12 physical groups, aggregates them into a latent representation, and evolves atmospheric dynamics with a Swin Transformer backbone.Variable-specific tokenization preserves physical identity, group-wise and global attention produce a latent token of dimension C = 1536, and the backbone contains 24 Swin Transformer blocks.
  • C.3 The BaguanHR Forecasting Model: The forecasting decoder predicts dynamical increments rather than absolute states and adds them to current fields through a global residual connection.Accumulated variables such as total precipitation are generated directly.
  • C.4 Experimental Setups: Training uses latitude-weighted MAE across a four-stage pipeline spanning super-resolution synthesis, analysis-data pretraining, synthetic fine-tuning, and rollout training.The models are trained on 32 NVIDIA A800 GPUs, with synthetic data generated from 2007–2016 and combined with real 2017–2024 analysis data in later stages.
  • C.5 Self-Adaptive Rollout and Replay Buffer Strategy: The self-adaptive rollout strategy weights the k-th autoregressive loss by 1/k, emphasizing short-range accuracy while retaining long-horizon supervision.This addresses the tendency of long-range errors to dominate optimization during multi-step fine-tuning.
  • C.5 Self-Adaptive Rollout and Replay Buffer Strategy: Replay-buffer-based rollout training mitigates narrow lead-time coverage and optimization instability under very small per-GPU batch sizes.The high-resolution model often uses N = 1 per GPU, making FIFO buffering prone to concentrating the distributed batch on limited lead times.
  • D Large-Scale System Engineering and Computational Cost: High-resolution training creates a severe storage-to-CPU I/O bottleneck because 0.1° data require 32-bit storage and individual files typically occupy 4 to 5 GB.A dense 1801 × 3600 grid makes storage bandwidth, rather than GPU FLOPs, the limiting factor.
  • D.1 Hybrid I/O and Caching Strategies: The pipeline masks I/O latency through three customized data-loading strategies and achieves a training speedup of over 4× versus naive network storage access.The design combines machine-level dataset sharding, time-based hybrid storage, and in-memory replay of cached batches.

D.2 System-Level Profiling … Experiment II: Input Perturbation Sensitivity

The section profiles BaguanHR’s system bottleneck and experimentally compares super-resolution with forecasting, finding that SR is simpler, more accurate, and substantially less sensitive to input perturbations. These results support using SR for robust resolution transfer.

  • D.2 System-Level Profiling: 364.7 TFLOPs per optimization step per GPU coexist with 28.7 seconds of I/O wait and a 54.0% GPU active ratio under NAS loading.Profiling used 200 optimization steps on a single operational 8 × A800 GPU node to isolate storage effects from cross-node communication.
  • D.2 System-Level Profiling: For 0.1° training, the dominant limitation is storage and data delivery efficiency rather than raw computation.The benchmark isolates the storage-to-CPU bottleneck from inter-node network communication overhead.
  • Experiment I: Task Difficulty Comparison: The protocol compares same-time low-resolution-to-high-resolution reconstruction against high-resolution prediction of the future 6-hour state.SR uses X_0.25°_t to reconstruct X_0.1°_t, whereas forecasting predicts X_0.1°_t+6h from the high-resolution state.
  • Experiment I: Task Difficulty Comparison: 13.8 vs. 23.6 for z500: SR yields significantly lower RMSE than forecasting, indicating lower conditional entropy for same-time spatial reconstruction.The comparison motivates treating SR as an intrinsically simpler task than temporal forecasting.
  • Experiment II: Input Perturbation Sensitivity: Input perturbation sensitivity is measured by adding Gaussian noise at σ ∈ {0.01, 0.05, 0.10, 0.20, 0.25} and dividing output-error change by injected-noise magnitude.The same Amplification Factor definition is applied to both SR and forecasting models.
  • Experiment II: Input Perturbation Sensitivity: 0.6 ∼0.8×: Amplification_SR remains below 1.0, while Amplification_FC rises above 1.0 as forecasting magnifies small input perturbations.The contrast is attributed to minimal, reversible information transfer in SR versus irreversible magnification from chaotic temporal evolution.

E.2 Mathematical Formulation of the Data Scaling Law … E.7 The Performance of BaguanHR in Tracking Tropical Cyclones

The appendix formalizes data scaling, examines SR physical consistency, clarifies forecasting baselines and evaluation protocols, and reports BaguanHR’s performance in station verification, qualitative forecasts, and tropical-cyclone tracking.

  • E.2 Mathematical Formulation of the Data Scaling Law: E(D) = αD^-β + E∞ models forecast error as a power law of training-data volume with an irreducible saturation bound.Here, D is training data volume, β > 0 is the scaling decay exponent, and E∞ is the theoretical saturation bound.
  • E.2 Mathematical Formulation of the Data Scaling Law: 18-year curves flattening indicate a saturation threshold, while additional SR data improve representation of rare meso-scale anomalies before atmospheric chaos and model capacity impose limits.The scaling models were trained only for 6-hour single-step prediction; R2 > 0.95 across all variables isolates the scaling mechanism from rollout error accumulation.
  • E.3 Extended Discussion on Physical Consistency of SR-Generated Fields: SR-generated fields preserve moisture, dynamical, and thermodynamic diagnostic relationships despite independent per-variable generation.The diagnostics use relative humidity, geostrophic wind speed, and potential-temperature-based stratification.
  • E.4 Extended Discussion on Baseline Selection and Forecasting Dynamics: Zero-shot +swin2sr baselines freeze 0.25° forecasting models and upscale their outputs to 0.1°, exposing information loss in decoupled post-processing.At short lead times, coarse models miss high-frequency structures and sharp mesoscale gradients that post-processing cannot fully reconstruct.
  • E.4 Extended Discussion on Baseline Selection and Forecasting Dynamics: BaguanHR’s advantage over post-processing baselines narrows by 10 days as deterministic autoregressive models accumulate error, lose variance, and produce smoother states.Baguan+GHR is included as a coarse-to-fine fine-tuned baseline for fair comparison within the critical 3-day operational window.
  • E.5 Comparison Against Station Observation: 13,058 global weather stations provide direct grid-to-point verification using hourly temperature and wind-speed observations from January 1, 2025 to December 20, 2025.Model outputs are bilinearly interpolated to station coordinates before forecast errors are computed.
  • E.6 Visualization of High-Resolution Forecasts: BaguanHR generally appears closer to 0.1° analysis and has lower overall error than Baguan+swin2sr in visual cases spanning T2M, U10, and Q850.The visualizations also compare IFS-HRES and include predicted-minus-analysis error maps.
  • E.7 The Performance of BaguanHR in Tracking Tropical Cyclones: Mean tracking errors across multiple initialization times show consistently strong BaguanHR tropical-cyclone performance for individual storms and basin means.Figure 14 covers six representative storms and the Western North Pacific and North Atlantic cyclogenesis regions.

F Limitations and Broader Impact

BaguanHR offers a data-centric route to broaden access to high-resolution forecasting, but its independently generated variable-wise super-resolution data may weaken cross-variable physical consistency. The framework also cannot resolve microscale convective extremes, and forecast errors accumulate with lead time, limiting gains at 10–15 days.

  • Broader impact: Public gridded meteorological datasets avoid human subjects, personal data, and sensitive individual information while enabling synthetic 0.1° training data from long-term coarse-resolution reanalyses.This provides a practical route for improving high-resolution forecasting under limited compute and data availability.
  • Limitations: Independently trained per-variable super-resolution models may not fully preserve cross-variable coupling and multivariate physical consistency at the finest scales.The design improves variable-specific fidelity but introduces a limitation in joint physical consistency.
  • Limitations: 0.1° resolution, roughly 9 km, remains insufficient to explicitly resolve microscale convective extremes such as tornadoes, hail, and highly localized severe storms.The limitation concerns phenomena smaller than the model’s effective resolution.
  • Limitations: Forecast errors continue accumulating during auto-regressive rollout, while BaguanHR provides limited additional gains at 10–15 day lead times.Long-lead forecasting therefore remains a challenge despite the framework’s high-resolution data scaling.
Loading 2608.14652v1…