Diffusion Precipitation Generation with Medium-range Weather Forecasting

The gist Accurate daily precipitation forecasts are important for weather-sensitive decisions, and this study introduces Dynamics–Precipitation Decoupling, a plug-and-play framework that uses a separately trained generator to map non-precipitation forecasts from a frozen weather model to observation-aligned daily precipitation <ref:2610.12976#pg6>.

Motivation and Problem Definition

Accurate daily precipitation forecasting is a core task in weather prediction, playing a pivotal role in disaster preparedness, agricultural production, transportation management, and water-resource planning <ref:2610.12976#pg5>. Traditional approaches mainly rely on Numerical Weather Prediction (NWP), which solves complex atmospheric dynamical and thermodynamic equations to simulate atmospheric evolution, or statistical methods that mine relationships between historical precipitation and meteorological factors <ref:2610.12976#pg4>. Recent development of Weather Foundation Models (WFMs) trained on ERA5 has enabled the exploration of global medium-range precipitation forecasting <ref:2610.12976#pg4>. However, previous evaluations have documented regional and intensity-dependent biases in ERA5 precipitation, leading to systematic over- or underestimation and distorted spatiotemporal distributions <ref:2610.12976#pg4>.

The study addresses two key challenges in Dynamics–Precipitation Decoupling:

**: Daily precipitation exhibits heterogeneous spatial patterns and skewed accumulation distributions, which deterministic models trained with pointwise regression can favor conditional averages, smoothing local accumulation maxima and rainband structures <ref:2610.12976#pg3>. 24-hour Evolution is captured through a multi-scale recurrent encoder summarizing atmospheric evolution, followed by cross-field alignment to guide generation <ref:2610.12976#pg4>. **

**: Atmospheric conditioning poses a nonlocal spatiotemporal alignment challenge, as precipitation reflects the cumulative outcome of evolving atmospheric processes that may occur earlier and over surrounding regions <ref:2610.12976#pg4>. This necessitates a solution where the generator learns from paired ERA5 and IMERG precipitation, using only non-precipitation states as conditioning inputs <ref:2610.12976#pg5>. **

How it works

The framework is structured into two main steps:

1. Obtain future atmospheric forecasts: Forecasts of non-precipitation atmospheric variables are obtained from a base weather model, which can be either an NWP system or a WFM, and the base model remains frozen <ref:2610.12976#pg5>. The initial global weather state is represented by Xt ∈ R H×W×V <ref:2610.12976#pg4>.

2. Generate precipitation fields via conditioned diffusion: This step employs a conditional diffusion model to generate the 24-hour accumulated precipitation field conditioned on (Xˆt−12, Xˆt) <ref:2610.12976#pg5>. The target distribution is formulated as pθ(Y 0:K tXˆt−12, Xˆt), where Y K t ∼ N(0, I) <ref:2610.12976#pg4>.

The generation process involves a multi-scale atmospheric encoder and a cross-attention mechanism in the denoiser:

**: The Atmospheric Encoder transforms atmospheric forecasts (Xˆt−12, Xˆt) into an intermediate latent representation τ tθ = τθ(Xˆt−12, Xˆt) <ref:2610.12976#pg5>. This encoder consists of ResNet blocks and a Convolutional GRU-based gating mechanism to model the temporal evolution of atmospheric states <ref:2610.12976#pg4>. **

**: The Denoiser augments the U-Net backbone with a cross-attention mechanism to maintain spatial consistency during denoising, allowing precipitation features to selectively attend to atmospheric conditions at the matching spatial scale <ref:2610.12976#pg5>. The prediction function is expressed as ϵˆθ = Denoiser(Y k t, τ tθ, k) = fout Cross-Attn(Y k t, τ tθ) <ref:2610.12976#pg4>. **

Key Contributions and Evaluation

The main contributions of the work are summarized as follows:

**: We highlight Dynamics–Precipitation Decoupling, separating atmospheric evolution from observation-supervised precipitation generation and enabling IMERG-aligned learning without retraining a complete weather forecasting system <ref:2610.12976#pg5>. **

**: We develop a forecast-conditioned diffusion architecture that combines a multi-scale atmospheric encoder with an attention-conditioned denoiser to model 24-hour precipitation distributions <ref:2610.12976#pg4>. **

**: We reuse one ERA5-trained generator with aligned forecasts from frozen Pangu-Weather, GraphCast, and HRES, outperforming available native precipitation forecasts without training a full weather model from scratch <ref:2610.12976#pg5>. **

Extensive experiments cover three complementary levels: global forecasting at 1.5◦, regional forecasting at 0.25◦, and a heavy-rainfall case study using observations from 2,288 stations in China <ref:2610.12976#pg4>. Global ensemble evaluation shows lower Brier scores than ENS and training-from-scratch GenCast at all reported lead times, with lower CRPS in most comparisons <ref:2610.12976#pg4>. A heavyrainfall case study using 2,288 stations in China further supports improved daily accumulation patterns and event detection against ground observations <ref:2610.12976#pg4>.

Performance Results

The framework demonstrates improved deterministic skill and competitive probabilistic performance across various backbones <ref:2610.12976#pg4>. For instance, GraphCast+Ours raises its ACC by 0.103 while reducing its RMSE by 0.623 on average <ref:2610.12976#pg4>. In the probabilistic forecast comparison, GraphCast+Ours performs best at every reported lead time, with its CRPS at Day 7 being 2.180, compared with 2.484 for ENS and 2.368 for GenCast <ref:2610.12976#pg4>. Furthermore, station observations support improved local rainfall intensity patterns and event detection during the evaluated heavy-rainfall event <ref:2610.12976#pg4>.

Conclusion

We present Dynamics–Precipitation Decoupling for daily precipitation generation at medium-range lead times <ref:2610.12976#pg4>. The framework samples precipitation conditional on a fixed atmospheric trajectory, and its skill remains dependent on the backbone forecast <ref:2610.12976#pg4>. This approach demonstrates the potential of decoupled precipitation generation as a reusable component of medium-range forecasting systems <ref:2610.12976#pg4>. The results demonstrate the potential of decoupled precipitation generation as a reusable component of medium-range forecasting systems <ref:2610.12976#pg4>.

Limitations

Cross-backbone reuse is empirically validated only for the aligned forecast systems considered in this work, while the 1.5◦ and 0.25◦ settings require separately trained models <ref:2610.12976#pg5>. Training on ERA5 states also introduces a distribution shift when conditioning on imperfect forecasts, which is probed through noise perturbation and forecast-field fine-tuning experiments <ref:2610.12976#pg5>. Future work should further evaluate transfer across other reanalyses, such as JRA and MERRA-2, and operational products such as GFS <ref:2610.12976#pg5>. IMERG retrieval uncertainty and the additional cost of diffusion sampling remain further limitations <ref:2610.12976#pg4>.

AI Use Statement

In this work, we used LLMs as an auxiliary tool to polish the writing and improve the manuscript’s clarity <ref:2610.12976#pg4>. Specifically, we use it to review the entire text for grammatical accuracy, improve sentence structure, and ensure consistent phrasing and tone throughout the paper <ref:2610.12976#pg4>.

Ablation Study Findings

The temporal-density ablation shows that denser conditioning reduces RMSE by only 0.002 (0.03%), while the two-state input reduces inference overhead from 0.84 s to 0.33 s under the evaluated configuration <ref:2610.12976#pg5>. The diagnostic fine-tuning variant increases RMSE at every evaluated lead day for both backbones, suggesting that fitting the generator to backbone-specific forecast errors may weaken the learned relationship <ref:2610.12976#pg5>. Direct reuse with GraphCast achieves a substantially lower mean RMSE than the original GraphCast forecast (6.522 versus 7.143) <ref:2610.12976#pg4>.

Computational Overhead

The proposed generator introduces 110.

1. Bold header: Direct cross-backbone reuse for generator training

The framework does not use the backbone’s precipitation outputs and One generator per resolution is reused across aligned backbones without backbone-specific retraining, allowing a single diffusion generator to be applied to frozen models like GraphCast or Pangu-Weather, which avoids the prohibitive cost of training a complete weather model from scratch.

2. Bold header: Enhanced atmospheric feature encoding

The multi-scale recurrent encoder captures atmospheric evolution in compact, daily precipitation features by using a Convolutional GRU cell as the 'convolutional operator' to model temporal evolution between two forecast states, which is shown to consistently outperform ConvLSTM under both conditioning backbones.

3. Bold header: Cross-field feature alignment during denoising

The system implements a cross-attention mechanism where precipitation features selectively attend to atmospheric conditions at the matching spatial scale, allowing the model to perform a constrained synthesis guided by the evolution of atmospheric states and bridge the gap between atmospheric and precipitation features.

4. Bold header: Improved localized accumulation pattern representation

The framework is shown to preserve its organization and recover sharper, heterogeneous daily rainfall patterns over Southern Africa and northern Australia compared to existing models which oversmooth rainfall, as evidenced by better matching with IMERG reference fields in regional evaluations at 0.25◦.

5. Bold header: Robustness against input perturbations

The model demonstrates controlled sensitivity to synthetic input errors because performance changes are limited for small perturbations and only degrade gradually as the standard deviation of Gaussian noise increases, indicating a stable mapping from atmospheric states to precipitation fields.

Accurate daily precipitation forecasts are important for weather-sensitive decisions. Although recent Weather Foundation Models (WFMs) have achieved significant progress, they are trained on reanalysis data, which exhibit regional and intensity-dependent biases. Since retraining WFMs on observation-based targets is costly, we study Dynamics-Precipitation Decoupling, a plug-and-play framework that uses a separately trained generator to map non-precipitation forecasts from a frozen weather model to observation-aligned daily precipitation. Its key challenges are to capture heterogeneous rainfall distributions and align daily accumulations with atmospheric signals that evolve across space and time. To address them, we train a diffusion generator conditional on atmospheric states and satellite-gauge estimates of precipitation. A multi-scale recurrent encoder summarizes atmospheric evolution, while cross-attention aligns atmospheric and precipitation features during denoising. Extensive evaluation of various backbones, including Pangu-Weather, GraphCast, and HRES, on global and regional forecasting shows improved daily precipitation accuracy and event detection, with better representation of localized accumulation patterns. Global ensemble evaluation shows lower Brier scores than ENS and training-from-scratch GenCast at all reported lead times, with lower CRPS in most comparisons. Remarkably, a heavy-rainfall case study using 2,288 stations in China further supports improved daily accumulation patterns and event detection against ground observations.