PaperlayerView exact revision on arXiv

Source-linked AI summary

Diffusion Precipitation Generation with Medium-range Weather Forecasting

Guowen Li, Yang Liu, Mengxuan Chen, Jiashun Cheng, Haoyuan Liang, Shilei Cao, Yujie Wang, Juepeng Zheng, Fang Wang, Yerong Feng, Meng Jin, Hong Cheng, Yutong Lu, Haohuan Fu

arXiv:2610.12976v1cs.LG

TL;DR

Daily precipitation forecasting matters, but precipitation reflects atmospheric processes evolving across locations and times. The paper decouples precipitation generation from a frozen weather model, using a conditional diffusion generator; evaluations report improved deterministic skill and support for a reusable forecasting component.

  • Problem

    Daily precipitation reflects evolving atmospheric processes, so conditions at the same grid point and time alone cannot explain it.

    Source: §1 · p. 2

  • Method

    A separately trained conditional diffusion generator maps non-precipitation forecasts from a frozen weather model to observation-aligned daily precipitation.

    Source: §1 · p. 2

  • Results

    Across global and regional evaluations, the generator improves deterministic skill, and it outperforms available native precipitation forecasts across frozen Pangu-Weather, GraphCast, and HRES backbones.

    Source: §1 · p. 3, §6 · p. 10

  • Takeaways & Limitations

    Decoupled precipitation generation has potential as a reusable component of medium-range forecasting systems without retraining a complete weather forecasting system.

    Source: §1 · p. 3, §6 · p. 11

    Further evaluation across forecasting systems and heavy-rainfall events remains important.

    Source: §6 · p. 11

Abstract

from arXiv · show

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.

1 INTRODUCTION

The paper separates atmospheric forecasting from daily precipitation generation to address biased reanalysis precipitation targets and the cost of retraining weather models. Its conditional diffusion generator combines temporal atmospheric encoding with cross-field alignment and is evaluated across forecast backbones and station observations.

1 INTRODUCTION

  • Reanalysis precipitation has documented regional and intensity-dependent biases, while replacing the target and retraining weather models is described as prohibitively costly.The paper uses the satellite-gauge IMERG product for a more direct observation-aligned target.Source: §1 · pp. 1–2
  • Dynamics–Precipitation Decoupling maps non-precipitation forecasts from a frozen backbone to observation-aligned daily precipitation using a separately trained generator.The generator learns from ERA5–IMERG pairs without backbone-specific retraining.Source: §1 · pp. 2–3
  • Daily rainfall’s heterogeneous patterns and skewed accumulations challenge pointwise regression, while precipitation can reflect earlier atmospheric signals over surrounding or upstream regions.The paper notes that same-grid, same-time atmospheric conditions alone cannot explain precipitation at a location.Source: §1 · p. 2
  • The framework combines a multi-scale recurrent encoder of atmospheric evolution with cross-attention that aligns atmospheric and precipitation features during denoising.The conditioning is intended to address heterogeneous rainfall distributions and nonlocal, evolving atmospheric signals.Source: §1 · pp. 2–3
  • The ERA5-trained generator is reused with frozen Pangu-Weather, GraphCast, and HRES forecasts, outperforming available native precipitation forecasts without training a full weather model from scratch.The cited evaluations span global and regional forecasts, while the 2,288-station China case supports improved local rainfall intensity patterns and event detection.Source: §1 · p. 3

2 RELATED WORK

Prior precipitation forecasting approaches differ in domain, lead time, and whether they forecast, correct, or refine precipitation fields. This paper instead conditions an independently supervised generator on non-precipitation forecast states from an existing weather model.

Precipitation forecasting with machine learning

  • Radar-based nowcasting captures local precipitation evolution over regional domains but generally targets only a few hours, whereas WFMs extend data-driven forecasting to global, medium-range lead times.The passage notes that precipitation products from WFMs remain tied to system-specific supervision and can have ERA5-related biases against satellite–gauge observations.Source: §2 · p. 3

Diffusion models for weather forecasting

  • Diffusion weather studies variously forecast full atmospheric states, emulate regional systems, or extrapolate precipitation history; this framework instead retains an existing backbone and generates precipitation from its non-precipitation states.The generator is independently supervised with IMERG.Source: §2 · p. 3

Precipitation post-processing and downscaling

  • Unlike precipitation post-processing and downscaling, the framework generates observation-aligned precipitation directly from non-precipitation atmospheric states, without requiring a precipitation forecast as input.The paper distinguishes this cross-variable generation task from correcting an existing precipitation field or spatial super-resolution.Source: §2 · p. 3

3 BACKGROUND

The task is to forecast 24-hour accumulated precipitation at future horizons from initial global non-precipitation atmospheric conditions. The background describes a diffusion process that learns to reverse noise corruption of precipitation fields.

3 BACKGROUND

  • The forecasting setup represents global non-precipitation atmospheric conditions on a latitude-longitude grid and seeks 24-hour accumulated precipitation at future horizons.The grid is defined over the latitude and longitude ranges, with atmospheric variables comprising upper-air and surface quantities.Source: §3 · p. 4
  • Predictions cover preceding 24-hour windows at 24-hour intervals, with the study considering forecasts up to seven days ahead.The horizons are t = 24, 48, …, T, where T = 168 hours.Source: §3 · p. 4
  • The diffusion model learns the precipitation distribution by reversing a fixed-length Markov chain, progressively adding Gaussian noise in the forward process and denoising in reverse.The denoising model is optimized to predict noise in a degraded precipitation map.Source: §3 · p. 4

4 DIFFUSION PRECIPITATION GENERATION

The framework decouples atmospheric forecasting from precipitation generation, reusing frozen weather-model forecasts to produce observation-aligned daily precipitation. Conditional diffusion uses temporally encoded atmospheric states and cross-attention to guide precipitation denoising.

4.1 DYNAMICS–PRECIPITATION DECOUPLING

  • Frozen weather-model forecasts of non-precipitation variables condition a separately trained generator, avoiding end-to-end retraining for a new precipitation target.Training uses ERA5 states and prediction uses forecast states; direct reuse of the ERA5-trained generator performs better than forecast-field fine-tuning.Source: §4.1 · p. 5
  • Conditional diffusion generates 24-hour accumulated precipitation from two atmospheric states 12 hours apart, denoising from Gaussian noise under their atmospheric conditions.The two states compactly describe intraday atmospheric change, and denoising is guided by atmospheric-state evolution.Source: §4.1 · p. 5

4.2 PRECIPITATION GENERATION ARCHITECTURE

  • A convolutional atmospheric encoder compresses multilevel forcing into latent features to reduce computation while preserving precipitation-relevant signals.Its hierarchical blocks extract features across downsampling stages, and ConvGRU gating models the temporal evolution of the two atmospheric states.Source: §4.2 · pp. 5–6
  • Cross-attention lets precipitation features selectively attend to atmospheric conditions at matching spatial scales while the denoiser predicts precipitation-field noise.This design targets spatial alignment between atmospheric conditions and daily precipitation features.Source: §4.2 · p. 6

5 EXPERIMENTS

Experiments evaluate the framework on global and regional forecasts, probabilistic ensembles, and a North China station-based heavy-rainfall case. Results show improved forecast skill and localized precipitation patterns, with the full design performing best across reported metrics.

5.1 EXPERIMENTAL SETUP

  • The dataset pairs hourly ERA5 states with IMERG precipitation, uses 1998–2021 for training, 2023 for validation, and 2022 for testing, and reserves 2,288 China stations for external evaluation.HRES and ENS come from WeatherBench2; GraphCast and Pangu-Weather forecasts use their released models.Source: §5.1 · pp. 6–7

5.2 GLOBAL ASSESSMENT AT 1.5◦

  • For precipitation forecasting, GraphCast+Ours raises GraphCast’s ACC by 0.103 and reduces its RMSE by 0.623 on average, outperforming GenCast; HRES+Ours performs comparably and beats Pangu-Weather+Ours.Source: §5.2 · p. 8
  • At Day 7, GraphCast+Ours records CRPS 2.180 vs ENS 2.484 and GenCast 2.368, and BS 0.218 vs ENS 0.245 and GenCast 0.247; all variants beat both on BS at every lead time.GraphCast+Ours performs best at every reported lead time for CRPS.Source: Table 1 · p. 7, §5.2 · p. 8
  • Day-3 forecasts show sharper local accumulation maxima; in Boxes A and C, patterns and daily amounts better match IMERG, while Box B has clearer rainband boundaries.All systems capture large-scale patterns, while baseline forecasts are smoother locally.Source: §5.2 · p. 8, Figure 4 · p. 8

5.3 REGIONAL EVALUATION AT 0.25◦

  • Across the United States, Southern Africa, and Australia at 0.25°, GraphCast+Ours achieves the best overall performance and better represents localized daily accumulation patterns.The framework recovers sharper, heterogeneous rainfall patterns over Southern Africa and northern Australia.Source: §5.3 · p. 9, Figure 5 · p. 9
  • At 2,288 North China stations during the early-July 2022 heavy-rainfall event, GraphCast+Ours preserves rainband placement and recovers accumulation maxima without domain-wide intensification; multi-threshold CSI gains support improved event detection.HRES and GraphCast capture the rainband but smooth contrasts in daily totals.Source: §5.3 · p. 9, Figure 6 · p. 9

5.4 MECHANISM AND ROBUSTNESS ANALYSIS

  • The full configuration performs best across reported metrics: ConvGRU beats ConvLSTM, and cross-attention beats gated fusion among their respective configurations.U-Net is competitive for low-intensity event detection but less accurate for the overall intensity distribution; CorrDiff's lower scores are consistent with an objective mismatch in this same-resolution setting.Source: §5.4 · p. 10, Table 2 · p. 10
  • Increasing sampling from 10 to 100 steps improves RMSE and ACC across all three backbones, with curves largely saturating beyond 100; small Gaussian input perturbations cause limited changes.The authors use 100 steps by default as a quality-computation trade-off; performance degrades gradually as perturbation magnitude increases.Source: Figure 7 · p. 9, §5.4 · p. 10

6 CONCLUSION

The paper presents Dynamics–Precipitation Decoupling for generating daily precipitation at medium-range lead times. Its evaluations report improved deterministic skill and support from a heavy-rainfall station case study, while identifying further evaluation as important.

6 CONCLUSION

  • The framework maps non-precipitation atmospheric states to daily accumulations using a diffusion generator trained on ERA5–IMERG pairs.It uses recurrent encoding and cross-field alignment, does not use backbone precipitation outputs, and reuses one generator per resolution across aligned backbones.Source: §6 · p. 10
  • Global and regional evaluations show improved deterministic skill, with competitive probabilistic performance in the global evaluation.A case study across 2,288 stations further supports improved daily accumulation patterns and event detection.Source: §6 · p. 10
  • Further evaluation across forecasting systems and heavy-rainfall events remains important.The results demonstrate the potential of decoupled precipitation generation as a reusable component of medium-range forecasting systems.Source: §6 · p. 11

A.1 BASELINES •

The baselines span AI weather models and operational numerical and ensemble systems. They provide deterministic, probabilistic, and physical-forecast comparisons for the paper’s evaluations.

A.1 BASELINES •

  • Pangu-Weather combines four 3D Earth-specific Transformer models for 1-, 3-, 6-, and 24-hour forecasts through hierarchical temporal aggregation.The four models have approximately 64 million parameters each, or 256 million in total.Source: §A.1 · p. 15
  • GraphCast is a Google DeepMind mid-range forecasting model based on Graph Neural Networks.Its Grid-to-Mesh-to-Grid architecture maps atmospheric variables onto a multi-scale icosahedral mesh.Source: §A.1 · p. 15
  • GenCast is a global deep-learning model that generates probabilistic forecasts and is described as surpassing traditional ensemble systems.The passage also describes high-resolution, physically consistent precipitation distributions and extreme-weather warnings.Source: §A.1 · p. 15
  • HRES is ECMWF’s numerical weather prediction system, using fluid dynamics and physical parameterizations alongside global observations.The passage presents it as a benchmark for validating physical reliability.Source: §A.1 · p. 15
  • ECMWF’s ENS generates multiple forecast members through perturbations to initial conditions or model parameters to quantify atmospheric uncertainty.Its ensemble mean or probabilistic distributions serve as operational precipitation-forecast benchmarks.Source: §A.1 · p. 15

A.2 METRICS

The appendix defines metrics for forecast error, anomaly agreement, event detection, and probabilistic performance. It also specifies precipitation thresholds and evaluation conventions for selected scores.

A.2 METRICS

  • RMSE measures the average size of forecast errors, while ACC measures the correlation between predicted and observed anomalies.Higher ACC indicates better forecasting performance.Source: §A.2 · pp. 15–16
  • CSI measures overlap between observed and predicted precipitation events, while HSS measures binary-forecast accuracy relative to random chance.Global CSI pools contingency counts over test cases and grid cells using latitude weights; HSS accounts for hits, misses, false alarms, and correct negatives.Source: §A.2 · pp. 16–17
  • CRPS is estimated for ensembles with 50 members using the traditional estimator under matched ensemble size.Gridpoint CRPS values are averaged over evaluation cases and the spatial domain using latitude weights.Source: §A.2 · p. 17
  • The Brier Score averages squared differences between predicted event probabilities and binary observations.For BS, the precipitation threshold is 0.1 mm; the score ranges from 0 to 1, with 0 indicating perfect event-probability accuracy, and is used as an aggregate skill score rather than decomposed into calibration or resolution.Source: §A.2 · pp. 17–18

A.3 ADDITIONAL RESULTS

Additional results quantify computational overhead, compare temporal-conditioning choices, and test direct cross-backbone reuse against fine-tuning. The appendix also provides further global and station-based visual comparisons.

A.3 ADDITIONAL RESULTS

  • The generator adds 110.3M parameters, 0.33 s per generated sample, and approximately 8 GB of peak GPU memory under the reported configuration.The overhead mainly arises from iterative diffusion sampling; fewer sampling steps are an option when lower latency is preferred.Source: §A.3 · p. 18
  • Two-state conditioning achieves a mean RMSE of 6.520 versus 6.518 with four states, while reducing inference overhead from 0.84 s to 0.33 s.Both configurations outperform native GraphCast at 7.143 mean RMSE; the denser input reduces RMSE by only 0.002 (0.03%).Source: §A.3 · p. 18, Figure 10 · p. 19
  • Direct reuse of the ERA5-trained generator achieves lower mean RMSE than fine-tuning for both GraphCast and Pangu-Weather.Fine-tuning raises mean RMSE from 6.522 to 6.947 for GraphCast and from 6.825 to 7.088 for Pangu-Weather; direct reuse with GraphCast reaches 6.522 versus 7.143 for native GraphCast.Source: §A.3 · p. 19
  • Additional visual comparisons show global precipitation fields at 1-, 3-, 5-, and 7-day lead times against IMERG.The global forecasts are initialized at 00:00 UTC on July 6, 2022; station-based comparisons cover three valid times for forecasts initialized June 29, 2022.Source: §A.3 · p. 19, Figure 11 · p. 20, Figure 12 · p. 21
Loading 2610.12976v1…