Guided Diffusion Sampling for Precipitation Forecast Interventions

Abstract: Extreme precipitation causes severe societal and economic damage, and weather control has long been discussed as a potential mitigation strategy. However, to the best of our knowledge, perturbation-based interventions for weather control using data-driven weather forecasting models have not yet been explored. While adversarial attacks also generate perturbations that alter forecasts, they aim to exploit model artifacts and do not account for physical plausibility. In this paper, we propose a gradient-based guidance framework for precipitation-reduction interventions through diffusion sampling in diffusion-based weather forecasting models. Instead of directly perturbing atmospheric states, our method steers the diffusion sampling trajectory, enabling precipitation reduction while maintaining consistency with the atmospheric distribution. To assess physical plausibility, we evaluate from three perspectives: (i) vertical and variable-wise perturbation profiles, (ii) latent-space trajectory deviation, and (iii) cross-model transferability. Experiments on extreme precipitation events from WeatherBench2 demonstrate that our method achieves effective precipitation reduction while yielding more physically plausible interventions than adversarial perturbations.

• The paper introduces a gradient-guided diffusion framework that steers generative weather models to achieve targeted precipitation reduction.
• It employs an efficient inference scheme by applying guidance at select noise levels, resulting in sparse, smooth, and physically plausible perturbations.
• The method outperforms adversarial approaches by maintaining global forecast integrity and demonstrating strong cross-model transferability.

Guided Diffusion Sampling for Precipitation Forecast Interventions

Introduction

This work introduces a novel framework for physically plausible perturbation-based precipitation-reduction interventions in data-driven weather forecasting using diffusion models. Traditional weather control interventions with numerical weather prediction (NWP) systems are computationally prohibitive for real-time, large-scale experimentation, while adversarial attacks on data-driven models yield unrealistic, model artifact-driven perturbations. To address these limitations, this study steers the diffusion sampling process of GenCast—a state-of-the-art diffusion-based weather forecasting model—with a gradient-based guidance technique targeting precipitation reduction, achieving interventions that are better aligned with atmospheric dynamics.

Methodology

Problem Formulation

The atmospheric state, $\mathbf{X}_t$, comprising multiple variables over a high-dimensional spatial grid, is forecasted recursively using a data-driven model $f$. The intervention goal is formalized as minimizing precipitation in a target region at forecast step $t+T$ by applying a small, time-constrained perturbation $\delta_{t+1}$ at step $t+1$. Unlike adversarial attacks that directly optimize $\delta_{t+1}$ via PGD or similar approaches, often resulting in physically implausible states, this work proposes an indirect approach.

Diffusion-based Guidance

Instead of acting on the atmospheric state, the intervention uses gradient guidance during the diffusion denoising process of GenCast. At each diffusion step, the denoised residual $Z_{t+1}$ is adjusted based on the gradient of a region-specific precipitation loss, modulated by a guidance scale $\lambda$. This guided diffusion trajectory inherently constrains the forecasted state to remain close to the data manifold learned by the generative model, leading to interventions that better respect the underlying atmospheric distribution.

Computational Considerations

To alleviate the prohibitive computational cost of full backpropagation through long diffusion chains, an efficient approximate inference scheme is implemented: gradient guidance acts at only a few representative noise levels, partitioned across the total $N$ steps, allowing for rapid generation of guided interventions at wall-clock times approximately half that required by state-of-the-art adversarial methods.

Empirical Evaluation

Dataset and Experimental Setup

Experiments rely on a new dataset of 219 global extreme precipitation events from WeatherBench2 (2022, 1° resolution), where event selection is based on 99th percentile climatology-corrected anomalies. The task is to suppress precipitation in ±2° neighborhoods around these extreme events. The baseline for comparison is AOWF, a published adversarial attack framework for weather models.

Quantitative Precipitation Control

The guided diffusion intervention achieves a tunable trade-off between strength of precipitation reduction in the target region and out-of-region forecast degradation. While AOWF obtains the highest reduction ratios (up to ~95% reduction, 100% success rate in suppressing extremes), it induces large out-of-region RMSE and MAE, indicating severe global forecast inconsistency. By contrast, the proposed method, at moderate guidance strengths ($\lambda$), yields substantial reduction (up to ~85% at extreme settings) while keeping out-of-region RMSE an order of magnitude lower, with success rates nearing or matching those of AOWF as guidance increases.

Qualitative and Structural Plausibility

Intervention profiles are analyzed in both variable-pressure space and latent space. Key findings include:

• Vertical Structure: The guided diffusion approach localizes interventions primarily in the lower-to-mid troposphere (surface–700 hPa), corresponding with physical mechanisms of mesoscale precipitation. Conversely, AOWF perturbations show large amplitudes at upper levels inconsistent with precipitation physics.
• Sparsity and Smoothness: The proposed method generates sparse, spatially smooth perturbations that do not excite high-frequency or gravity-wave artifacts, further supporting plausibility.

Latent Space Consistency

Analysis of posterior denoiser representations in GenCast reveals that the guided diffusion's latent trajectory remains closely aligned with both reanalysis and the model's unperturbed forecast, as evidenced by low RMSE and high cosine similarity in the latent space—even at later diffusion steps and high guidance strength. By contrast, AOWF causes abrupt and large deviations in latent space, especially in late diffusion steps, indicative of departures from physically admissible atmospheric states.

Cross-model Transferability

To test physical plausibility beyond the idiosyncrasies of a single model, the transferability of perturbations is evaluated by applying those optimized for GenCast into GraphCast, a distinct, independently trained GNN-based weather model with different architecture and time resolution. The guided diffusion method consistently achieves transferable precipitation reduction effects across cases (Japan, Brazil, India, Australia, Mozambique) and temporal lead times (+6h, +12h), whereas AOWF's impact is inconsistent or absent, highlighting its lack of physical generalizability. Quantitatively, guided diffusion maintains higher reduction ratios and success rates under transfer, with minimal out-of-region forecast distortion.

Spatial Pattern Consistency

Using the Fractions Skill Score (FSS) at multiple spatial scales, guided diffusion maintains FSS above 0.96 for all guidance values, indicating robust preservation of surrounding precipitation patterns outside the intervention zone. In contrast, AOWF's FSS saturates at much lower values (~0.48–0.58), further confirming non-local breakdown of forecast coherence.

Implications and Future Prospects

This framework marks a pivotal methodological advance for real-time, data-driven weather intervention studies. The essential insight is that guidance in diffusion-space, as opposed to direct state perturbation, exploits generative model priors to regularize interventions, yielding targeted effects while maintaining fidelity to global atmospheric structure. The strong cross-model transferability and smoothness of interventions suggest that this approach could inform the design of future operational weather modification protocols and potentially interact with NWP systems, pending further validation on dynamical stability and physical feasibility in full physics-based simulations.

At a theoretical level, this work positions diffusion models not only as generative surrogates for weather prediction but as differentiable controllers for forecast-constrained optimization in chaotic dynamical systems. Moreover, the framework is agnostic to the variables or regions of intervention, paving the way for experiments on humidity, storm track modification, or even multi-objective control.

Key future directions include:

• Integration and validation of guided diffusion interventions within high-resolution NWP systems, addressing the challenge of dynamically balanced initial states and long-term nonlinear sensitivity.
• Enhancing spatial and variable flexibility of the interventions, to enable more granular, operationally feasible weather control scenarios.
• Systematic study of ethical and environmental risk in the context of real-world deployment, although beyond the technical focus of the present work.

Conclusion

Guided diffusion sampling constitutes a robust, evaluation-supported framework for precipitation-reduction interventions in data-driven atmospheric models. By steering generative model trajectories rather than direct state-space optimization, physically plausible and transferable interventions are achieved, overcoming the limitations of adversarial perturbation approaches. The method's computational efficiency, latent- and field-level plausibility, and broad extensibility underscore its importance as a new tool for scientific and operational studies in weather modification and controlled forecast interventions.