Progress Report
Last updated:
Artificial generation of upstream maritime heavy rains to govern intense-rain-induced disasters over land (AMAGOI)[2] Deep Neural Network for Weather
Progress until FY2025
1. Outline of the project
To achieve model predictive control of extreme weather phenomena such as linear precipitation bands, it is effective to perform efficient numerical experiments using fast and accurate data-driven forecasting models. During 2025, we advanced our data-driven regional weather prediction model from a deterministic to a probabilistic framework enabling ensemble prediction. By introducing perturbations in latent space including the Bred Vector method and a distribution-aware loss function, we significantly improved ensemble spread growth and reduced underestimation of extreme rainfall compared to Gaussian noise perturbations.
We also conducted research on intervention-based forecasting to suppress heavy rainfall in data-driven prediction models. By guiding the sampling process of a diffusion model with gradients, our method controls the trajectory of the generative process rather than directly perturbing atmospheric states, achieving physically consistent interventions aligned with atmospheric distributions.
2. Outcome so far
We extended a Swin-Unet-based deterministic weather model to two probabilistic ensemble models: Swin-Unet-BV, incorporating the BV method, and Probabilistic Swin-Unet, which introduces perturbations within decoder layers. The latter employs the Continuous Ranked Probability Score (CRPS) as a loss function. Both models demonstrated substantially enhanced ensemble spread across all physical variables compared to conventional Gaussian perturbations of initial conditions. Validation using a heavy rainfall event in the Kyushu region further showed that incorporating the CRPS loss reduces the underestimation of extreme precipitation relative to the deterministic model (Fig.1).
We proposed a gradient-guided intervention method for diffusion-based prediction models that steers the sampling process toward desired atmospheric states. Unlike the existing methods such as AOWF, which directly perturb atmospheric states, our approach controls the generative trajectory, enabling precipitation reduction while preserving physical consistency. Numerical experiments on a heavy rainfall event in the Kyushu region show that, whereas conventional methods nearly eliminate the rainfall, our approach shifts it southward (Fig.2). Further evaluations across vertical structure, variable-wise perturbations, latent-space trajectory deviation, and model transferability demonstrate both effectiveness and physical plausibility on extreme precipitation cases in WeatherBench2.
3. Future plans
To enable large-scale ensemble data generation across multiple scenarios at practical computational cost, we will further optimize the previously developed surrogate models for speed and memory efficiency. This includes reducing model parameters while maintaining accuracy through input variable selection and extraction of effective subnetworks.
For intervention prediction, we will develop methods that minimize intervention magnitude using adversarial perturbation techniques such as projected gradient descent. We will also explore optimization approaches with regularization that accounts for both the magnitude and simplicity of interventions to enhance real-world feasibility.