💻 computer science

RainODE: Continuous-Time Precipitation Forecasting with Latent Neural ODEs

RainODE is a continuous-time precipitation forecasting framework that combines latent Neural ODEs for capturing large-scale advective dynamics with a Brownian Bridge-based stochastic module to refine non-advective intensity variations, thereby enabling accurate, sharp, and arbitrary-time predictions.

Original authors: Yeeun Seong, Doyi Kim, Minseok Seo, Changick Kim

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yeeun Seong, Doyi Kim, Minseok Seo, Changick Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Stop-Motion" vs. "Real Life" Gap

Imagine you are trying to predict how a crowd of people will move through a park.

Old Methods (Discrete Models): These models take a photo every 5 minutes and try to guess where everyone will be in the next photo. If they guess wrong in the first photo, that mistake gets worse in the second, and even worse in the third. It's like playing "Telephone" with a crowd; the message (or the position) gets garbled over time.
The Limitation: Real life doesn't happen in 5-minute jumps. Rain doesn't just "snap" from one spot to another; it flows, grows, and shrinks continuously. But because computers usually only look at snapshots, they struggle to predict what happens between those snapshots or to predict far into the future without getting blurry and wrong.

The Solution: RainODE (The "Smooth Movie" Approach)

The researchers at KAIST created RainODE. Instead of taking snapshots, they treat rain forecasting like watching a smooth, continuous movie.

They use a mathematical tool called a Neural ODE (Ordinary Differential Equation). Think of this as a "physics engine" for the weather. Instead of guessing the next frame, the model learns the rules of motion (how rain moves, drifts, and flows).

The Analogy:

Old Way: Guessing where a car will be in 10 minutes by looking at where it was 5 minutes ago, then 5 minutes before that.
RainODE Way: Understanding the car's speed and direction, then using a calculator to say, "If it keeps going this way, it will be exactly here at 10:05, 10:06, or 10:07."

This allows the model to predict rain at any time you ask (e.g., 12 minutes from now, or 47 minutes from now), not just at the specific times it was trained on.

The Two-Part Engine

The paper explains that rain is tricky because it does two things at once:

Moving: Rain clouds drift across the sky (like a train on a track).
Changing: Rain can suddenly get heavier, lighter, or start/stop in specific spots (like a sprinkler turning on/off).

RainODE handles these with two distinct modules:

1. The Deterministic ODE (The "Train Track")

This part handles the movement. It learns the smooth, large-scale flow of the rain.

The Metaphor: Imagine a train moving along a track. The ODE knows exactly where the train is at any second. It's very good at predicting where the rain is going.
The Problem: If you only use this, the prediction looks too smooth. It's like a cartoon where the rain just slides around without any texture. It misses the sudden "puffs" of heavy rain or the tiny spots where it stops.

2. The Stochastic Source Modeling (The "Sprinkler")

This is the paper's secret sauce. Since the "train track" is too smooth, they add a Stochastic Source Modeling (SSM) module.

The Metaphor: Imagine the train is carrying a sprinkler system that adds random, fine-grained details. This module uses a "Brownian Bridge" (a fancy math way of saying "controlled randomness") to fill in the gaps. It adds the "grit," the sudden heavy downpours, and the tiny details that the smooth train track missed.
The Result: You get the smooth movement of the train plus the realistic, sharp details of the sprinkler.

The New "Gym" for Testing: RAPID

To prove their model works, the researchers built a new testing ground called RAPID.

The Old Gym (SEVIR): Everyone has been training on this dataset, which only tests predictions in 5-minute chunks for one hour.
The New Gym (RAPID): This is a new dataset based on real radar data from Korea. It allows testing predictions at 10, 30, and 60-minute intervals, all the way up to 6 hours into the future.
Why it matters: It's like moving from a gym with only short, light weights to one with heavy, long-distance weights. It tests if the model can actually handle long, complex forecasts.

What Did They Find?

Better Long-Term Predictions: When predicting 6 hours ahead, old models (like the "stop-motion" ones) fall apart. They get blurry or predict rain everywhere at once. RainODE stays sharp and coherent.
Arbitrary Time: You can ask RainODE, "What will the rain look like in 17 minutes?" or "What about in 23 minutes?" and it gives a clear answer. Old models can only answer for 15, 20, or 25 minutes.
No "Blurry" Rain: Because of the "Sprinkler" (SSM) module, the rain predictions look realistic. They don't just look like a smooth, gray fog; they have the sharp edges and heavy spots of real storms.

The Catch (Limitations)

The authors are honest about the limits.

Radar Only: The model only looks at radar images. It doesn't see the wind, temperature, or humidity.
The "Clear Sky" Problem: If it's currently sunny and a storm is about to form out of nowhere, the model can't see it coming because it only sees what's already on the radar.
Long Haul: Predicting 6 hours ahead is still very hard. The further out you go, the more the "rules of motion" get complicated by things the radar can't see.

Summary

RainODE is a new way to forecast rain that stops treating time like a series of disconnected photos and starts treating it like a continuous movie. By combining a smooth "motion engine" with a "detail injector," it can predict rain at any time you want, keeping the forecast sharp and realistic even hours into the future.

Technical Summary: RainODE

Problem Statement

Precipitation forecasting faces a fundamental tension between accuracy and temporal resolution. While data-driven approaches have advanced significantly using large-scale benchmarks like SEVIR, they are inherently constrained by discrete-time formulations. Existing models—whether autoregressive, many-to-many, or implicit—struggle when temporal resolution increases. Autoregressive models suffer from error accumulation during dense rollouts, many-to-many models incur rapidly increasing computational costs, and implicit models often fail to enforce dynamical consistency across time points. Furthermore, purely deterministic continuous-time models (e.g., standard Neural ODEs) tend to produce over-smoothed predictions, failing to capture non-advective intensity changes such as localized growth, decay, and sub-grid variability.

Methodology: RainODE Framework

The authors propose RainODE, a framework that reformulates precipitation forecasting as a continuous-time dynamical system using Latent Neural ODEs. The approach is grounded in the physical decomposition of precipitation dynamics into advection (transport), diffusion/mixing, and source/sink processes.

1. Continuous-Time Latent Dynamics (Deterministic Component)

RainODE models the coherent, transport-like evolution of precipitation in a latent space.

Latent Encoding: Input radar sequences are encoded into latent representations using a 2D spatial encoder.
Neural ODE: A Translator network predicts an initial future latent state, which serves as the initialization for a Neural ODE. The latent trajectory $z(s)$ evolves continuously over a normalized time variable $s$ according to:
$\frac{dz(s)}{ds} = f_{\text{ODE}}(z(s), z(S), s)$
This formulation enforces derivative-consistent temporal dynamics, allowing the model to capture large-scale advective motion without relying on discrete frame-to-frame mappings.
Inference: The continuous trajectory is obtained via numerical integration (Runge–Kutta), enabling predictions at arbitrary time points without retraining.

2. Stochastic Source Modeling (SSM)

To address the limitation of deterministic ODEs in capturing fine-grained, non-advective variability (e.g., turbulence, localized intensification), RainODE introduces a Stochastic Source Modeling (SSM) module.

Brownian Bridge Formulation: The SSM treats the residual evolution between the deterministic ODE prediction ( $\bar{y}_0$ ) and the ground truth ( $\bar{y}_K$ ) as a conditional diffusion process.
Refinement: A forward diffusion process bridges the ODE prediction to the target state, injecting controlled stochastic perturbations based on a time-dependent variance. This restores high-frequency spatial structures and localized intensity variations while preserving the large-scale advective consistency learned in the deterministic stage.

3. Training Strategy

The framework employs a dual-objective training strategy:

Reconstruction Loss: Minimizes the discrepancy between the discrete teacher predictions and ground truth.
Consistency Loss: Enforces alignment between the continuous ODE trajectory and the discrete teacher predictions to ensure smooth temporal evolution.
SSM Integration: The stochastic refinement is applied to recover fine details that the deterministic ODE smooths out.

Key Contributions

Continuous-Time Reformulation: The paper proposes a shift from discrete frame prediction to continuous-time latent trajectory modeling, enabling arbitrary-time inference and avoiding the error accumulation typical of autoregressive dense rollouts.
Hybrid Deterministic-Stochastic Modeling: By combining a Neural ODE (for advective consistency) with a Brownian Bridge-based SSM (for intensity refinement), the model achieves both sharp predictions and dynamical consistency.
RAPID Dataset: The authors introduce RAPID (Radar-based Precipitation Integrated Dataset), a new benchmark constructed from Korean operational radar data. Unlike SEVIR, RAPID supports evaluation across multiple temporal intervals (10, 30, and 60 minutes) and extends forecasting horizons up to 6 hours, facilitating the assessment of models under heterogeneous temporal regimes.
Generalization to Unseen Intervals: The framework demonstrates the ability to generalize from training on coarse intervals (e.g., 1-hour) to fine-grained inference (e.g., 10-minute) without retraining, a capability lacking in discrete baselines.

Experimental Results

Datasets and Baselines

Experiments were conducted on SEVIR (standard benchmark) and the new RAPID dataset. Baselines included 3D U-Net, SimVP, Earthformer, PreDiff, CasCast, and exPreCast.

Quantitative Performance

SEVIR: RainODE achieved state-of-the-art performance across multiple metrics (CSI-M, CSI-th, RMSE, FSS, HSS), outperforming both discrete and generative baselines.
RAPID: RainODE showed consistent improvements across 10-, 30-, and 60-minute intervals. Notably, on the RAPID-60min (6-hour lead time) setting, discrete baselines like Earthformer degraded significantly (near-zero CSI), whereas RainODE maintained stable categorical and spatial skill scores.
High-Intensity Detection: The inclusion of SSM significantly improved performance on high-intensity thresholds (CSI-219), addressing the over-smoothing issue common in MSE-based models.

Qualitative Analysis

Spatial Coherence: RainODE produced forecasts with better alignment of precipitation location and intensity compared to baselines. While MSE-based models underestimated intensity and diffusion-based models exhibited spatial misalignment, RainODE maintained coherent structures.
Long-Horizon Stability: Visualizations up to 6 hours showed that RainODE preserved precipitation structures and intensity patterns, whereas discrete models (e.g., SimVP) suffered from progressive dissipation of precipitation intensity.
Latent Consistency: t-SNE visualizations of latent trajectories revealed that RainODE's latent space evolution was more coherent and less erratic than discrete baselines, even when evaluated at 10-minute intervals without additional training.

Significance and Claims

The paper claims that RainODE offers a promising paradigm for precipitation forecasting by moving beyond fixed temporal grids. Its primary significance lies in:

Temporal Flexibility: Enabling inference at arbitrary time points and resolutions, decoupling prediction capability from the specific training grid.
Robustness: Demonstrating superior stability in long-horizon forecasting (up to 6 hours) and across varying temporal intervals where discrete models fail.
Physical Consistency: Successfully modeling precipitation as a continuous dynamical process that balances advective transport with stochastic local evolution.

The authors acknowledge limitations, noting that long-term prediction from radar-only observations is inherently constrained by the inability to capture large-scale atmospheric dynamics that emerge from clear conditions. They suggest future work will extend the framework to multi-modal settings incorporating additional atmospheric variables.