FlowCast: Advancing Precipitation Nowcasting with Conditional Flow Matching
Bernardo Perrone Ribeiro, Jana Faganeli Pucer
TL;DR
FlowCast introduces a novel probabilistic model using Conditional Flow Matching for rapid, high-fidelity precipitation nowcasting, surpassing diffusion models in accuracy and efficiency for short-term weather prediction.
Contribution
This paper presents FlowCast, the first end-to-end probabilistic precipitation nowcasting model utilizing CFM, offering faster sampling and improved accuracy over diffusion-based methods.
Findings
FlowCast achieves state-of-the-art probabilistic nowcasting performance.
CFM outperforms diffusion models in accuracy and efficiency.
FlowCast enables rapid, high-quality precipitation forecasts.
Abstract
Radar-based precipitation nowcasting, the task of forecasting short-term precipitation fields from previous radar images, is a critical problem for flood risk management and decision-making. While deep learning has substantially advanced this field, two challenges remain fundamental: the uncertainty of atmospheric dynamics and the efficient modeling of high-dimensional data. Diffusion models have shown strong promise by producing sharp, reliable forecasts, but their iterative sampling process is computationally prohibitive for time-critical applications. We introduce FlowCast, the first end-to-end probabilistic model leveraging Conditional Flow Matching (CFM) as a direct noise-to-data generative framework for precipitation nowcasting. Unlike hybrid approaches, FlowCast learns a direct noise-to-data mapping in a compressed latent space, enabling rapid, high-fidelity sample generation.…
Peer Reviews
Decision·ICLR 2026 Poster
1. **Innovation and Practicality of the Methodology:** - The paper innovatively introduces the highly efficient CFM generative framework to the high-dimensional spatiotemporal prediction problem of precipitation nowcasting. This represents a valuable exploration of cutting-edge methods in the field. - A direct comparison and ablation study with diffusion models clearly demonstrates the significant advantage of CFM in terms of efficiency, showing that it can maintain or close
1. **Incomplete Experimental Results and Lack of Convincing Evidence:** - **Inadequate Metric Presentation:** The experimental evaluation lacks quantitative verification of structured forecasts. For instance, the absence of analysis using spatial verification methods (e.g., FSS at different scales) prevents a thorough validation of the model's ability to predict the structure and location of precipitation systems. Furthermore, the authors did not provide complete performance curves for
Significance: improving the performance and democratising access to precipitation forecasting is an important problem as addressed in this paper. Clarity: the paper is in general clearly written Quality: the methodology is sound, the description seems correct and reproducible. The ablation studies show the justification for experimental decisions.
Quality: The main argument is that swapping the training procedure from the diffusion model to conditional flow matching improves the performance, both in terms of accuracy and in terms of computational efficiency. The problem I see with this argument, however, is that there is previous evidence, in other domains, that the CFM may improve upon the efficiency and performance [1]. I would expect in this case for the authors to link it to the background and then say something that goes beyond the
1. This work is the first to apply Conditional Flow Matching to precipitation nowcasting, learning a direct noise-to-data mapping. It achieves high-fidelity forecasting with only 10 sampling steps. 2. This work adopts VAE latent space compression and Cuboid Attention architecture, balancing the efficiency of high-dimensional radar data processing and the ability of spatiotemporal dynamic modeling.
1. Conditional Flow Matching achieves the mapping from noise to data by learning a vector field. However, the paper does not design a prior structure for the vector field in combination with the physical laws of precipitation, relying entirely on data-driven learning. This may lead the model to generate results lacking physical condition constraints. 2. This work applies CFM to precipitation nowcasting for the first time, but the idea of CFM has already been validated in some tasks within the fi
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsPrecipitation Measurement and Analysis · Meteorological Phenomena and Simulations · Flood Risk Assessment and Management