Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generalizing Weather Forecast to Fine-grained Temporal Scales via Physics-AI Hybrid Modeling

Wanghan Xu, Fenghua Ling, Wenlong Zhang, Tao Han, Hao Chen, Wanli Ouyang, Lei Bai

TL;DR

The paper tackles the limitation of black-box ML weather forecasts in generalizing to finer temporal scales. It introduces WeatherGFT, a physics-AI hybrid that uses a PDE kernel to drive short-time physical evolution and a learnable router to adaptively correct AI-based residuals, with a lead-time conditional decoder and multi-lead-time training to enable forecasts at finer horizons than the training data. The approach demonstrates that WeatherGFT can generalize to 30-minute forecasts from hourly data and provides competitive performance on medium-range tasks, while also delivering improved nowcasting results using 30-minute ground truth for validation. The work offers a principled pathway to bridge nowcasting and mid-range forecasting, with interpretable physics-driven evolution and adaptive AI corrections that enhance generalization and robustness.

Abstract

Data-driven artificial intelligence (AI) models have made significant advancements in weather forecasting, particularly in medium-range and nowcasting. However, most data-driven weather forecasting models are black-box systems that focus on learning data mapping rather than fine-grained physical evolution in the time dimension. Consequently, the limitations in the temporal scale of datasets prevent these models from forecasting at finer time scales. This paper proposes a physics-AI hybrid model (i.e., WeatherGFT) which generalizes weather forecasts to finer-grained temporal scales beyond training dataset. Specifically, we employ a carefully designed PDE kernel to simulate physical evolution on a small time scale (e.g., 300 seconds) and use a parallel neural networks with a learnable router for bias correction. Furthermore, we introduce a lead time-aware training framework to promote the generalization of the model at different lead times. The weight analysis of physics-AI modules indicates that physics conducts major evolution while AI performs corrections adaptively. Extensive experiments show that WeatherGFT trained on an hourly dataset, effectively generalizes forecasts across multiple time scales, including 30-minute, which is even smaller than the dataset's temporal resolution.

Generalizing Weather Forecast to Fine-grained Temporal Scales via Physics-AI Hybrid Modeling

TL;DR

The paper tackles the limitation of black-box ML weather forecasts in generalizing to finer temporal scales. It introduces WeatherGFT, a physics-AI hybrid that uses a PDE kernel to drive short-time physical evolution and a learnable router to adaptively correct AI-based residuals, with a lead-time conditional decoder and multi-lead-time training to enable forecasts at finer horizons than the training data. The approach demonstrates that WeatherGFT can generalize to 30-minute forecasts from hourly data and provides competitive performance on medium-range tasks, while also delivering improved nowcasting results using 30-minute ground truth for validation. The work offers a principled pathway to bridge nowcasting and mid-range forecasting, with interpretable physics-driven evolution and adaptive AI corrections that enhance generalization and robustness.

Abstract

Data-driven artificial intelligence (AI) models have made significant advancements in weather forecasting, particularly in medium-range and nowcasting. However, most data-driven weather forecasting models are black-box systems that focus on learning data mapping rather than fine-grained physical evolution in the time dimension. Consequently, the limitations in the temporal scale of datasets prevent these models from forecasting at finer time scales. This paper proposes a physics-AI hybrid model (i.e., WeatherGFT) which generalizes weather forecasts to finer-grained temporal scales beyond training dataset. Specifically, we employ a carefully designed PDE kernel to simulate physical evolution on a small time scale (e.g., 300 seconds) and use a parallel neural networks with a learnable router for bias correction. Furthermore, we introduce a lead time-aware training framework to promote the generalization of the model at different lead times. The weight analysis of physics-AI modules indicates that physics conducts major evolution while AI performs corrections adaptively. Extensive experiments show that WeatherGFT trained on an hourly dataset, effectively generalizes forecasts across multiple time scales, including 30-minute, which is even smaller than the dataset's temporal resolution.
Paper Structure (32 sections, 26 equations, 10 figures, 6 tables)

This paper contains 32 sections, 26 equations, 10 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Data-driven Weather Forecasting.
  4. Physical Neural Networks.
  5. Method
  6. Problem Formulation
  7. WeatherGFT Overview
  8. PDE Kernel
  9. HybridBlock with Adaptive Router
  10. Lead Time Conditional Decoder
  11. Multiple Lead Time Training
  12. Experiment
  13. Experimental Setup
  14. Dataset.
  15. Tasks.
  16. ...and 17 more sections

Figures (10)

  • Figure 1: Learnable router weight. The role of physics and AI at different lead times: major evolution and adaptive correction (details in Sec. \ref{['sec:weight']}).
  • Figure 2: Overview of WeatherGFT. HybridBlock serves as the fundamental unit of the model, consisting of three PDE kernels, a parallel Attention Block, and a subsequent learnable router. A lead time conditional decoder is employed to generate forecasts for different lead times.
  • Figure 3: Router in HybridBlock. $B$ represents batch size, $L$ is the number of tokens with $D$ dimension.
  • Figure 4: Medium-Range Forecast. The x-axis represents the lead time in hours, while the y-axis represents the RMSE for different variables. The smaller RMSE the better.
  • Figure 5: Visualization of z500 Predictions.
  • ...and 5 more figures