Enhanced Precipitation Forecasting: Integrating Physics, Atmospheric Dynamics, and Deep Learning Using Omega-GNN and Omega-EGNN Models

Introduction

In the realm of artificial intelligence (AI), the fusion of data-driven meteorological models and climate models has yielded remarkable progress, rivaling and even surpassing the accuracy of traditional numerical models. However, prevailing deep learning models often grapple with low physical consistency and suboptimal forecasting of divergent winds. These limitations impede the predictive capabilities for intricate weather and climate phenomena, particularly precipitation.

Addressing Challenges in Precipitation Forecasting

To surmount the challenges in precipitation forecasting, notably for heavy rainfall events, researchers have explored a myriad of approaches. A promising strategy entails the integration of physics, atmospheric dynamics, and deep learning models. This approach capitalizes on the strengths of each discipline to bolster the accuracy and physical interpretability of precipitation forecasts.

Leveraging EarthLab and Graph Neural Networks

In the pursuit of enhancing precipitation forecasting, a research team spearheaded by Prof. Huang Gang from the Institute of Atmospheric Physics (IAP) at the Chinese Academy of Sciences harnessed the capabilities of EarthLab, a novel Earth System Science Numerical Simulator Facility conceived by IAP. EarthLab provided comprehensive data and computational resources, facilitating the development and evaluation of numerical models.

The research team focused on coupling physical variables via graph neural networks (GNNs) to introduce physical constraints and augment the precision of precipitation forecasts. GNNs, a type of deep learning model, are adept at capturing intricate relationships between data points arranged in a graph structure.

Constructing a Variable Coupling Graph

To construct a variable coupling graph, the researchers drew upon the omega equation and water vapor equation for variable selection. These equations epitomize vertical motion and water vapor variations, respectively, both of which are critical factors influencing precipitation. The graph network distilled these equations into a network structure, reflecting the nonlinear combinations of fundamental physical quantities and the interconnections between key precipitation factors.

Incorporating Sparse Data and Localizing ChebNet Graph Neural Network

Considering the profound influence of climate factors on weather scales, particularly systematic disparities in model errors under varying climate backgrounds, the study incorporated sparse data, encompassing seasonality, El Niño Southern Oscillation (ENSO), and initialization time, utilizing entity embedding techniques to calibrate the model.

Furthermore, the research team localized the ChebNet graph neural network for precipitation, preserving its efficacy while significantly reducing computational complexity by eschewing global operations.

Comparison Results and Performance Evaluation

The comparison outcomes of the proposed models, omega-GNN and omega-EGNN, vis-à-vis numerical models unveiled a substantial improvement in precipitation forecasting skills across diverse categories. The performance of these models eclipsed that of mainstream physics-unconstrained deep learning models, such as U-NET and 3D-CNN.

The omega-GNN and omega-EGNN models demonstrated superior prowess in capturing the spatial distribution and intensity of precipitation, especially for heavy rainfall events. Moreover, these models exhibited superior generalization capabilities across diverse regions and seasons, attesting to their robustness and adaptability.

Conclusion

The research findings provide invaluable insights into refining precipitation forecasting through a novel approach that harmonizes physics, atmospheric dynamics, and deep learning. The proposed omega-GNN and omega-EGNN models effectively integrate physical constraints and sparse data to augment the accuracy and physical interpretability of precipitation forecasts.

The study ushers in new avenues for further research in weather and climate modeling, underscoring the potential of synergizing data-driven and physics-based approaches to tackle intricate meteorological and climate challenges.