TL;DR
This paper introduces PASSAT, a novel deep learning model for weather prediction that uniquely combines physical laws (advection and Navier-Stokes equations) with the Earth's spherical topology. By addressing limitations of existing models that either ignore physics or treat the Earth as a flat plane, PASSAT achieves state-of-the-art performance in medium-term weather forecasting.
What Problem Does This Paper Address?
Accurate weather prediction is crucial for societal and economic well-being. Traditional Numerical Weather Prediction (NWP) models, while physics-based, struggle with the complexity of Earth-atmosphere interactions and are computationally intensive. More recently, deep learning (DL) models have shown promise, offering faster predictions by learning from historical data. However, many DL models overlook the underlying physics governing weather evolution or the Earth's spherical geometry, leading to unreliable predictions and distortions, especially near the poles. This paper aims to bridge this gap by developing a DL model that explicitly incorporates both physical principles and the Earth's spherical topology.
Methods
The proposed model, PASSAT (Physics-ASSisted And Topology-informed), tackles weather prediction by:
- Physics Integration: It numerically solves the advection equation (describing how weather variables move) and the Navier-Stokes equation (describing the evolution of velocity fields) on a spherical manifold. This ensures that the model adheres to fundamental physical laws.
- Topology Awareness: Instead of using planar latitude-longitude grids, PASSAT operates directly on the Earth's spherical surface. This is achieved by using a spherical graph neural network (GNN) that inherently understands the spherical geometry, avoiding distortions that arise from projecting a sphere onto a plane.
- Earth-Atmosphere Interaction Modeling: A key challenge in NWP is modeling Earth-atmosphere interactions. PASSAT uses a dedicated "interaction branch" within its spherical GNN to estimate these complex effects.
- Velocity Field Estimation: The model also employs a "velocity branch" within the GNN to estimate initial velocity fields, which are then updated by solving the Navier-Stokes equation.
The overall architecture involves a backbone spherical GNN with separate branches for estimating Earth-atmosphere interactions and initial velocity fields. The model then iteratively updates weather variables and velocity fields by solving the advection and Navier-Stokes equations on the spherical manifold.
Datasets
The experiments were conducted using the ERA5 dataset, a widely used reanalysis dataset for weather and climate. Specifically, the authors used data with a 5.625° resolution, spanning from 1979 to 2018. The data was split into training (1979-2015), validation (2016), and testing (2017-2018) sets. The weather variables considered were temperature at 2m (t2m), temperature at 850hPa (t850), geopotential at 500hPa (z500), and the u and v components of wind at 10m (u10, v10).
Key Insights and Ideas
- Physics and Topology are Synergistic: The paper strongly suggests that both incorporating physical laws and respecting the Earth's spherical topology are crucial for achieving high-performance weather prediction. Ablation studies show that neglecting either significantly degrades performance.
- Spherical GNNs for Earth Science: The use of spherical GNNs is a key innovation, enabling models to naturally handle the Earth's geometry and avoid distortions that plague planar-based approaches.
- Bridging NWP and DL: PASSAT represents a significant step towards integrating the strengths of physics-based NWP models (accuracy, robustness) with the efficiency and data-learning capabilities of deep learning.
- Navier-Stokes on a Sphere: The paper demonstrates a practical method for numerically solving the Navier-Stokes equation on a spherical manifold within a deep learning framework, a non-trivial task.
Key Results
PASSAT demonstrated competitive performance, outperforming state-of-the-art deep learning models and the operational NWP model IFS T42 on the 5.625° resolution ERA5 dataset. The paper highlights that PASSAT achieved better RMSE (Root Mean Square Error) and ACC (Anomaly Correlation Coefficient) across various weather variables and lead times compared to models like Pangu, GraphCast, and FourCastNet. Notably, GraphCast, which also considers Earth's topology, was closely matched by PASSAT, but PASSAT's inclusion of physics provided an edge. The IFS T42 model, despite being a finer-resolution physical model, was surpassed by PASSAT.
Limitations and Future Work
The authors acknowledge that while PASSAT achieves impressive results, it comes with higher computational complexity compared to purely data-driven DL models. Future work includes:
- Expanding Variable Scope: Incorporating more weather variables into the model.
- Higher Resolution Data: Training on datasets with finer spatial resolutions.
- Improved Time Integration: Exploring more efficient time integration schemes than the Euler method used in the current implementation.
The authors express optimism that PASSAT's approach of combining physics, topology, and historical data will inspire further research in data-driven weather prediction.
Summary updated