TL;DR
PASSAT is a novel deep learning model that outperforms state-of-the-art AI and operational weather models by solving fundamental fluid dynamics equations (Advection and Navier-Stokes) directly on a spherical graph. By respecting both the physical laws of the atmosphere and the actual shape of the Earth, it provides more robust and accurate medium-term weather forecasts.
What Problem Does This Paper Address?
Modern weather prediction generally falls into two camps: Numerical Weather Prediction (NWP) and Deep Learning (DL).
NWP models are built on rigorous physics but struggle with "messy" processes like cloud formation and radiation (Earth-atmosphere interactions), which are computationally expensive and difficult to model. On the other hand, DL models are incredibly fast and learn directly from data, but they have two major "blind spots":
- Physics Ignorance: They often treat weather as a sequence of images, ignoring the physical constraints (like fluid dynamics) that govern how air actually moves.
- Topological Distortion: Most DL models treat the Earth as a flat 2D map. This creates massive distortions at the poles, where a single point on a sphere is stretched across the entire top edge of a rectangular grid, leading to biased predictions.
The authors of PASSAT (Physics-ASSisted And Topology-informed) aim to combine the best of both worlds.
Methods
PASSAT decomposes weather evolution into two components: the advection process (the movement of air and moisture) and Earth-atmosphere interactions (complex local effects).
- The Physics Engine: Unlike "black-box" AI, PASSAT numerically solves the Advection equation and the Navier-Stokes equation. Crucially, it solves these on a spherical manifold rather than a flat grid, ensuring the math matches the Earth's geometry.
- The Neural Engine: It uses a Spherical Graph Neural Network (SGNN) to handle the parts physics equations struggle with. The SGNN does two things:
- It estimates the initial velocity fields needed to start the physics simulations.
- It learns to predict the "tendency" of Earth-atmosphere interactions that aren't captured by the fluid equations.
- Time Integration: The model uses an iterative approach (Euler’s method) with a small 0.2-hour step size to evolve the weather state forward in time, alternating between neural network updates and physical equation solving.
Datasets
The researchers evaluated PASSAT using the ERA5 dataset, the industry-standard "gold record" of global climate data.
- Resolution: 5.625° (a standard benchmark for global modeling).
- Timeframe: Training on data from 1979–2015, validating on 2016, and testing on 2017–2018.
- Variables: The model predicts five key variables: temperature at 2m (t2m), temperature at 850hPa (t850), geopotential at 500hPa (z500), and wind components at 10m (u10, v10).
Key Insights and Ideas
- Navier-Stokes is the Missing Link: Previous hybrid models (like ClimODE) used neural networks to update velocity fields. PASSAT shows that using the actual Navier-Stokes equation to evolve velocity leads to significantly better medium-term stability.
- Topology is a First-Class Citizen: By building the model on a spherical graph, the authors eliminated the "pole problem." Their ablation studies proved that for variables like wind and geopotential, respecting the Earth's shape was actually more important for accuracy than the physics equations themselves.
- Efficiency through Hybridization: Because the physics equations handle the heavy lifting of "moving" the weather, the underlying Graph Neural Network can be much simpler (fewer nodes and edges) than purely data-driven models like GraphCast while achieving better results.
Key Results
PASSAT was compared against heavyweights like GraphCast, Pangu-Weather, and FourCastNet, as well as the operational IFS T42 model.
- Superior Accuracy: PASSAT outperformed all other DL models across all variables and lead times (up to 144 hours).
- Beating the Pros: It consistently beat the operational NWP model (IFS T42). While a higher-resolution version of the operational model (IFS T63) performed better in some metrics, it requires vastly more computational power than PASSAT.
- Robustness: In ablation tests, the "full" PASSAT model significantly outperformed versions that lacked either the physics constraints or the spherical topology, proving that the hybrid approach is synergistic.
Limitations and Future Work
While impressive, PASSAT currently has a few hurdles to clear before it replaces your local weather app:
- Variable Scope: The current study only predicts five variables. Real-world forecasting requires dozens more (humidity, precipitation, etc.).
- Resolution: The 5.625° resolution is relatively coarse. The authors plan to test the model on finer-grained data in the future.
- Integration Speed: The model currently uses Euler’s method for time integration. The authors acknowledge that more sophisticated, efficient integration schemes could further speed up training and inference.
Paper Citation: Zheng, J., Ling, Q., & Feng, Y. (2025). Physics-Assisted and Topology-Informed Deep Learning for Weather Prediction. arXiv:2505.04918v1.
Summary updated