[Paper Review] Forecasting Global Weather with Graph Neural Networks

Ryan Keisler|arXiv (Cornell University)|Feb 15, 2022

Meteorological Phenomena and Simulations126 citations

TL;DR

This paper presents a data-driven weather forecasting model using graph neural networks that evolves the global atmospheric state in 6-hour steps, achieving competitive performance to operational models on 1-degree grids and enabling integration with live GFS forecasts.

ABSTRACT

We present a data-driven approach for forecasting global weather using graph neural networks. The system learns to step forward the current 3D atmospheric state by six hours, and multiple steps are chained together to produce skillful forecasts going out several days into the future. The underlying model is trained on reanalysis data from ERA5 or forecast data from GFS. Test performance on metrics such as Z500 (geopotential height) and T850 (temperature) improves upon previous data-driven approaches and is comparable to operational, full-resolution, physical models from GFS and ECMWF, at least when evaluated on 1-degree scales and when using reanalysis initial conditions. We also show results from connecting this data-driven model to live, operational forecasts from GFS.

Motivation & Objective

• Develop a scalable data-driven model to forecast global weather by moving the current 3D atmospheric state forward by six hours.
• Leverage graph neural networks to handle spherical geometry and dense, multi-variable atmospheric data.
• Evaluate performance against previous data-driven approaches and against operational, physics-based models.
• Demonstrate compatibility with live, operational forecast systems (GFS) to form a hybrid physics+ML workflow.

Proposed method

• Use a three-component architecture (Encoder, Processor, Decoder) where the Encoder maps latitude/longitude grid data to an icosahedron latent grid, the Processor performs iterative message-passing on latent features, and the Decoder maps back to the latitude/longitude grid to predict 6-hour state changes.
• Model operates on 78 channels (6 variables x 13 pressure levels) on a 1-degree grid, across 5,882 latent nodes on an icosahedron grid.
• Train with a multi-step rollout loss (4, 8, 12 steps corresponding to 1–3 day rollouts) and normalize losses by the temporal-difference variance of each variable.
• Train on ERA5 reanalysis data and, separately, on historical GFS forecasts, using static land-sea mask and orography plus solar radiation as inputs.
• Utilize gradients and memory-efficient strategies on NVIDIA A100 hardware; 6-hour step latency is 0.04 seconds per step, enabling ~0.8 seconds for a 5-day forecast.

Experimental results

Research questions

• RQ1 Can a data-driven, graph-neural-network-based model forecast global weather with higher fidelity than prior data-driven methods at 1-degree resolution?
• RQ2 How does a 6-hour-step, autoregressive rollout perform for multi-day forecasts when trained with multi-step loss on dense atmospheric data?
• RQ3 To what extent can a GNN-based forecaster match or approach operational NWP models (GFS/ECMWF) on reanalysis-initialized benchmarks?
• RQ4 Is it feasible to couple a data-driven forecast engine with live operational forecasts (e.g., GFS) to produce accurate, timely hybrid forecasts?

Key findings

• The model learns to predict 6-hour state changes across 78 channels on a 1-degree grid and can roll out multiple steps to produce forecasts over several days.
• On ERA5-based tests, the data-driven forecasts outperform previous data-driven models and are comparable to production-scale GFS/ECMWF performance on 1-degree scales with reanalysis initial conditions.
• Compared to GFS v15.2 and ECMWF production in 2020, the data-driven model generally outperforms GFS and is comparable to ECMWF when evaluated on extratropics with smoothing to mimic truncation.
• The data-driven model can be connected to live GFS data to create a hybrid system that anticipates forecast-to-forecast changes ahead of published updates, demonstrated with a 3-day rollout.
• The approach shows stability up to ~6 days of rollout, with hexagonal instabilities emerging beyond that, potentially addressable by methodological tweaks.