GraphCast: A Breakthrough in Weather Forecasting with Machine Learning

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Nov 16 2023

In a paper published in the journal Science, researchers presented "GraphCast," a novel approach that revolutionized global medium-range weather forecasting. Unlike traditional numerical weather prediction relying on increased computational power, GraphCast harnessed machine learning directly from reanalysis data. This method forecasted hundreds of weather variables globally at 0.25° resolution over 10 days, achieving unparalleled speed in under a minute.

With GraphCast surpassing the accuracy of current operational deterministic systems in 90% of 1380 verification targets, it notably enhanced severe event prediction, encompassing tropical cyclone paths, atmospheric rivers, and extreme temperatures. This pioneering advancement signified a crucial stride in precise and efficient weather forecasting, showcasing the potential of machine learning in modeling intricate dynamical systems.

GraphCast: Revolutionizing Weather Forecasting

In mid-October 2022, at the European Centre for Medium-Range Weather Forecasts (ECMWF) in Bologna, Italy, the Integrated Forecasting System (IFS) commenced its advanced calculations, predicting Earth's weather for the upcoming days and weeks. Operating at the new High-Performance Computing Facility, the IFS began disseminating its initial forecasts, a process recurrent every six hours to provide the world with highly accurate weather predictions.

The IFS and modern weather forecasting represent monumental achievements in Science and Engineering, grappling with the intricacies of one of Earth's most complex phenomena: weather systems. These forecasts significantly impact everyday decisions, from personal choices like what to wear to critical safety decisions during hazardous weather events. Numerical weather prediction (NWP), the prevailing method, involves solving weather equations using supercomputers, steadily enhancing accuracy through ongoing research and computational advancements.

However, traditional NWP needs help leveraging historical weather data to boost accuracy. It relies on expert-driven innovations in models and algorithms, a time-consuming and costly process. Machine learning-based weather prediction (MLWP) emerges as an alternative, capable of training forecast models directly from historical data, potentially capturing elusive patterns not easily represented in equations.

While NWP systems like ECMWF's High-RESolution forecast (HRES) remain most accurate for medium-range weather forecasting for up to ten days, machine learning weather prediction (MLWP) methods have steadily progressed in this domain. GraphCast introduced as an MLWP approach, stands out by producing highly accurate 10-day forecasts in under a minute using a single Google Cloud Tensor Processing Unit (TPU) v4 device. Its applications include predicting tropical cyclone tracks, atmospheric rivers, and extreme temperatures, showcasing the promise of machine learning in weather forecasting.

GraphCast predicts the subsequent weather state six hours ahead based on the two most recent weather states, represented on a 0.25° latitude/longitude grid. Implemented as a neural network architecture relying on graph neural networks (GNNs), GraphCast utilizes an "encode-process-decode" configuration comprising an encoder, processor, and decoder components.

The training involved historical data from ECMWF's ERA5 reanalysis archive, minimizing the mean squared error between predicted and actual states. The model's development and training spanned 39 years of historical data, focusing on reducing prediction errors. The training took approximately four weeks on 32 Cloud TPU v4 devices, mirroring real deployment scenarios by evaluating GraphCast on held-out data from 2018 onward.

GraphCast's Forecast Evaluation Overview

GraphCast underwent an extensive evaluation against HRES, covering diverse variables, levels, and lead times using Root-Mean-Squared Error (RMSE) and anomaly correlation coefficient (ACC) metrics. Among the 227 forecasted combinations, GraphCast's assessment of 69 variables exhibited consistent superiority over HRES across multiple lead times, mainly showcasing improved RMSE scores of up to 14% on the meteorologically significant Z500 field.

The meticulous methodology, which involved using ERA5 as the benchmark for GraphCast and creating an "HRES forecast at step 0" dataset for fair comparisons, highlighted GraphCast's dominance across 90.3% of 1380 targets. Its strength was evident across various atmospheric regions, except for stratospheric levels, where it consistently outperformed HRES, marking superiority on 89.9% of targets.

Expanding beyond variable assessments, GraphCast excelled in predicting severe weather events. Its prowess in cyclone tracking showcased notably lower errors compared to HRES, especially within 18 hours to 4.75 days. Furthermore, GraphCast demonstrated superior performance in predicting atmospheric rivers' intensity and provided enhanced precision-recall curves for extreme heat events at longer lead times, solidifying its position as an advanced forecasting tool.

These holistic evaluations, spanning variables, lead times, and severe weather event predictions confirm GraphCast's resilience and superiority over traditional forecasting techniques. Such superiority underscores its potential to significantly augment weather prediction accuracy, offering crucial support for critical decision-making processes.

Conclusion

To summarize, the comparison between GraphCast and HRES signifies a significant milestone in MLWP competitiveness. GraphCast showcased remarkable efficiency and accuracy across diverse variables and lead times, even excelling in forecasting severe weather events despite needing explicit training. This achievement marks a leap forward, promising heightened accuracy, accessibility, and tailored applications across various sectors. However, handling uncertainty remains challenging, emphasizing the need for probabilistic forecasts akin to ensemble forecasts to address increasing uncertainty over longer lead times.

While GraphCast operates at a relatively modest scale regarding parameters and resolution, it holds promise as a model family that can scale further with advancements in computing resources and higher-resolution data. The reliance on data-driven MLWP on robust datasets emphasizes the importance of rich data sources like ECMWF's MARS archive. It's crucial to note that GraphCast aims to complement rather than replace traditional forecasting methods, acknowledging their extensive development while offering potential avenues to enhance forecasting accuracy and expand machine learning's role in scientific domains beyond weather forecasting.