Architectural insights into and training methodology optimization of Pangu-Weather

<p>Data-driven medium-range weather forecasts have recently outperformed classical numerical weather prediction models, with Pangu-Weather (PGW) being the first breakthrough model to achieve this. The Transformer-based PGW introduced novel architectural components including the three-dimension...

Full description

Saved in:
Bibliographic Details
Main Authors: D. To, J. Quinting, G. A. Hoshyaripour, M. Götz, A. Streit, C. Debus
Format: Article
Language:English
Published: Copernicus Publications 2024-12-01
Series:Geoscientific Model Development
Online Access:https://gmd.copernicus.org/articles/17/8873/2024/gmd-17-8873-2024.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1846125660063399936
author D. To
J. Quinting
G. A. Hoshyaripour
M. Götz
M. Götz
A. Streit
C. Debus
author_facet D. To
J. Quinting
G. A. Hoshyaripour
M. Götz
M. Götz
A. Streit
C. Debus
author_sort D. To
collection DOAJ
description <p>Data-driven medium-range weather forecasts have recently outperformed classical numerical weather prediction models, with Pangu-Weather (PGW) being the first breakthrough model to achieve this. The Transformer-based PGW introduced novel architectural components including the three-dimensional attention mechanism (3D Transformer) in the Transformer blocks. Additionally, it features an Earth-specific positional bias term which accounts for weather states being related to the absolute position on Earth. However, the effectiveness of different architectural components is not yet well understood. Here, we reproduce the 24 h forecast model of PGW based on subsampled 6-hourly data. We then present an ablation study of PGW to better understand the sensitivity to the model architecture and training procedure. We find that using a two-dimensional attention mechanism (2D Transformer) yields a model that is more robust to training, converges faster, and produces better forecasts compared to using the 3D Transformer. The 2D Transformer reduces the overall computational requirements by 20 %–30 %. Further, the Earth-specific positional bias term can be replaced with a relative bias, reducing the model size by nearly 40 %. A sensitivity study comparing the convergence of the PGW model and the 2D-Transformer model shows large batch effects; however, the 2D-Transformer model is more robust to such effects. Lastly, we propose a new training procedure that increases the speed of convergence for the 2D-Transformer model by 30 % without any further hyperparameter tuning.</p>
format Article
id doaj-art-f99173ad8cfe45169c71bce43b7a1aac
institution Kabale University
issn 1991-959X
1991-9603
language English
publishDate 2024-12-01
publisher Copernicus Publications
record_format Article
series Geoscientific Model Development
spelling doaj-art-f99173ad8cfe45169c71bce43b7a1aac2024-12-13T10:46:13ZengCopernicus PublicationsGeoscientific Model Development1991-959X1991-96032024-12-01178873888410.5194/gmd-17-8873-2024Architectural insights into and training methodology optimization of Pangu-WeatherD. To0J. Quinting1G. A. Hoshyaripour2M. Götz3M. Götz4A. Streit5C. Debus6Scientific Computing Center (SCC), Karlsruhe Institute of Technology (KIT), Karlsruhe, GermanyInstitute of Meteorology and Climate Research, Karlsruhe Institute of Technology (KIT), Karlsruhe, GermanyInstitute of Meteorology and Climate Research, Karlsruhe Institute of Technology (KIT), Karlsruhe, GermanyScientific Computing Center (SCC), Karlsruhe Institute of Technology (KIT), Karlsruhe, GermanyHelmholtz AI, Karlsruhe, GermanyScientific Computing Center (SCC), Karlsruhe Institute of Technology (KIT), Karlsruhe, GermanyScientific Computing Center (SCC), Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany<p>Data-driven medium-range weather forecasts have recently outperformed classical numerical weather prediction models, with Pangu-Weather (PGW) being the first breakthrough model to achieve this. The Transformer-based PGW introduced novel architectural components including the three-dimensional attention mechanism (3D Transformer) in the Transformer blocks. Additionally, it features an Earth-specific positional bias term which accounts for weather states being related to the absolute position on Earth. However, the effectiveness of different architectural components is not yet well understood. Here, we reproduce the 24 h forecast model of PGW based on subsampled 6-hourly data. We then present an ablation study of PGW to better understand the sensitivity to the model architecture and training procedure. We find that using a two-dimensional attention mechanism (2D Transformer) yields a model that is more robust to training, converges faster, and produces better forecasts compared to using the 3D Transformer. The 2D Transformer reduces the overall computational requirements by 20 %–30 %. Further, the Earth-specific positional bias term can be replaced with a relative bias, reducing the model size by nearly 40 %. A sensitivity study comparing the convergence of the PGW model and the 2D-Transformer model shows large batch effects; however, the 2D-Transformer model is more robust to such effects. Lastly, we propose a new training procedure that increases the speed of convergence for the 2D-Transformer model by 30 % without any further hyperparameter tuning.</p>https://gmd.copernicus.org/articles/17/8873/2024/gmd-17-8873-2024.pdf
spellingShingle D. To
J. Quinting
G. A. Hoshyaripour
M. Götz
M. Götz
A. Streit
C. Debus
Architectural insights into and training methodology optimization of Pangu-Weather
Geoscientific Model Development
title Architectural insights into and training methodology optimization of Pangu-Weather
title_full Architectural insights into and training methodology optimization of Pangu-Weather
title_fullStr Architectural insights into and training methodology optimization of Pangu-Weather
title_full_unstemmed Architectural insights into and training methodology optimization of Pangu-Weather
title_short Architectural insights into and training methodology optimization of Pangu-Weather
title_sort architectural insights into and training methodology optimization of pangu weather
url https://gmd.copernicus.org/articles/17/8873/2024/gmd-17-8873-2024.pdf
work_keys_str_mv AT dto architecturalinsightsintoandtrainingmethodologyoptimizationofpanguweather
AT jquinting architecturalinsightsintoandtrainingmethodologyoptimizationofpanguweather
AT gahoshyaripour architecturalinsightsintoandtrainingmethodologyoptimizationofpanguweather
AT mgotz architecturalinsightsintoandtrainingmethodologyoptimizationofpanguweather
AT mgotz architecturalinsightsintoandtrainingmethodologyoptimizationofpanguweather
AT astreit architecturalinsightsintoandtrainingmethodologyoptimizationofpanguweather
AT cdebus architecturalinsightsintoandtrainingmethodologyoptimizationofpanguweather