FourCastNet: Accelerating Global High-Resolution Weather Forecasting Using Adaptive Fourier Neural Operators

Kurth, Thorsten; Subramanian, Shashank; Harrington, Peter de B.; Pathak, Jaideep; Mardani, Morteza; Hall, D.; Miele, Andrea; Kashinath, Karthik; Anandkumar, Anima

doi:10.1145/3592979.3593412

Cited by 72 publications

(42 citation statements)

References 13 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Deep learning has been investigated as a full replacement for NWP models in, for example, (Bihlo, 2021; Dueben & Bauer, 2018; Pathak et al., 2022; Weyn et al., 2019). While still not competitive in comparison with traditional NWP at high resolution, the results obtained are often comparable at coarser resolution (Pathak et al., 2022). Deep learning has also been proposed as an alternative to subgrid‐scale parameterization (Gentine et al., 2018), it has been used extensively for precipitation nowcasting (Shi et al., 2015), downscaling of meteorological fields (Baño‐Medina et al., 2020) and most recently also in combination with differential equations‐based methodologies (Bihlo & Popovych, 2022).…”

Section: Related Workmentioning

confidence: 99%

“…Meteorology, as other fields of science, has also seen an unprecedented rise in use of machine learning, and specifically deep learning, to areas that have traditionally been tackled using methods of numerical analysis and scientific computing. Deep learning has been investigated as a full replacement for NWP models in, for example, (Bihlo, 2021; Dueben & Bauer, 2018; Pathak et al., 2022; Weyn et al., 2019). While still not competitive in comparison with traditional NWP at high resolution, the results obtained are often comparable at coarser resolution (Pathak et al., 2022).…”

Section: Related Workmentioning

confidence: 99%

“…(2021) neural networks are trained on reanalysis data to produce purely data‐driven weather forecasts. While still in its infancy, if successful these data‐driven approaches would enable issuing weather forecasts at several orders of magnitude faster than conventional NWP models (Pathak et al., 2022).…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Computing the Ensemble Spread From Deterministic Weather Predictions Using Conditional Generative Adversarial Networks

Brecht

Bihlo

2023

Geophysical Research Letters

View full text Add to dashboard Cite

To accurately predict weather and climate numerical weather prediction (NWP) models are being used (Bauer et al., 2015). Recently, machine learning methods have received more attention as an alternative approach for weather and climate prediction. For example, in Bihlo (2021), Bauer (2018), andWeyn et al. (2021) neural networks are trained on reanalysis data to produce purely data-driven weather forecasts. While still in its infancy, if successful these data-driven approaches would enable issuing weather forecasts at several orders of magnitude faster than conventional NWP models (Pathak et al., 2022).Owing to the inherent uncertainty in the atmospheric dynamical system (Lorenz, 1963) it has been recognized early on in the development of NWP that a measure of uncertainty of a numerical weather forecast can substantially enhance the value of these forecasts. This gave rise to the field of ensemble weather prediction (Leutbecher & Palmer, 2008), which aims to quantify the various sources of uncertainty in NWP models, chief of which are uncertainty in the initial conditions and errors in the numerical model formulation. To overcome these uncertainties, in addition to the single deterministic weather forecast, an ensemble of perturbed forecasts is generated whose overall divergence, or spread, ideally will provide a measure of the uncertainty in the deterministic prediction. In some applications one is interested in the potential worst cases scenarios . Here, the ensemble members themselves are needed as the spread only would not provide the spatial correlation between the different points. The main limiting factor in generating the ensemble is still computational in nature as each ensemble run takes up computational resources thereby limiting the total number of such ensembles, typically less then 100, that can be computed on an operational basis.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Computing the Ensemble Spread From Deterministic Weather Predictions Using Conditional Generative Adversarial Networks

Brecht

Bihlo

2023

Geophysical Research Letters

View full text Add to dashboard Cite

show abstract

“…Most current ML‐based atmospheric emulators don't predict all the fields needed for many types of analysis. For instance, PANGU (Bi et al., 2022) doesn't predict precipitation, and FourCastNet (Pathak et al., 2022) doesn't predict top‐of‐atmosphere radiative fluxes. It is relatively straightforward to train emulators that predict a broader set of outputs that facilitate interpretability, and that may become standard practice despite adding computational overhead.…”

Section: Hybrid Rc Versus Full‐model Emulationmentioning

confidence: 99%

“…A23 use the intermediate‐complexity SPEEDY AGCM, which has 3.75° × 5° grid resolution (15–20 times coarser than recent FMEs), eight vertical levels (also less than recent FMEs), and simplified physical parameterizations. SPEEDY, run on 1,152 CPU processors, simulates a day per 2 s of execution time (Arcomano et al., 2022), comparable to current FME approaches (Lam et al., 2022; Pathak et al., 2022) run on small GPU systems. The ML element (RC) with which SPEEDY is combined uses a much longer time step than SPEEDY, so it doesn't significantly slow down simulations, while it is surprisingly effective in counteracting SPEEDY's considerable systematic weather and climate biases.…”

Section: Hybrid Rc Versus Full‐model Emulationmentioning

confidence: 99%

Old Dog, New Trick: Reservoir Computing Advances Machine Learning for Climate Modeling

Bretherton

2023

Geophysical Research Letters

View full text Add to dashboard Cite

Physics‐informed machine learning (ML) applied to geophysical simulation is developing explosively. Recently, graph neural net and vision transformer architectures have shown 1–7 days global weather forecast skill superior to any conventional model with integration times over 1,000 times faster, but longer simulations rapidly degrade. ML that achieves high skill in both weather and climate applications is a tougher goal. This Commentary was inspired by Arcomano et al. (2023, https://doi.org/10.1029/2022GL102649), who show impressive progress toward that goal using hybrid ML, combining reservoir computing (RC) to a coarse‐grid climate model and coupling to a separate data‐driven RC model that interactively predicts sea‐surface temperature. This opens new horizons; where will the next ML breakthrough come from, and is conventional climate modeling about to be disrupted?

show abstract

A shifting climate: New paradigms and challenges for (early career) scientists in extreme weather research

Kretschmer,

Jézéquel,

Labe

et al. 2024

Atmospheric Science Letters

View full text Add to dashboard Cite

Research on weather and climate extremes has become integral to climate science due to their increasing societal relevance and impacts in the context of anthropogenic climate change. In this perspective we examine recent changes and evolving paradigms in the study of extreme events, emphasizing the increasingly interdisciplinary nature of research and their societal implications. We discuss the importance of understanding the physical basis of extreme events and its linkages to climate impacts, highlighting the need for collaboration across multiple disciplines. Furthermore, we explore the challenge of big climate data analysis and the application of novel statistical methods, such as machine learning, in enhancing our understanding of extreme events. Additionally, we address the engagement with different stakeholder groups and the evolving landscape of climate services and private‐sector involvement. We conclude with reflections on the risks and opportunities for early career researchers in navigating these interdisciplinary and societal demands, stressing the importance of meaningful scientific engagement, and removing barriers to inclusivity and collaboration in climate research.

show abstract

FourCastNet: Accelerating Global High-Resolution Weather Forecasting Using Adaptive Fourier Neural Operators

Cited by 72 publications

References 13 publications

Computing the Ensemble Spread From Deterministic Weather Predictions Using Conditional Generative Adversarial Networks

Computing the Ensemble Spread From Deterministic Weather Predictions Using Conditional Generative Adversarial Networks

Old Dog, New Trick: Reservoir Computing Advances Machine Learning for Climate Modeling

A shifting climate: New paradigms and challenges for (early career) scientists in extreme weather research

Contact Info

Product

Resources

About