Author: Sakura

Department: School of Earth Science, Zhejiang University

Latested Update: 2024/6/6

© Sakura, 2024. All rights reserved.

This document is protected by copyright law. No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author, except in the case of brief quotations embodied in critical reviews and certain other noncommercial uses permitted by copyright law. For permission requests, please contact the author at [email protected].

Outline:

• Introduction
• Advanced Data-driven Models for Weather Forecasting
  • I. Pangu-Weather
  • II. GraphCast
  • III. MetNet-3
  • IV. PreDiff
  • V. CorrDiff
• Conclusion
• References
• Appendix

Abstract:

Weather forecasting is important for science and society. At present, the commonly used forecasting models are numrical ones, which represents atmospheric states as discretized grids and numerically solves partial differential equations that describe the transition between those states. Though effective, they face limitations in computational efficiency and accuracy, particularly for small-scale and extreme weather phenomena. The integration of artificial intelligence (AI) into modeling represents a significant advancement in weather forecasting. This paper explores the potential of data-driven methods to address these challenges by introducing several novel deep-learning-based models: Pangu-Weather, GraphCast, MetNet-3, PreDiff, and CorrDiff. These models leverage advanced machine learning algorithms and vast datasets that outperforms the state-of-the-art numerical models. We highlight the potential of AI in weather forecasting, offering insights into future research directions and applications.

Key words: Weather forecasting; deep nueral network; Transformer; U-Net; diffusion model; prediction.

Weather forecasting aims at predicting atmospheric conditions over short periods (eg. from a few hours up to 2 weeks), using real-time observational data, including satellite data, radar data, and ground-based observations. In the past decade, high-performance computing systems have greatly accelerated research in the field of numerical weather prediction (NWP) methods (Bauer et al., 2015). However, these methods primarily use partial differential equations (PDEs) and then solve them with numerical simulations, which are time-consuming and compute-intensive (Skamarock et al., 2005). In addition, conventional NWP algorithms largely rely on parameterization, which uses approximate functions to capture unresolved processes, where errors can be introduced by approximation (Allen et al., 2006; Palmer et al., 2005).

With the development of artificial intelligence, weather models based on deep neural networks that use direct atmospheric observations for training offer an alternative modeling paradigm (Andrychowicz et al., 2023). Neural models can learn atmospheric phenomena directly from the observation without explicitly using complex physics equation even go beyond the usual domain of weather. Besides, high spatial resolution prediction is dramatically accelerated in an end-to-end neural network model. These advantageous properties make neural models a strong contender for an alternative paradigm for atmospheric modeling.

In the following chapter, we are going to introduce five novel deep-learning-based models for weather forecasting, namely Pangu-Weather (Bi et al., 2023), GraphCast (Lam et al., 2023), MetNet-3 (Andrychowicz et al., 2023), PreDiff (Gao et al., 2023), and CorrDiff (Mardani et al., 2023).

Figure. 1: Overview of Pangu-Weather. a, 3DEST architecture. b, Hierarchical temporal aggregation.

The paper "Accurate medium-range global weather forecasting with 3D neural networks" by Bi et al. (2023) introduces Pangu-Weather, an AI-based weather forecasting model leveraging 3D deep neural networks and Earth-specific priors to enhance medium-range global weather forecasting accuracy and efficiency. Unlike traditional NWP methods, which are computationally intensive and rely on approximations, Pangu-Weather uses a hierarchical temporal aggregation strategy to reduce error accumulation and significantly improve speed.

Pangu-Weather was trained using 39 years of global weather data (Hersbach et al., 2020). This extensive training allows it to perform better than the current system Integrated Forecasting System (IFS) used by the European Centre for Medium-Range Weather Forecasts (ECMWF) (Bougeault et al., 2010). Pangu-Weather is faster, gaining 10-15 hours in forecast time over ECMWF's system and up to 40 hours over other AI-based models like FourCastNet (Pathak et al., 2022).

The architecture of Pangu-Weather, known as the 3D Earth-specific transformer (3DEST), integrates height information and utilizes patch embedding and Earth-specific positional bias within an encoder-decoder framework, enhancing forecast accuracy without increasing computational costs. This results in smoother and more consistent weather predictions.

• 3D Neural Networks: Pangu-Weather incorporates height information into a new dimension, enabling the model to capture the relationship between atmospheric states at different pressure levels. The model utilizes a 3DEST architecture (see Figure. 1a).
• Hierarchical Temporal Aggregation: This strategy trains a series of models with increasing forecast lead times, significantly reducing the number of iterations needed for medium-range forecasts and mitigating cumulative forecast errors (see Figure. 1b).

Pangu-Weather outperforms the operational IFS of ECMWF in all tested weather variables. The model achieves lower root mean square error (RMSE) and higher anomaly correlation coefficient (ACC) than both IFS and FourCastNet (see Figure. 2). The model also excels in forecasting extreme weather events, including tropical cyclones, with higher accuracy than ECMWF’s high-resolution forecast (HRES) (see Figure. 3). As an AI-based method, Pangu-Weather is more than 10,000-times faster than the operational IFS. This offers an opportunity for performing large-member ensemble forecasts with small computational costs (see Figure. 4).

Figure. 2: Pangu-Weather produces higher accuracy than the operational IFS and FourCastNet in deterministic forecasts on the ERA5 data (Hersbach et al., 2020). Figure. 3: Visualization of tracking tropical cyclones. a) The tracking results of cyclone eyes for Hurricane Michael (2018–13) and Typhoon Ma-on (2022–09) by Pangu-Weather and ECMWF-HRES, with a comparison to the ground-truth by IBTrACS (Kenneth et al., 2019). b) An illustration of the tracking process, where we used Pangu-Weather as an example. c) The procedural tracking results of Typhoon Kong-rey (2018–25). Figure. 4: Ensemble forecast results of Pangu-Weather. Here, Z500, Q500 and U500 indicate the geopotential, temperature and the u-component of wind speed at 500 hPa, respectively. T2M indicates the 2-m temperature and U10 indicates the u-component of 10-m wind speed.

Despite its promising results, Pangu-Weather has limitations, such as its reliance on reanalysis data, exclusion of certain weather variables like precipitation, potential underestimation of extreme events, and possible temporal inconsistencies. The paper suggests future improvements could include incorporating more data dimensions, using deeper networks, and combining AI with traditional NWP methods.

Figure. 5: Model schematic.

• (A) The input weather state(s) are defined on a 0.25° latitude-longitude grid comprising a total of 721 × 1440 = 1,038,240 points. Yellow layers in the close-up pop-out window represent thefivesurfacevariables, and blue layers represent the six atmospheric variables that are repeated at 37 pressure levels (5 + 6 × 37 = 227 variables per point in total), resulting in a state representation of 235,680,480 values.

• (B) GraphCast predicts the next state of the weather on the grid.

• (C) A forecast is made by iteratively applying GraphCast (GC) to each previous predicted state, to produce a sequence of states that represent the weather at successive lead times.

• (D) The encoder component of the GraphCast architecture maps local regions of the input (green boxes) into nodes of the multimesh graph representation (green, upward arrows that terminate in the green-blue node).

• (E) The processor component updates each multimesh node using learned message-passing (heavy blue arrows that terminate at a node).

• (F) The decoder component maps the processed multimesh features (purple nodes) back onto the grid representation (red, downward arrows that terminate at a red box).

• (G) The multimesh is derived from icosahedral meshes of increasing resolution, from the base mesh (M0, 12 nodes) to the finest resolution (M6, 40,962 nodes), which has uniform resolution across the globe. It contains the set of nodes from M6 and all the edges from M0 to M6. The learned message-passing over the different meshes’ edges happens simultaneously, so that each node is updated by all of its incoming edges.

  The Earth texture in the figure is used under CC BY 4.0 from https://www.solarsystemscope.com/textures/

GraphCast is a state-of-the-art machine learning based method for medium-range global weather forecasting. Developed by Lam et al. (2023), GraphCast leverages historical weather data to predict hundreds of weather variables for the next 10 days with high accuracy. Unlike traditional NWP methods, which rely on solving complex equations using supercomputers, GraphCast utilizes graph neural networks (GNNs) and modern deep learning hardware to achieve more efficient and accurate forecasts. This innovative approach significantly outperforms the most accurate operational deterministic systems, providing better predictions for severe weather events, including tropical cyclones, atmospheric rivers, and extreme temperatures.

GraphCast was trained using 39 years of historical weather data from ECMWF’s ERA5 reanalysis archive (Hersbach et al., 2020), covering the period from 1979 to 2017. This extensive dataset includes detailed observations and analyses of various weather variables, allowing GraphCast to learn complex patterns and dynamics of the atmosphere. The model's training objective was to minimize the mean squared error (MSE) between predicted and actual weather states over multiple autoregressive steps, ranging from 6 hours to 3 days. This incremental training process enabled GraphCast to develop robust forecasting capabilities for both short-term and medium-range weather predictions.

GraphCast employs a sophisticated neural network architecture based on graph neural networks (GNNs) in an "encoder-processor-decoder" configuration, comprising 36.7 million parameters. The model operates on a 0.25° latitude-longitude grid, with each grid point representing a set of surface and atmospheric variables (see Figure. 5A,5B,5C). The encoder maps these variables into a learned node representation on an internal "multimesh" graph, which is a spatially homogeneous graph derived from refining a regular icosahedron (see Figure. 5D). The processor component performs learned message-passing across this multimesh graph, enabling efficient local and long-range information propagation (see Figure. 5E,5G). Finally, the decoder maps the processed features back onto the latitude-longitude grid, predicting the next state of the weather as a residual update to the most recent input state (see Figure. 5F). This autoregressive approach allows GraphCast to generate extended weather forecasts by iteratively feeding its own predictions back into the model.

Figure. 6: Global skill and skill scores for GraphCast and HRES in 2018.

GraphCast has demonstrated exceptional performance in weather forecasting, significantly outperforming ECMWF’s HRES on key metrics such as RMSE and ACC. The model excelled in predicting the geopotential at 500 hPa (Z500), a crucial variable for understanding synoptic-scale weather patterns. Across 1380 evaluated variables and pressure levels, GraphCast achieved better skill scores than HRES on 90.3% of the targets (see Figure. 6). Moreover, GraphCast showed superior skill in forecasting severe weather events (see Figure. 7), including tropical cyclones, where it consistently provided more accurate track predictions compared to HRES. It also outperformed HRES in predicting atmospheric rivers and extreme temperatures, highlighting its robustness and applicability in critical weather forecasting scenarios.

Figure. 7: Severe event prediction.

• (A) Cyclone tracking performances for GraphCast and HRES. The x axis represents lead times (in days), and the y axis represents median track error (in kilometers). Error bars represent bootstrapped 95% confidence intervals for the median.
• (B) Cyclone tracking paired error difference between GraphCast and HRES. The x axis represents lead times (in days), and the y axis represents median paired error difference (in kilometers). Error bars represent bootstrapped 95% confidence intervals for the median difference (see supplementary materials section 8.1).
• (C) Atmospheric river prediction (IVT) skills for GraphCast and HRES. The x axis represents lead times (in days), and the y axis represents RMSE. Error bars are 95% confidence intervals.
• (D) Extreme heat prediction precision-recall for GraphCast and HRES. The x axis represents recall, and the y axis represents precision. The curves represent different precision-recall trade-offs when sweeping over gain applied to forecast signals

Despite its promising results, GraphCast has some limitations. One key challenge is its handling of uncertainty, as the model primarily produces deterministic forecasts. This approach does not capture the probability distributions of extreme events as effectively as ensemble forecasting systems like ECMWF’s ensemble forecast (ENS), which generate multiple stochastic forecasts to approximate future weather distributions. Additionally, GraphCast's reliance on high-quality historical data means its performance is closely tied to the fidelity of the reanalysis datasets used for training. The model also operates on a 0.25° resolution grid due to engineering constraints, whereas higher resolution data could potentially improve forecast accuracy further. Future improvements may involve integrating probabilistic forecasting methods to better capture uncertainties and extending the model's capabilities with more recent and higher resolution data.

The paper "Deep Learning for Day Forecasts from Sparse Observations" by Andrychowicz et al. (2023) presents MetNet-3, an advanced neural network-based weather forecasting model. Unlike traditional NWP models that rely on complex physical simulations, MetNet-3 leverages deep learning to predict weather conditions up to 24 hours ahead. The model processes both dense and sparse observational data, including radar, satellite, and weather station inputs, to generate high-resolution forecasts with minimal computational latency.

MetNet-3 was trained using a diverse set of observational data sources spanning from July 2017 to September 2022. The training data includes radar estimates of precipitation, weather station reports, satellite images, and assimilated weather states. This comprehensive dataset allowed MetNet-3 to learn from the highest fidelity and lowest latency data available, covering a broad range of atmospheric conditions. The model uses data from 942 weather stations across the Continental United States (CONUS) and integrates dense radar and satellite observations to enhance forecast accuracy.

Figure. 8: MetNet-3 network architecture.

MetNet-3's architecture (see Figure. 8) consists of three main components: topographical embeddings, a U-Net (Ronneberger et al., 2015) backbone, and a MaxVit (Tu et al., 2022) block. The model processes high-resolution, small-context inputs (4 km resolution) and low-resolution, large-context inputs (8 km resolution). The U-Net backbone handles local interactions, while the MaxVit block captures long-range dependencies. The network has 227 million parameters and utilizes innovative techniques like densification (see Figure. 9) to handle sparse observational data, ensuring accurate and dense forecasts. The model predicts weather variables at 4 km and 1 km resolutions, providing high temporal resolution forecasts every 2 to 5 minutes.

Figure. 9: Abstract depiction of densification aspects.

• (a) During training a fraction of the weather stations are masked out from the input, while kept in the target.
• (b) To evaluate generalization to untrained locations, a set of weather stations represented by squares is never trained on and only used for evaluation.
• (c) To evaluate forecasts for the sparse locations for which data is available, these stations are fed as input during the evaluation as well.
• (d) The final forecasts uses the full set of training weather stations as input, and produces fully dense forecasts aided by spatial parameter sharing.

MetNet-3 demonstrated superior performance compared to state-of-the-art NWP models, including the High Resolution Rapid Refresh (HRRR) and ENS from ECMWF. The model achieved higher Continuous Ranked Probability Scores (CRPS) and Critical Success Index (CSI) for predicting precipitation, temperature, dew point, and wind speed. MetNet-3 consistently outperformed ENS for lead times up to 24 hours, particularly excelling in forecasting severe weather events and providing high-resolution spatial forecasts with lower computational costs. For more details, see Figure. 10,11,12.

Figure. 10: Case study for Sat Apr 23 2022 12:00 UTC featuring the Rocky Mountains of Colorado showing the mean of the ENS and MetNet-3 6 hour wind speed forecasts (top, left and center) along with the One Minute Observations (OMO) stations ground truth (top, right) and the error of ENS and MetNet-3 on the individual weather stations (bottom). Circles and squares denote, respectively, training and test stations with MAEs calculated on both training and test stations. This example shows MetNet-3’s ability to densify the targets, the higher spatial resolution of MetNet-3 as well as forecast precision on the weather stations..Figure. 11: Case study for Thu Jan 17 2019 00:00 UTC showing the probability of instantaneous precipitation rate being above 1 mm/h on CONUS. The maps also show the prediction threshold when optimized towards CSI (dark blue contours) as well as the CSI values (lower left corners) calculated on the evaluation mask (Figure 2 in Supplement C). This specific case study shows the formation of a new large precipitation pattern in central US and not just extrapolation of existing patterns.Figure. 12: Case study for Thu Jun 10 2021 00:00 UTC comparing a MetNet-3 forecast and an ENS forecast for a single location (117.22°W, 33.91°N): Bold lines depict the means of the forecast distributions, and shaded areas correspond to 80% confidence interval based on the 10th- and 90th-quantile of the forecasted distribution in the case of MetNet-3 and the ensemble distribution in the case of ENS.

Despite its advancements, MetNet-3 has some limitations. The model relies on high-quality observational data, which may not always be available or consistent. It also produces deterministic forecasts, which might not capture the full range of uncertainties in weather predictions as effectively as ensemble-based NWP models. Additionally, the reliance on assimilated NWP initial states means that the model's performance could be impacted by any inaccuracies in these inputs. Future improvements could focus on enhancing the model's probabilistic forecasting capabilities and expanding the range of input data sources to further improve forecast accuracy and reliability.

The paper "PreDiff: Precipitation Nowcasting with Latent Diffusion Models" by Gao et al. (2023) introduces PreDiff, a new model designed for precipitation nowcasting. PreDiff uses a novel two-stage pipeline that combines conditional latent diffusion models (LDMs) (Rombach et al., 2022) with a knowledge alignment mechanism to improve forecast accuracy and reliability. This approach enables PreDiff to handle uncertainty better and generate predictions that align with domain-specific physical constraints, addressing common issues in data-driven Earth system forecasting.

PreDiff was trained using two datasets: the N-body MNIST dataset, a synthetic dataset with chaotic behavior, and the SEVIR dataset, a real-world precipitation nowcasting dataset. The N-body MNIST dataset contains sequences of moving digits influenced by gravitational forces, providing a controlled environment to test the model's ability to adhere to physical laws like energy conservation. The SEVIR dataset includes various weather observation data types, such as radar and satellite images, over a four-hour period, offering a comprehensive basis for training and evaluating the model's performance in real-world conditions.

Figure. 13: Overview of PreDiff inference with knowledge alignment. An observation sequence y is encoded into a latent context z_cond by the frame-wise encoder E. The latent diffusion model p_θ(z_t|z_t+1, z_cond), which is parameterized by an Earthformer-UNet, then generates the latent future z₀ by autoregressively denoising Gaussian noise z_T conditioned on z_cond. It takes the concatenation of the latent context z_cond (in the blue border) and the previous-step noisy latent future z_t+1 (in the cyan border) as input, and outputs z_t. The transition distribution of each step from z_t+1 to z_t can be further refined as p_{θ, φ}(z_t|z_t+1, y, F₀) via knowledge alignment, according to auxiliary prior knowledge. This denoising process iterates from t = T to t = 0, resulting in a denoised latent future z₀. Finally, z₀ is decoded back to pixel space by the frame-wise decoder D to produce the final prediction x̂. (Best viewed in color).

PreDiff's architecture consists of two main components: an Earthformer-UNet and a knowledge alignment network (see Figure. 13). The Earthformer-UNet is used as the backbone for the latent diffusion model, which operates in a lower-dimensional latent space to improve computational efficiency. The model follows a hierarchical UNet architecture with self-cuboid attention layers, allowing it to capture complex spatiotemporal dependencies in the data. The knowledge alignment network guides the diffusion process by incorporating domain-specific constraints, ensuring that the generated forecasts are physically plausible.

• Earthformer-UNet: This component processes the input data and generates latent space representations, leveraging self-cuboid attention blocks for detailed spatiotemporal modeling.
• Knowledge Alignment: This mechanism adjusts the transition probabilities during each denoising step of the diffusion process to ensure the forecasts comply with physical constraints, such as energy conservation or anticipated precipitation intensity.

PreDiff outperforms several state-of-the-art spatiotemporal forecasting models in both synthetic and real-world datasets (see Table. 1). In the N-body MNIST dataset, PreDiff with knowledge alignment (PreDiff-KA) significantly reduced energy errors compared to other models, demonstrating better adherence to the conservation of energy. On the SEVIR dataset (Veillette et al., 2020), PreDiff achieved higher scores on key metrics such as the CSI and FVD (Unterthiner et al., 2019), indicating superior forecast accuracy and visual quality. The knowledge alignment mechanism further enhanced the model's ability to generate high-fidelity predictions that align with domain-specific knowledge.

Despite its strengths, PreDiff has several limitations. The model's reliance on high-quality training data means that its performance can be affected by the availability and accuracy of observational data. Additionally, while the knowledge alignment mechanism helps ensure physical plausibility, the model's forecasts can sometimes deviate from real-world observations due to noise in the training data. Finally, integrating physical principles and domain knowledge into deep learning models remains an ongoing challenge, requiring close collaboration between AI researchers and domain experts to develop more robust and reliable forecasting systems.

The paper "Residual Diffusion Modeling for Km-scale Atmospheric Downscaling" by Mardani et al. (2023) introduces a cost-effective stochastic downscaling model named CorrDiff. This model is designed to predict high-resolution (2 km) weather data over Taiwan from coarser global inputs (25 km ERA5 reanalysis data). CorrDiff uses a two-step approach, combining a UNet prediction of the mean with a diffusion correction step. This technique enhances the ability to forecast weather phenomena such as intense rainfall and typhoons accurately.

CorrDiff was trained using high-resolution weather data from the Central Weather Administration (CWA) of Taiwan. This data includes radar-assimilating Weather Research and Forecasting (WRF) model outputs at a 2 km resolution, conditioned on 25 km ERA5 reanalysis data. The training data spans four years, from 2018 to 2021, with additional data from the first four months of 2022. The dataset consists of various meteorological variables, including temperature, wind speed, and radar reflectivity.

Figure. 14: The workflow for training and sampling CorrDiff for generative downscaling. Top: Coarse-resolution global weather data at 25 km scale is used to first predict the mean μ using a regression model, which is then stochastically corrected using EDM diffusion r, together producing the probabilistic high-resolution 2 km-scale regional forecast. Bottom right: diffusion model is conditioned with the coarse-resolution input to generate the residual r after a few denoising steps. Bottom left: the score function for diffusion is learned based on the UNet architecture.

CorrDiff employs a two-step model architecture (see Figure. 14):

• UNet Regression: The first step involves a UNet model to predict the mean atmospheric state from low-resolution input data.
• Diffusion Correction: The second step uses a diffusion model to correct the predicted mean, ensuring the final output adheres to realistic atmospheric conditions. This step employs the concept of score-based generative models to iteratively refine the predictions.

CorrDiff outperforms several baseline models in terms of both deterministic and probabilistic skill scores. The model demonstrates superior performance in terms of RMSE and CRPS across various atmospheric variables. CorrDiff also excels in capturing the power spectra and probability distributions of key weather phenomena, including coherent structures like frontal systems and typhoons (see Figure. 15,16,17).

Figure. 15: Examining the downscaling of a cold front on Feb 2, 2022 at 20 UTC. Left to right: prediction of ERA5, CorrDiff and Target for different fields, followed by their averaged cross section from 20 lines parallel to the thin dashed line in the contour figures. Top to bottom: 2 meter temperature (arrows are true wind vectors), along front wind (arrows are along front wind component) and across front wind (arrows are across front wind component).Figure. 16: A comparison of the 10m windspeed maps (m/s), distributions and the axisymmetric cross section from typhoon Chanthu (2021) on 2021/09/11:12:00:00UTC. Panels (a),(b),(c) show the 10m windspeed from ERA5, CorrDiff downscaling of ERA5 and the target (WRF), respectively. The CorrDiff panels show the first ensemble member. The solid black contour indicates the Taiwan coastline. Storm center of the ERA5, CorrDiff and WRF are shown in red ‘+‘, orange diamond, and the black dot, respectively. Panels (d) and (e) show the distribution shift (normalized PDFs) for the entire CWA domain and for the typhoon selected region in the top panels. Panel (f) shows the axisymmteric structure of the typhoon about its center. For the CorrDiff curves, line is the ensemble mean and the shading shows one standard deviation around the mean.Figure. 17: A comparison of a live km-scale forecast made by linking GFS predictions to CorrDiff donwscaling, validated against WRF-CWA operational predictions made prior to the typhoon Loinu landfall. Forecast is initialized on 20231004-12:00 UTC, top to bottom show lead time 0, 3, 6, 9, 12 hours. Left to right: forecast from the GFS; CorrDiff downscaling of GFS and the target (WRF-CWA). The fourth column compares the axisymmetric windspeed profile of the typhoon (shading for GFS+CorrDiff shows one standard deviation). The domain mean of the 1-hr maximum radar reflectivity is indicated at the top left of each of the reflectivity maps.

Despite its strong performance, CorrDiff has some limitations. The model relies on high-quality training data, which may not always be available or consistent. Additionally, while the diffusion step helps correct biases, it may not fully eliminate them, especially in cases with significant data variability. Further research is needed to optimize the model for different geographic regions and improve its ability to handle out-of-sample inputs.

In this course paper, we have explored five cutting-edge data-driven weather forecasting models: Pangu-Weather, GraphCast, MetNet-3, PreDiff, and CorrDiff. Each of these models represents a significant advancement in the integration of artificial intelligence and meteorological science, offering promising alternatives to traditional NWP methods.

• Pangu-Weather leverages 3D neural networks and Earth-specific priors to enhance medium-range weather forecasting accuracy and efficiency. Its hierarchical temporal aggregation strategy significantly reduces computational costs and error accumulation, making it a robust tool for predicting extreme weather events.

• GraphCast utilizes graph neural networks to provide high-accuracy, medium-range weather forecasts. By efficiently capturing both local and long-range dependencies within atmospheric data, GraphCast offers superior performance in forecasting severe weather phenomena such as tropical cyclones and atmospheric rivers, setting a new benchmark in the field.

• MetNet-3 employs a combination of U-Net and MaxVit transformer blocks to process dense and sparse observational data, enabling it to deliver high-resolution, short-term forecasts. Its ability to integrate various data sources and generate dense predictions with low computational latency makes it particularly effective for nowcasting applications.

• PreDiff introduces a novel approach to precipitation nowcasting using latent diffusion models. By incorporating a knowledge alignment mechanism, PreDiff ensures that its forecasts are both accurate and physically plausible. This model excels in handling uncertainty and generating high-fidelity predictions aligned with domain-specific constraints.

• CorrDiff offers a cost-effective solution for downscaling coarse global weather data to high-resolution local forecasts. Combining UNet regression with a diffusion correction step, CorrDiff accurately predicts weather phenomena at a fine scale, making it an invaluable tool for regions prone to intense weather events.

Despite their strengths, each model has its limitations. The reliance on high-quality training data, handling of uncertainty, and potential biases remain challenges that need further research and refinement. Future improvements could involve integrating probabilistic forecasting methods, expanding the range of input data sources, and optimizing models for different geographic regions.

In conclusion, the integration of AI into weather forecasting holds immense potential for improving the accuracy and efficiency of predictions. The advancements presented by Pangu-Weather, GraphCast, MetNet-3, PreDiff, and CorrDiff demonstrate that data-driven models can complement and, in some cases, surpass traditional NWP methods. Continued research and development in this field will undoubtedly lead to even more sophisticated and reliable forecasting tools, enhancing our ability to anticipate and respond to weather-related challenges.

Bauer, P., Thorpe, A., & Brunet, G. (2015). The quiet revolution of numerical weather prediction. Nature, 525(7567), 47–55. https://doi.org/10.1038/nature14956

Skamarock, W., Klemp, J., Dudhia, J., Gill, D., Barker, D., & Wang, W. (2005). A Description of the Advanced Research WRF Version 2 (p. 2500 KB) [Application/pdf]. UCAR/NCAR. https://doi.org/10.5065/D6DZ069T

Andrychowicz, M., Espeholt, L., Li, D., Merchant, S., Merose, A., Zyda, F., Agrawal, S., & Kalchbrenner, N. (2023). Deep Learning for Day Forecasts from Sparse Observations. https://doi.org/10.48550/ARXIV.2306.06079

Allen, M., Frame, D., Kettleborough, J., & Stainforth, D. (2006). Model error in weather and climate forecasting. In T. Palmer & R. Hagedorn (Eds.), Predictability of Weather and Climate (1st ed., pp. 391–427). Cambridge University Press. https://doi.org/10.1017/CBO9780511617652.016

Palmer, T. N., Shutts, G. J., Hagedorn, R., Doblas-Reyes, F. J., Jung, T., & Leutbecher, M. (2005). REPRESENTING MODEL UNCERTAINTY IN WEATHER AND CLIMATE PREDICTION. Annual Review of Earth and Planetary Sciences, 33(1), 163–193. https://doi.org/10.1146/annurev.earth.33.092203.122552

Bi, K., Xie, L., Zhang, H., Chen, X., Gu, X., & Tian, Q. (2023). Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619(7970), 533–538. https://doi.org/10.1038/s41586-023-06185-3

Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Stott, J., Pritzel, A., Mohamed, S., & Battaglia, P. (2023). Learning skillful medium-range global weather forecasting. Science, 382(6677), 1416–1421. https://doi.org/10.1126/science.adi2336

Gao, Z., Shi, X., Han, B., Wang, H., Jin, X., Maddix, D., Zhu, Y., Li, M., & Wang, Y. (Bernie). (2023). PreDiff: Precipitation Nowcasting with Latent Diffusion Models. Advances in Neural Information Processing Systems, 36, 78621–78656. https://proceedings.neurips.cc/paper_files/paper/2023/hash/f82ba6a6b981fbbecf5f2ee5de7db39c-Abstract-Conference.html

Mardani, M., Brenowitz, N., Cohen, Y., Pathak, J., Chen, C.-Y., Liu, C.-C., Vahdat, A., Kashinath, K., Kautz, J., & Pritchard, M. (2023). Residual Diffusion Modeling for Km-scale Atmospheric Downscaling (arXiv:2309.15214). https://doi.org/10.48550/arXiv.2309.15214

Bougeault, P., Tóth, Z., Bishop, C., Brown, B., Burridge, D., Chen, D., Ebert, B., Fuentes, M., Hamill, T., Mylne, K., Nicolau, J., Paccagnella, T., Park, Y.-Y., Parsons, D., Raoult, B., Schuster, D., Silva Dias, P., Swinbank, R., Takeuchi, Y., & Worley, S. (2010). The THORPEX interactive grand global ensemble. Bulletin of the American Meteorological Society, 91, 1059–1072. https://doi.org/10.1175/2010BAMS2853.1

Pathak, J., Subramanian, S., Harrington, P., Raja, S., Chattopadhyay, A., Mardani, M., Kurth, T., Hall, D., Li, Z., Azizzadenesheli, K., Hassanzadeh, P., Kashinath, K., & Anandkumar, A. (2022). FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators (arXiv:2202.11214). arXiv. https://doi.org/10.48550/arXiv.2202.11214

Kenneth, R., Howard, J., James, P., Michael, C., & Carl, J. (2019). International Best Track Archive for Climate Stewardship (IBTrACS) Project, Version 4 [dataset]. NOAA National Centers for Environmental Information. https://doi.org/10.25921/82TY-9E16

Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D., Simmons, A., Soci, C., Abdalla, S., Abellan, X., Balsamo, G., Bechtold, P., Biavati, G., Bidlot, J., Bonavita, M., … Thépaut, J.-N. (2020). The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730), 1999–2049. https://doi.org/10.1002/qj.3803

Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. In N. Navab, J. Hornegger, W. M. Wells, & A. F. Frangi (Eds.), International Conference on Medical Image Computing and Computer-Assisted Intervention (Vol. 9351, pp. 234–241). Springer International Publishing. https://doi.org/10.1007/978-3-319-24574-4_28

Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., & Li, Y. (2022). MaxViT: Multi-axis Vision Transformer. In S. Avidan, G. Brostow, M. Cissé, G. M. Farinella, & T. Hassner (Eds.), Computer Vision – ECCV 2022 (pp. 459–479). Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-20053-3_27

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022). High-Resolution Image Synthesis with Latent Diffusion Models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 10674–10685. https://doi.org/10.1109/CVPR52688.2022.01042

Veillette, M., Samsi, S., & Mattioli, C. (2020). SEVIR: A Storm Event Imagery Dataset for Deep Learning Applications in Radar and Satellite Meteorology. Advances in Neural Information Processing Systems, 33, 22009–22019. https://proceedings.neurips.cc/paper/2020/hash/fa78a16157fed00d7a80515818432169-Abstract.html

Wang, Y., Wu, H., Zhang, J., Gao, Z., Wang, J., Yu, P. S., & Long, M. (2023). PredRNN: A Recurrent Neural Network for Spatiotemporal Predictive Learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2), 2208–2225. https://doi.org/10.1109/TPAMI.2022.3165153

Guen, V. L., & Thome, N. (2020). Disentangling Physical Dynamics From Unknown Factors for Unsupervised Video Prediction. 11474–11484. https://openaccess.thecvf.com/content_CVPR_2020/html/Le_Guen_Disentangling_Physical_Dynamics_From_Unknown_Factors_for_Unsupervised_Video_Prediction_CVPR_2020_paper.html

Wang, Y., Jiang, L., Yang, M.-H., Li, L.-J., Long, M., & Fei-Fei, L. (2018, September 27). Eidetic 3D LSTM: A Model for Video Prediction and Beyond. International Conference on Learning Representations. https://openreview.net/forum?id=B1lKS2AqtX

Bai, C., Sun, F., Zhang, J., Song, Y., & Chen, S. (2022). Rainformer: Features Extraction Balanced Network for Radar-Based Precipitation Nowcasting. IEEE Geoscience and Remote Sensing Letters, 19, 1–5. https://doi.org/10.1109/LGRS.2022.3162882

Gao, Z., Shi, X., Wang, H., Zhu, Y., Wang, Y. (Bernie), Li, M., & Yeung, D.-Y. (2022). Earthformer: Exploring Space-Time Transformers for Earth System Forecasting. Advances in Neural Information Processing Systems, 35, 25390–25403. https://proceedings.neurips.cc/paper_files/paper/2022/hash/a2affd71d15e8fedffe18d0219f4837a-Abstract-Conference.html

Yan, W., Zhang, Y., Abbeel, P., & Srinivas, A. (2021). VideoGPT: Video Generation using VQ-VAE and Transformers (arXiv:2104.10157). arXiv. https://doi.org/10.48550/arXiv.2104.10157

Mardani, M., Brenowitz, N., Cohen, Y., Pathak, J., Chen, C.-Y., Liu, C.-C., Vahdat, A., Kashinath, K., Kautz, J., & Pritchard, M. (2023). Residual Diffusion Modeling for Km-scale Atmospheric Downscaling (arXiv:2309.15214). https://doi.org/10.48550/arXiv.2309.15214

NVIDIA (Director). (2024, March 19). GTC March 2024 Keynote with NVIDIA CEO Jensen Huang. https://www.youtube.com/watch?v=Y2F8yisiS6E

NVIDIA Earth 2 Platform. (n.d.). NVIDIA. Retrieved 3 June 2024, from https://www.nvidia.com/en-us/high-performance-computing/earth-2/

NVIDIA/earth2mip. (2024). [Python]. NVIDIA Corporation. https://github.com/NVIDIA/earth2mip (Original work published 2023)