PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning and Foundation Models for Weather Prediction: A Survey

Jimeng Shi, Azam Shirali, Bowen Jin, Sizhe Zhou, Wei Hu, Rahuul Rangaraj, Shaowen Wang, Jiawei Han, Zhaonan Wang, Upmanu Lall, Yanzhao Wu, Leonardo Bobadilla, Giri Narasimhan

arXiv:2501.06907v1cs.LG

TL;DR

Weather prediction needs methods that complement computationally intensive physics-based models and address their simplified assumptions and incomplete uncertainty representation. This survey organizes deep learning and foundation models by training paradigm, reviews their architectures and applications, and compiles datasets, code, and research directions. Across paradigms, deterministic, probabilistic, and foundation models offer strong performance or adaptability, while generalization to rare extremes and data quality remain important boundaries.

  • Problem

    Physics-based weather models are computationally intensive, rely on simplified atmospheric assumptions, and do not explicitly capture model uncertainty, motivating improved data-driven approaches.

    Source: §1 · p. 1

  • Method

    The survey categorizes weather models into predictive learning, generative learning, and pre-training followed by fine-tuning, while reviewing architectures, applications, resources, and future directions.

    Source: §1 · p. 2, §7 · p. 14

  • Results

    GenCast reported greater skill than IFS ENS on 97.4% of 1320 evaluated targets, while NeuralGCM outperformed other state-of-the-art deep learning models and was comparable with ECMWF’s IFS for geopotential, temperature, and wind.

    Source: §6 · p. 14

  • Takeaways & Limitations

    Deterministic, probabilistic, and foundation models provide complementary strengths for prediction accuracy, uncertainty modeling, and adaptation across weather tasks.

    Source: §6 · p. 14

    Deep learning models often perform poorly on rare extreme or anomalous events outside the training distribution, and reanalysis-data quality remains a concern.

    Source: §5.1 · p. 11, §5.5 · p. 13

Abstract

from arXiv · show

Physics-based numerical models have been the bedrock of atmospheric sciences for decades, offering robust solutions but often at the cost of significant computational resources. Deep learning (DL) models have emerged as powerful tools in meteorology, capable of analyzing complex weather and climate data by learning intricate dependencies and providing rapid predictions once trained. While these models demonstrate promising performance in weather prediction, often surpassing traditional physics-based methods, they still face critical challenges. This paper presents a comprehensive survey of recent deep learning and foundation models for weather prediction. We propose a taxonomy to classify existing models based on their training paradigms: deterministic predictive learning, probabilistic generative learning, and pre-training and fine-tuning. For each paradigm, we delve into the underlying model architectures, address major challenges, offer key insights, and propose targeted directions for future research. Furthermore, we explore real-world applications of these methods and provide a curated summary of open-source code repositories and widely used datasets, aiming to bridge research advancements with practical implementations while fostering open and trustworthy scientific practices in adopting cutting-edge artificial intelligence for weather prediction. The related sources are available at https://github.com/JimengShi/ DL-Foundation-Models-Weather.

1 Introduction

Weather prediction is increasingly important as climate change intensifies extreme events, while existing physics-based and statistical approaches face computational, representational, and uncertainty limitations. This survey organizes deep learning weather models by training paradigm and provides resources and future directions for research and practice.

Limitations of Existing Methods

  • Physics-based models are computationally intensive, rely on simplified atmospheric assumptions, and typically produce deterministic forecasts that do not explicitly capture model uncertainty.Their governing equations are high-dimensional and nonlinear, and perturbing initial conditions represents input uncertainty rather than the full uncertainty in weather evolution.Source: §1 · pp. 1–2
  • Statistical methods such as ARIMA struggle with nonlinear relationships and are usually restricted to univariate or low-dimensional weather responses.ARIMA also requires careful parameter selection and is sensitive to outliers.Source: §1 · p. 2

Motivation

  • Deep learning models have achieved state-of-the-art medium-range global prediction while reducing computational costs by up to three orders of magnitude.Pangu and GraphCast match or surpass traditional methods in accuracy, although point-wise training can produce blurry predictions.Source: §1 · p. 2

Emerging Model Paradigms

  • Probabilistic generative models use diffusion-based sampling to provide high-quality predictions with calibrated uncertainty estimates.CasCast and GenCast apply this approach to precipitation nowcasting and weather prediction.Source: §1 · p. 2
  • Foundation models are pre-trained on massive historical weather datasets and fine-tuned for diverse downstream applications.Their stated advantages are transferable weather representations and adaptation without training task-specific models from scratch.Source: §1 · p. 2

Contributions

  • The survey introduces a taxonomy based on predictive learning, generative learning, and pre-training followed by fine-tuning.It distinguishes deterministic predictive models, probabilistic generative models, and foundation-model adaptation as organizing paradigms.Source: §1 · p. 2
  • The survey compiles benchmark datasets, open-source code, real-world applications, and ten future research directions across five avenues.These resources are intended to connect research advances with practical implementations and support open scientific practice.Source: §1 · p. 2

2 Background and Preliminaries

Weather prediction uses complementary station observations and gridded reanalysis data, and can be formulated across temporal, spatial, application, and event dimensions. The preliminaries distinguish general-purpose, domain-specific, and foundation models while linking them to deterministic, probabilistic, and pre-training-based paradigms.

2.1 Weather Data Representation

  • Station observations provide high-resolution local measurements but have uneven coverage, especially across oceans, mountains, and deserts.They support localized forecasting, trend analysis, and model validation despite limited global representativeness.Source: §2.1 · p. 3
  • Gridded reanalysis combines observations, satellite measurements, and NWP models to provide consistent global spatial coverage at varying resolutions.Common grid sizes range from 1°×1° to 0.25°×0.25°, with hourly or daily temporal intervals.Source: §2.1 · p. 3

2.2 Weather Prediction Formulation

  • Weather forecasting spans temporal lead times from hours to years, spatially targets global or regional domains, predicts selected applications, and includes extreme or regular events.The temporal category includes nowcasting, medium-range, sub-seasonal, and seasonal forecasting.Source: §2.2 · p. 3
  • The forecasting formulation maps input variables X over an input window α to output variables Y over an output window β using a parameterized model F(θ).F(·) may also represent a probabilistic function, written as Y ∼ P(Y | X).Source: §2.2 · p. 3

2.3 Preliminaries

  • General-purpose large models use diverse global datasets and multiple meteorological variables, whereas domain-specific models focus on a single variable and regional prediction.These definitions distinguish broad application scope from specialized regional modeling.Source: §2.3 · p. 3
  • Foundation models are large models pre-trained on diverse massive datasets and subsequently fine-tuned or adapted for downstream tasks.General-purpose and domain-specific models may use deterministic or probabilistic learning, while foundation models follow pre-training and fine-tuning.Source: §2.3 · p. 3

3 Overview and Taxonomy

The survey organizes weather-prediction models by training paradigm, model scope, and architecture, covering predictive, generative, and foundation-model approaches. It also distinguishes general-purpose models from domain-specific systems and compares their architectures, applications, and limitations.

3 Overview and Taxonomy

  • The taxonomy classifies models by deterministic predictive learning, probabilistic generative learning, and pre-training followed by fine-tuning.Source: §3 · pp. 3–4

3.1 Predictive Learning

  • Predictive models estimate future weather states from observations by minimizing point-wise losses, spanning general-purpose large models and domain-specific models.Source: §3.1 · p. 4

3.1.1 General-Purpose Large Models

  • General-purpose architectures include Transformers, GNNs, and physics-AI hybrids that model global variables, spatial dependencies, or physically constrained dynamics.Source: §3.1.1 · pp. 4–6
  • Iterative forecasting can accumulate errors over longer prediction windows, motivating hierarchical, multi-stage, and time-conditioned designs such as Pangu-Weather, FuXi, and Stormer.Source: §3.1.1 · pp. 4–5

3.1.2 Domain-Specific Models

  • Domain-specific predictive models target regional or single-variable forecasts using Transformer, GNN, RNN/CNN, Mamba, and physics-AI architectures.Source: §3.1.2 · pp. 6–7

3.2 Generative Models

  • Probabilistic generative models produce ensemble forecasts and represent uncertainty; diffusion approaches include global systems such as GenCast and domain-specific precipitation models.Source: §3.2 · p. 7, §3.2.1 · p. 7, §3.2.2 · p. 8

3.3 Foundation Models

  • Foundation models are pre-trained on broad weather or climate data and fine-tuned for downstream tasks, including forecasting, nowcasting, and downscaling.Source: §3.3 · p. 8

4 Applications and Resources

The survey reviews deep-learning applications across precipitation, air quality, sea-surface temperature, floods, droughts, tropical cyclones, and wildfire. Across these domains, models combine convolutional, recurrent, graph, Transformer, generative, and physics-guided approaches for prediction and risk assessment.

4 Applications and Resources

  • Deep-learning weather applications span precipitation, air quality, sea-surface temperature, floods, droughts, tropical cyclones, and wildfire.Source: §4 · p. 9, §4.1 · p. 9, §4.2 · p. 9

4.1 Precipitation

  • Precipitation prediction uses CNNs, RNNs, Transformers, hybrid architectures, GANs, and diffusion models for nowcasting and synthesis.Source: §4.1 · p. 9

4.2 Air Quality

  • Air-quality forecasting applies ANN, CNN-Bi-LSTM, graph neural networks, spatiotemporal models, and Transformers across cities and countries.Source: §4.2 · p. 9

4.3 Sea Surface Temperature

  • Sea-surface-temperature and ENSO prediction uses recurrent, convolutional, graph, residual, and Transformer models, including approaches extending lead times to 23 months.Source: §4.3 · p. 9

4.4 Flood

  • Flood forecasting combines CNN, RNN, ConvLSTM, graph, Transformer, diffusion, GAN, and physics-guided models for prediction, mapping, and risk assessment.Source: §4.4 · pp. 9–10

4.5 Drought

  • Drought studies predict meteorological and soil-moisture indices with LSTM, CNN, hybrid, Transformer, and GAN architectures across multiple temporal scales.Source: §4.5 · p. 10

4.6 Tropical Storms/Cyclones and Hurricanes

  • Tropical-cyclone models address formation, intensity, track, rainfall, trajectories, and storm surge using CNN, hybrid, GAN, diffusion, and graph-recurrent architectures.Source: §4.6 · p. 10

4.7 Wildfire

  • Wildfire applications model spread, risk, fire weather, and landscape heterogeneity with CNN, recurrent, hybrid, ConvLSTM, GAN, GNN, and Transformer methods.Source: §4.7 · pp. 10–11

5 Challenges and Future Directions

The survey identifies challenges spanning trustworthy AI, foundation-model augmentation, constrained generation, multimodal integration, and data processing, and proposes targeted research directions across these areas.

5 Challenges and Future Directions

  • Ten research directions are organized across model- and data-focused challenges, covering trustworthy AI, retrieval augmentation, constrained generation, multimodal learning, and data management.The section explicitly frames its roadmap around DL-model and data perspectives.Source: §5 · p. 11

5.1 Trustworthy AI

  • Rare extreme or anomalous events remain out of distribution for AI models, motivating physics-informed methods and diffusion-based augmentation with additional extreme samples.The proposed approaches target reliability and generalization under sparse representation of extreme events.Source: §5.1 · p. 11
  • Diffusion models can represent both aleatoric and epistemic uncertainty through stochastic sampling and varied initial noise, while biased observations require correction, assimilation, or adversarial training.These directions address uncertainty quantification and robustness to collection, spatial, temporal, and sensor biases.Source: §5.1 · pp. 11–12

5.2 Retrieval-augmented Foundation Models

  • Retrieval augmentation could supply similar historical weather patterns to diffusion models and relevant documents to weather-text LLMs, improving contextual information for prediction and analysis.The proposed uses are especially relevant to rare extreme-weather cases and context-rich textual tasks.Source: §5.2 · p. 12

5.3 Generative AI with Weather Constraints

  • Weather generation can be conditioned on physical equations, geographic laws, and heterogeneous modalities, although diffusion models’ computational overhead makes training and inference acceleration important.These constraints are proposed as prior knowledge for more robust and precise weather representations.Source: §5.3 · p. 13

5.4 Multi-Modal Learning

  • Multimodal learning can combine observations, reanalysis, imagery, and text, but aligning numerical data with textual descriptions remains a central challenge.Knowledge graphs built from environmental news are proposed as one way to link weather events with meteorological raster data.Source: §5.4 · p. 13

5.5 Data Processing and Management

  • Variational autoencoders can compress massive weather datasets while preserving comparable downstream forecasting accuracy, reducing storage requirements substantially.Reported examples reduce 8.61 TB to 204 GB and 226 TB to 0.7 TB.Source: §5.5 · p. 13
  • Reanalysis quality is constrained by empirical assumptions, motivating data assimilation that calibrates model outputs with observations from stations, satellites, and other sources.Examples include conditional diffusion, Swin Transformer, and masked-autoencoder approaches for assimilating heterogeneous Earth observations.Source: §5.5 · p. 13

6 Discussion

The survey compares three training paradigms and finds distinct strengths, trade-offs, and unresolved challenges across deterministic, probabilistic, and foundation models.

Deterministic Predictive Models

  • NeuralGCM outperforms other state-of-the-art deep-learning models and is comparable with ECMWF’s IFS for geopotential, temperature, and wind.GraphCast, Pangu, and Fuxi are competitive with or better than ECMWF HRES.Source: §6 · p. 14
  • Deterministic models suit different forecast ranges and architectures, but MSE training can produce blurry outputs, especially for extreme events.Transformer models support temporal prediction, GNNs model spatial relations, and hybrid models capture spatiotemporal dependencies but may take longer to train.Source: §6 · p. 14

Probabilistic generative models

  • GenCast, CoDiCast, and CasCast model aleatoric and epistemic uncertainty through probabilistic noise sampling, supporting extreme-weather decision-making.GenCast reported greater skill than IFS ENS on 97.4% of 1320 evaluated targets.Source: §6 · p. 14
  • Probabilistic generative models require more training and inference resources than deterministic models, although they remain faster than physics-based models.Source: §6 · p. 14

Foundation models

  • Foundation models such as Aurora, ClimaX, and Prithvi WxC provide adaptable transfer learning across diverse tasks after fine-tuning.Current foundation models mainly use deterministic predictive learning during pre-training, and no probabilistic generative architectures were identified.Source: §6 · p. 14
  • Foundation models face barriers from large parameter counts and pre-training requirements, while weather-forecast fine-tuning remains an early-stage area.The survey points to natural-language-processing research as a potential source of fine-tuning insights.Source: §6 · p. 14

Comparison

  • The comparison evaluates predictive learning, generative learning, and pre-training and fine-tuning models for global medium-range weather prediction.Source: Table 2 · p. 14

7 Conclusions

The survey synthesizes deep-learning and foundation models for weather prediction, organizes them by training paradigm, and connects the review to datasets, codebases, applications, and future research directions.

7 Conclusions

  • The paper presents a comprehensive, up-to-date survey of data-driven deep-learning and foundation models for weather prediction.Source: §7 · p. 14
  • Its taxonomy categorizes models by training paradigms and supports in-depth review, analysis, and comparison of methodologies within each category.Source: §7 · p. 14
  • The survey summarizes datasets, open-source codebases, and real-world applications through a related GitHub repository.Source: §7 · p. 14
  • It outlines ten critical research directions across five avenues for advancing AI-driven weather prediction.Source: §7 · p. 14
  • The survey focuses on weather prediction and excludes climate downscaling, climate emulation, and climate trend prediction because of limited space.Source: §7 · p. 14

A Datasets

The dataset overview summarizes publicly available weather datasets by their domains, names, coverage, collection methods, resolutions, time spans, and introducing papers.

A Datasets

  • Each dataset is described by domain, name, coverage, collection method, spatial and temporal resolution, time span, and introducing paper.Source: A Datasets · p. 31
  • Table 3 summarizes publicly available weather datasets and identifies CAM5 as Community Atmospheric Model v5.Source: Table 3 · p. 31

B.1 Convolutional Neural Networks

Convolutional neural networks process structured grid data with convolutional kernels, often followed by pooling, and are used across several vision-style tasks.

B.1 Convolutional Neural Networks

  • CNNs process structured grid data using convolutional kernels that extract features such as edges, textures, and patterns.Source: §B.1 · pp. 31–32
  • Pooling can reduce feature-map dimensions, improve computational efficiency, and focus processing on important information.Source: §B.1 · p. 32
  • CNNs are used for image classification, object detection, and segmentation, with Conv1D, Conv2D, and Conv3D variants distinguished by kernel-sliding dimension.Source: §B.1 · p. 32

B.2 Recurrent Neural Networks

RNNs model sequential weather data by updating an internal hidden state from current inputs and prior states, but long sequences can cause gradient instability.

B.2 Recurrent Neural Networks

  • RNNs recurrently learn from current inputs and previous hidden states, making them suited to time-dependent and sequential data.Their applications include time-series forecasting, natural-language processing, and speech recognition.Source: §B.2 · p. 32
  • The update rule produces hidden states and outputs using weight matrices, bias terms, and an activation function such as tanh or ReLU.The hidden state summarizes information at each time step for subsequent updates and outputs.Source: §B.2 · p. 32
  • RNNs often suffer from vanishing or exploding gradients when modeling long sequences, motivating gated LSTM and GRU variants.These architectures use gates to forget and filter information.Source: §B.2 · p. 32

B.3 Graph Neural Networks

GNNs represent weather domains as graphs whose nodes encode spatial locations and whose edges encode relationships, while ST-GNNs model evolving spatial and temporal dependencies through message passing.

B.3 Graph Neural Networks

  • GNNs operate on graph-structured data with nodes representing entities and edges representing dependencies; ST-GNNs extend this representation across time.In weather applications, nodes correspond to spatial locations and edges to spatial relationships.Source: §B.3 · p. 32
  • Message passing updates each node’s hidden state using feature messages from neighboring nodes to capture spatial dependencies.The neighborhood N(i) specifies which nodes contribute to node i’s update.Source: §B.3 · p. 32
  • After k iterative updates, a readout function combines node hidden states to produce the graph-level forecast at time t+k.The output is computed from the updated states across all graph nodes.Source: §B.3 · p. 32

B.4 Transformer and Vision Transformer

This section surveys Transformer, Mamba, GAN, and diffusion architectures used to model weather data, covering attention, state-space sequence modeling, adversarial generation, and iterative denoising.

B.4 Transformer and Vision Transformer

  • Transformers replace recurrent sequential processing with parallel attention, allowing dependencies between any sequence elements to be captured.Attention uses query, key, and value representations derived by linear transformations of the input sequence.Source: §B.4 · pp. 32–33
  • Vision Transformers divide images into non-overlapping patches, add positional embeddings, and process the resulting patch tokens with Transformer layers.This preserves spatial layout while applying sequence-based attention to image data.Source: §B.4 · p. 33

B.5 Mamba and Vision Mamba

  • State Space Models map input sequences to outputs through latent states, while Mamba makes sequence interactions input-dependent through selective parameters and hardware-aware acceleration.Mamba’s selection mechanism uses input-dependent ∆, A, and B parameters to capture contextual information in long sequences.Source: §B.5 · p. 33
  • Vision Mamba splits images into patch tokens and uses bidirectional Mamba blocks instead of attention, making it suitable for two-dimensional grid weather data.The survey identifies MetMamba as an example of this adaptation.Source: §B.5 · p. 33

B.6 Generative Adversarial Networks

  • GANs generate weather samples through adversarial training between a generator that produces samples and a discriminator that distinguishes generated from real data.The approach has been used for image-based weather forecasting.Source: §B.6 · p. 34

B.7 Diffusion Models

  • Diffusion models learn generation through fixed forward noising followed by learned reverse denoising that reconstructs realistic samples from Gaussian noise.Weather applications include precipitation nowcasting, atmospheric downscaling, and weather forecasting.Source: §B.7 · p. 34
Loading 2501.06907v1…