PaperlayerView exact revision on arXiv

Source-linked AI summary

A Foundation Model for the Earth System

Cristian Bodnar, Wessel P. Bruinsma, Ana Lucic, Megan Stanley, Anna Vaughan, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan A. Weyn, Haiyu Dong, Jayesh K. Gupta, Kit Thambiratnam, Alexander T. Archibald, Chun-Chieh Wu, Elizabeth Heider, Max Welling, Richard E. Turner, Paris Perdikaris

arXiv:2405.13063v3physics.ao-phcs.LG

TL;DR

Earth system forecasting needs accurate predictions, but existing systems are computationally demanding and AI coverage remains limited beyond weather. Aurora is a large-scale foundation model pretrained on diverse Earth system data and fine-tuned across forecasting domains, where it outperforms multiple operational systems at much lower computational cost. Its flexible adaptation supports broader Earth system prediction, while it still relies on traditional data-assimilation initial conditions.

  • Problem

    Existing Earth system forecasting systems are computationally demanding, while machine-learning advances have left several important domains underexplored.

    Source: §1 · p. 1, §6 · p. 8

  • Method

    Aurora is a foundation model pretrained on diverse Earth system data and fine-tuned for multiple forecasting tasks using a flexible 3D architecture.

    Source: §1 · p. 2, §2 · p. 3, §8.1 · p. 13

  • Results

    Aurora achieves state-of-the-art results across air quality, ocean waves, tropical cyclone tracks, and high-resolution weather forecasting while outperforming multiple operational systems at a fraction of the computational cost.

    Source: §1 · pp. 2–3, §7 · p. 9

  • Takeaways & Limitations

    Aurora can be fine-tuned for diverse Earth system applications at modest computational cost, making actionable predictions more accessible.

    Source: §7 · p. 9

    Aurora still relies on initial conditions from traditional data-assimilation systems.

    Source: §7 · p. 9

Abstract

from arXiv · show

Reliable forecasts of the Earth system are crucial for human progress and safety from natural disasters. Artificial intelligence offers substantial potential to improve prediction accuracy and computational efficiency in this field, however this remains underexplored in many domains. Here we introduce Aurora, a large-scale foundation model for the Earth system trained on over a million hours of diverse data. Aurora outperforms operational forecasts for air quality, ocean waves, tropical cyclone tracks, and high-resolution weather forecasting at orders of magnitude smaller computational expense than dedicated existing systems. With the ability to fine-tune Aurora to diverse application domains at only modest computational cost, Aurora represents significant progress in making actionable Earth system predictions accessible to anyone.

1 Introduction

Earth system forecasts are vital but existing systems are computationally demanding, difficult to maintain, and limited by physical approximations. Aurora addresses these challenges as a pretrained foundation model that achieves state-of-the-art results across several forecasting domains.

1 Introduction

  • Existing Earth system models require substantial computational resources, engineering effort, and approximations that can limit accuracy.These limitations motivate alternative forecasting approaches.Source: §1 · p. 2
  • Recent AI forecasting advances have focused mainly on 0.25° global medium-range weather, leaving ocean dynamics, wave modelling, atmospheric chemistry, and extreme-weather prediction less explored.The cited gap concerns applying machine learning beyond the domains where recent advances have concentrated.Source: §1 · p. 2
  • Aurora is pretrained on over one million hours of Earth system data and fine-tuned for diverse forecasting tasks.Its reported applications include air pollution, ocean waves, tropical cyclone tracks, and high-resolution weather.Source: §1 · pp. 2–3
  • Aurora is introduced to provide one foundation model capable of addressing multiple Earth system forecasting tasks.The paper presents this as an AI model that can outperform multiple existing operational systems while being orders of magnitude faster.Source: §1 · p. 2
  • 74% of air-pollution targets, 86% of ocean-wave targets, 100% of tropical-cyclone targets, and 92% of high-resolution weather targets are outperformed by Aurora’s reported comparisons.The corresponding domains use 0.4°, 0.25°, 0.25°, and 0.1° resolutions, respectively.Source: §1 · pp. 2–3

2 Aurora: a flexible 3D foundation model for the Earth system

Aurora uses a flexible three-part architecture to map heterogeneous Earth system inputs into a universal 3D representation, evolve it, and decode predictions. Pretraining learns general-purpose dynamics, while fine-tuning adapts the model to new tasks and improves performance with scale.

2 Aurora: a flexible 3D foundation model for the Earth system

  • Aurora encodes heterogeneous inputs into a universal 3D representation, evolves it with a 3D Swin Transformer, and decodes task-specific predictions.Perceiver-based encoder and decoder modules support different variables and resolutions, while forecasts are generated recursively.Source: §2 · p. 3
  • Pretraining combines forecasts, analyses, reanalyses, and climate simulations to learn general-purpose atmospheric and oceanic dynamics.Fine-tuning then adapts these learned representations to new tasks, datasets, and variables using less data and computation.Source: §2 · p. 3
  • Pretraining on more diverse data systematically improves validation performance, especially for extreme values, while validation performance rises approximately 6% for every 10× increase in model size.These scaling results support jointly increasing data diversity and model size.Source: §2 · p. 3
  • Aurora outperforms IFS and GraphCast on over 91% of targets for 0.25° medium-range weather forecasting.This comparison measures the benefits of Aurora’s data and model scaling against numerical and AI baselines.Source: §2 · p. 3

3 Modelling atmospheric chemistry and air quality

Global atmospheric-composition forecasting is difficult because pollution is heterogeneous, chemically complex, and costly to simulate. Aurora is fine-tuned on CAMS-related data and matches or outperforms CAMS on most evaluated targets while using far less computation.

3 Modelling atmospheric chemistry and air quality

  • Air-quality forecasting requires modelling hundreds of stiff chemical equations and heterogeneous pollution driven by anthropogenic emissions.CAMS extends IFS with aerosol, reactive-gas, and greenhouse-gas modules, increasing computational costs by approximately tenfold.Source: §3 · p. 3
  • Aurora is fine-tuned using CAMS analysis data from 2017–2022 together with lower-resolution CAMS reanalysis data extending back to 2003.The reanalysis is explicitly considered lower quality because it uses lower resolution and an older underlying model.Source: §3 · p. 4
  • Aurora matches or outperforms CAMS on 74% of all targets and 89% of all variables at a three-day lead time.It is within 20% RMSE of CAMS on 95% of all targets and on all variables at three days.Source: Figure 2 · p. 4, §3 · p. 5
  • Aurora produces atmospheric-composition predictions in approximately 1.1 s per hour lead time on one A100 GPU, about 50,000× faster than CAMS.CAMS outperforms Aurora for upper-atmosphere ozone and twelve-hour predictions of all species in the lower atmosphere.Source: §3 · p. 5
  • Aurora predicts a severe Iraqi sandstorm one day in advance with similar accuracy to CAMS at a fraction of the computational cost.The case study supports generalisation to an extreme event involving interactions between atmospheric dynamics and pollutants.Source: §3 · p. 5

4 Modeling ocean wave dynamics

Global wave forecasting must represent directional, spectral, and intermittently missing variables over a short observational record. After fine-tuning on HRES-WAM data, Aurora matches or outperforms HRES-WAM on most wave comparisons and captures a major typhoon’s wave field.

4 Modeling ocean wave dynamics

  • Wave forecasting involves directional, temporal, and spectral variables that can be undefined over land, sea ice, or arbitrary spatial locations.Aurora was extended to support missing data for these conditions.Source: §4 · p. 5
  • Aurora is fine-tuned jointly on wave and meteorological variables from HRES-WAM analysis and HRES T0, regridded to 0.25°.Training uses 2016–2021 data and evaluation uses 2022.Source: §4 · p. 5
  • Aurora matches or outperforms HRES-WAM on 86% of all wave variables and 91% of variables at a three-day lead time.It is within 20% RMSE on 96% of all targets and on all but PP1D at three days.Source: §4 · p. 5, Figure 3 · p. 6
  • Aurora accurately predicts significant wave height and mean wave direction during Typhoon Nanmadol, including the large waves associated with the storm.The predicted fields follow prevailing global wind patterns in the case study.Source: §4 · p. 5, Figure 3 · p. 6

5 Predicting tropical cyclone tracks

Aurora is evaluated against operational tropical-cyclone forecasting systems across global basins. It surpasses official forecasts and headline guidance models through five-day lead times, including high-impact cases.

Forecasting context

  • Official tropical-cyclone forecasts combine multiple dynamical and statistical models, consensus products, and human forecaster analysis.These systems include global ensembles and purpose-built, vortex-following high-resolution models.Source: §5 · p. 6
  • Earlier AI comparisons showed marginal gains at short lead times, while other large-scale models lagged official operational forecasts.Source: §5 · pp. 6–7

Operational comparison

  • Aurora outperforms official tropical-cyclone track forecasts across four basins and all evaluated lead times, surpassing full operational forecasts through five days.Improvements in the North Atlantic and Eastern Pacific reach 6% at one day and 20–25% at two to five days.Source: §5 · p. 7
  • Aurora outperforms all headline models used by human forecasters in North Atlantic and Eastern Pacific verification.This comparison addresses the broader guidance set rather than only the final official forecast.Source: §5 · p. 7

High-impact case

  • Aurora correctly predicts Typhoon Doksuri’s Philippine landfall four days ahead, unlike the official prediction placing the vortex off Northern Taiwan.Source: §5 · p. 7, Figure 4 · p. 7

6 High-resolution operational weather forecasting

Aurora is adapted to 0.1° operational weather forecasting, targeting the fine-scale processes needed for severe-weather prediction. It generally outperforms IFS HRES and uniquely captures an abrupt extreme-wind increase in Storm Ciarán.

High-resolution setup

  • Aurora is fine-tuned at 0.1° to resolve smaller-scale convective and boundary-layer processes that current AI models commonly evaluate at 0.25°.The 0.1° data record is shorter, extending only to 2016.Source: §6 · p. 8
  • Evaluation initializes Aurora with IFS HRES analysis and compares against HRES analysis, while scoring IFS HRES from its zero-hour forecast.Source: §6 · p. 8

Operational evaluation

  • 92% of target variables, pressure levels, and lead times have lower RMSE with Aurora than IFS HRES.At lead times beyond 12 hours, RMSE reductions reach 24%; IFS HRES remains better for many targets at the shortest lead times.Source: §6 · p. 9
  • Aurora outperforms IFS HRES on 10 m wind speed and 2 m temperature measurements for every lead time up to 10 days.The pretrained model is 25% better on average than training from scratch, reflecting the limited availability of 0.1° data.Source: §6 · p. 9

Storm Ciarán case study

  • Aurora is the only tested AI weather model that accurately predicts Storm Ciarán’s abrupt rise in maximum 10 m wind speed.Its prediction closely matches IFS analysis, treated as ground truth in the comparison.Source: §6 · p. 9

7 Discussion

Aurora delivers strong results across four Earth-system domains at a fraction of the computational cost. Its broader applicability is promising, but current operation still depends on traditional data-assimilation initial conditions.

Overall contribution

  • Aurora achieves state-of-the-art results for air quality, ocean waves, tropical-cyclone tracks, and high-resolution weather forecasting at a fraction of computational cost.Source: §7 · p. 9
  • Fine-tuning experiments took small engineering teams 4–8 weeks, compared with typically multiple years for dynamical baseline development.The accelerated timeline depends on data produced by decades of traditional numerical research.Source: §7 · p. 9

Limitations and future work

  • Aurora still relies on initial conditions from traditional data-assimilation systems, although it could eventually operate directly on observational data.Source: §7 · p. 9

Broader implications

  • Aurora could be fine-tuned for diverse Earth-system tasks, potentially making actionable predictions accessible at modest computational cost.Suggested applications include ocean circulation, regional weather, vegetation, floods, wildfires, agriculture, renewable energy, and sea ice.Source: §7 · p. 9

8 Methods

Aurora learns an autoregressive Earth-system simulator that maps successive observed states to future states through a flexible 3D latent representation. Its methods combine multi-scale processing, staged fine-tuning, diverse datasets, infrastructure for heterogeneous data, and latitude-weighted verification metrics.

8 Methods

  • Aurora’s simulator maps the previous and current observed states to the next predicted state, then repeatedly applies this operation for autoregressive roll-outs.The state includes surface and atmospheric variables arranged over latitude–longitude grids.Source: §8 · p. 13

8.1 The Aurora model

  • Aurora uses a flexible Perceiver encoder to standardize heterogeneous variables, pressure levels, and resolutions into a 3D latent representation.Variable-specific embeddings are aggregated into three latent pressure levels and augmented with encodings for position, patch area, and absolute time.Source: §8.1 · p. 13
  • A multi-scale 3D Swin Transformer U-Net processes the latent state with local windowed attention, shifted windows, and symmetric downsampling–upsampling stages.The backbone has 48 layers across three stages and is designed to operate at multiple resolutions.Source: §8.1 · p. 13

8.2 Training methods

  • Training proceeds through pretraining, short-lead-time fine-tuning, and roll-out fine-tuning using MAE as the objective.Long-roll-out adaptation uses LoRA, the pushforward trick, and an in-memory replay buffer to improve data, parameter, and memory efficiency.Source: §8.2 · pp. 14–15

8.3 Datasets

  • Aurora is trained and evaluated on diverse analysis, reanalysis, forecast, reforecast, and climate-simulation datasets spanning different variables, resolutions, and temporal coverage.Task-specific adaptations include density channels to represent spatially varying data availability in ocean-wave forecasting.Source: §8.3 · p. 15, §8.4 · p. 15

8.5 Data infrastructure

  • Large weather datapoints and heterogeneous resolutions, variables, and pressure levels make training primarily data-loading rather than model-computation limited.The authors address this with chunked storage, compression, dataset-specific batching, workload balancing, and a common BatchGenerator interface.Source: §8.5 · pp. 15–16

8.6 Verification metrics

  • Evaluation combines latitude-weighted RMSE and ACC with thresholded RMSE curves that assess errors across increasing weather-event intensities.Thresholds are computed from the mean and standard deviation of ERA5 data separately at each latitude–longitude point.Source: §8.6 · p. 16

Supplementary information.

The supplementary materials provide reproducibility details, additional text, and results supporting the paper’s main claims. They include comparisons with operational IFS HRES and GraphCast across typical and extreme weather conditions.

Supplementary information.

  • The supplementary document contains methodological details required to reproduce the work, alongside supplementary text and results supporting the main claims.Its listed contents include details on the Aurora model and extensive forecast comparisons.Source: Supplementary information · p. 17
  • Supplementary results compare Aurora with operational IFS HRES and GraphCast across weather conditions ranging from typical to extreme events.Source: Supplementary information · p. 17

Declarations.

The declarations report project funding, competing interests, approvals, availability statements, and author contributions. The project was supported by Microsoft Research AI for Science, with no competing interests reported.

Declarations.

  • The project was supported by Microsoft Research AI for Science, and the authors report no competing interests.Source: Declarations · p. 18
  • Ethics approval was not applicable, and all authors reviewed and consented to publication.Source: Declarations · p. 18
  • The declarations direct readers to separate sections for data and code availability and state that materials are not applicable.Source: Declarations · p. 18
  • Core contributors formulated, implemented, and evaluated Aurora’s architecture, training and fine-tuning pipelines, experiments, and evaluations, with additional contributors supporting infrastructure, experiments, comparisons, and supervision.Source: Declarations · p. 18

Notation and problem statement

This section defines Aurora’s Earth-system forecasting formulation and flexible encoder–backbone–decoder architecture, together with its application-specific extensions and configurations. It also describes the heterogeneous datasets used for pretraining, fine-tuning, and evaluation, including the adaptations required for wave forecasting.

Notation and problem statement

  • Aurora predicts a future state by iteratively applying a simulator Φ to consecutive observed or predicted states, with separate surface and atmospheric components.The state tensor has dimensions V × H × W, while atmospheric variables additionally span C pressure levels and V = V_S + V_A × C.Source: Notation and problem statement · p. 21

B.1 3D Perceiver encoder

  • The 3D Perceiver encoder standardises heterogeneous weather datasets by representing variables, pressure levels, resolutions, and available subsets in a common latent 3D tensor.It aggregates variable numbers of atmospheric pressure levels into C_L = 3 latent levels using Perceiver cross-attention, while retaining a latent surface level.Source: §B.1 · pp. 21–22

B.2 Multi-scale 3D Swin Transformer U-Net backbone

  • The backbone is a 48-layer 3D Swin Transformer U-Net that simulates atmospheric physics across multiple spatial scales.Encoder downsampling, decoder upsampling, shifted local windows, and direct communication across the spherical longitude boundary propagate information while avoiding vanilla Transformers’ quadratic complexity.Source: §B.2 · pp. 22–23, Figure B1 · p. 22

B.3 3D Perceiver decoder

  • The 3D Perceiver decoder reconstructs latent outputs into regular latitude–longitude patches and can produce predictions for arbitrary pressure levels and variable sets.Pressure-level embeddings are queried through a Perceiver layer, while variable-specific dynamic weights generate the output patches.Source: §B.3 · p. 23

B.4 Position, scale, level, and time encodings

  • Aurora encodes geometry, physical patch scale, pressure level, and absolute time using Fourier-style representations that remain compatible with varying spatial resolutions.Position uses patch-average latitude and longitude, scale uses physical patch area, pressure uses atmospheric pressure, and time uses hours since January 1, 1970.Source: §B.4 · pp. 23–24

B.5 Data normalisation

  • Aurora normalises each surface variable and atmospheric pressure level using spatially constant statistics estimated from ERA5, then unnormalises decoder outputs.Centres use empirical means and scales use empirical standard deviations computed over the whole ERA5 training data and reused across datasets.Source: §B.5 · p. 24

B.6 Extensions for 0.1° weather forecasting

  • For 0.1° forecasting, Aurora increases patch size from 4 to 10 and removes selected backbone layers so backbone resolution and memory remain manageable during fine-tuning.The modified encoder and decoder use stage sizes (6, 8, 8) and (8, 8, 6), respectively, without a significant initial-loss drop.Source: §B.6 · p. 24

B.7 Extensions for air pollution forecasting

  • For air pollution, Aurora adds temporal-difference modeling, pollution-specific normalisation and transformations, static covariates, pressure-dependent processing, and a separate pollution decoder.These adaptations address diurnal cycles, skewed concentration ranges, human and natural influences, differing pressure-level behaviour, and the absence of direct emissions inputs.Source: §B.7 · pp. 25–26
  • Aurora uses float32 before and after the backbone, with the backbone operating in bf16 mixed precision to better handle low-magnitude values.Source: §B.7 · p. 26

B.8 Extensions for wave forecasting

  • Wave adaptation also transforms angle-valued variables into sine and cosine components, initializes new variables with dedicated or transferred embeddings, and adds bathymetry, masks, and layer normalisation.The sine and cosine outputs are converted back with atan2; 10UN and 10VN reuse the pretrained 10U and 10V patch embeddings.Source: §B.8 · pp. 26–27
  • Wave forecasting adds density channels to distinguish missing values from zeros, predict data availability, and handle variables undefined over land or missing over water.Missing values are zeroed after normalisation, density channels are sigmoid-bounded to [0, 1], and autoregressive inputs are thresholded to avoid distribution mismatch.Source: §B.8 · p. 27

B.9 Model hyperparameters and configurations

  • The main Aurora configuration has 1.3B parameters, embedding dimension 512 initially, doubling backbone dimensions, 64 dimensions per attention head, a 1024-dimensional decoder, and 16 Perceiver cross-attention heads.Smaller scaling-experiment models reduce layers, embedding dimensions, and attention heads while preserving the 64-dimensional attention-head size.Source: §B.9 · p. 27, Table B1 · p. 27

C.1 Types of datasets

  • Aurora’s experiments use analysis, reanalysis, forecast, reforecast, and climate-simulation datasets that differ in observations, numerical-model use, and availability of future information.Reanalysis incorporates past and future information and is unavailable in real time, whereas analysis uses past observations and previous forecasts for operational initialization; reforecasts use reanalysis initial conditions, and climate simulations generate scenarios under different forcings.Source: §C.1 · p. 28

C.2 Dataset inventory

  • The dataset inventory combines global weather and atmospheric-composition products, ocean-wave data, ensemble forecasts, climate-model simulations, and multiple resolutions and pressure-level selections.Sources include ERA5, HRES, IFS ENS, GFS, GEFS, CMIP6, MERRA-2, HRES-WAM, CAMS, and CAMS reanalysis; collectively they expose variability from initial conditions, model parameterizations, and chaotic dynamics.Source: §C.2 · pp. 28–29

C.3 Motivation for CMIP6 dataset selection

Aurora’s data and training design combines heterogeneous Earth-system datasets with task-specific adaptation and preprocessing. The selected data provide complementary resolution, vertical coverage, temporal diversity, and handling for dataset-specific variables and missing values.

C.3 Motivation for CMIP6 dataset selection

  • CMCC-CM2-VHR4 and ECMWF-IFS-HR were selected for complementary resolutions and coverage of core meteorological variables across seven pressure levels.The resolutions are 0.25° and 0.45°, respectively.Source: §C.3 · pp. 29, 34
  • 1950–2014 temporal coverage in CMIP6 datasets provides climate variability that benefits extreme-event prediction, reducing tail errors by up to 35%.Source: §C.3 · p. 34

C.4 Training, validation, and test splits

  • The CAMS split accounts for distribution shifts from system updates by selecting a relatively stable period for fine-tuning and testing.Fine-tuning uses June 2021–May 2022 plus earlier post-fix data, while testing uses June–November 2022.Source: §C.4 · p. 34

C.5 Dataset-specific postprocessing

  • HRES-WAM neutral-wind components are reconstructed from wind magnitude and direction, with a minus sign because direction denotes where wind comes from.Source: §C.5 · p. 34

D.1 Training objective

  • Aurora trains with weighted MAE, adjusting variable and dataset weights and adding sine–cosine treatment for angles plus density channels for missing data.Dataset weights upweight ERA5 to 2.0 and GFS-T0 to 1.5.Source: §D.1 · p. 35

D.2 Pretraining methods

Aurora is pretrained extensively, then adapted through short-lead and roll-out fine-tuning. LoRA, replay buffers, and pushforward training make multi-step adaptation feasible under memory constraints.

D.2 Pretraining methods

  • Aurora pretraining uses 150 k steps on 32 A100 GPUs, exposing the model to about 4.8 million frames over approximately two and a half weeks.Training uses AdamW, cosine decay with warm-up, stochastic depth, activation checkpointing, gradient sharding, and bf16 precision.Source: §D.2 · p. 36
  • Aurora’s training cost is comparable to GraphCast and AIFS and lower than Pangu-Weather while supporting improved performance and additional capabilities.Source: §D.2 · p. 36

D.3 Short lead-time fine-tuning

  • Short-lead fine-tuning updates the entire architecture using one or two rollout steps, with task-dependent settings for resolution, memory, and data domain.Source: §D.3 · p. 36

D.4 Roll-out fine-tuning

  • Roll-out fine-tuning uses LoRA layers, propagates gradients only through the final rollout step, and trains with an in-memory replay buffer.The replay-buffer procedure avoids repeated expensive roll-out generation and extra memory or speed penalties.Source: §D.4 · p. 37, Figure D2 · p. 37

E Data infrastructure

Aurora’s infrastructure supports efficient training across large, heterogeneous datasets and evaluates forecasts with latitude-aware metrics. Extreme-event metrics extend RMSE to selected distribution tails.

E Data infrastructure

  • Aurora’s data pipeline addresses nearly 2 GB samples and heterogeneous resolutions by cloud storage, chunking, compression, and dataset-aware loading.Source: E Data infrastructure · p. 38
  • Dataset streams are sampled, shuffled, and sharded before expensive unpacking, so same-dataset samples remain batched while workloads can be balanced across GPUs.The pipeline performs dataset-specific downloading, decompression, checks, canonical variable mapping, batching, and device loading after sharding.Source: E Data infrastructure · p. 38, Figure E3 · p. 39

F.1 Main evaluation metrics

  • Evaluation uses latitude-weighted RMSE and ACC, with special handling for angular wave variables and missing wave data.Source: §F.1 · p. 39

F.2 Extreme weather metrics

  • Thresholded RMSE evaluates errors at grid points above or below thresholds defined from ERA5 means and standard deviations, with bootstrap 95% confidence intervals.Source: §F.2 · p. 40
  • Diverse pretraining improves extreme-value forecasting, while larger models reduce validation loss according to L(N) ∝N−0.026.At 5 k GPU hours, every 10× increase in model size corresponds to a 6% validation-loss reduction.Source: Figure G4 · p. 41

G Data diversity and model scaling improve atmospheric forecasting

Aurora benefits from diverse pretraining data and larger models, with gains extending across variables, atmospheric levels, and extreme-value prediction. The section also establishes the evaluation setup for comparisons at 0.25° resolution.

G.1 Effects of data diversity

  • Adding climate simulations lowers RMSE across variable groups except 2T, while ensemble and forecast data provide further gains at covered levels and improve atmospheric coverage.The results support positive transfer from mixed-fidelity datasets, although the ordering of configurations varies across atmospheric regions.Source: G Data diversity and model scaling… · pp. 41–42

G Data diversity and model scaling improve atmospheric forecasting

  • Aurora’s C4 data configuration performs best overall, showing that heterogeneous weather and climate datasets improve forecasting even on an overlapping ERA5 test year.C4 combines broader atmospheric coverage with one-day forecast samples, and its positive transfer persists despite evaluation on ERA5 itself.Source: G Data diversity and model scaling… · p. 42
  • Up to 35% improvement in surface extreme-value RMSE follows from adding diverse pretraining data, with C2, C3, and C4 progressively improving over ERA5-only C1.The largest gains relative to C1 occur in distributional extremes rather than ordinary aggregate performance.Source: G Data diversity and model scaling… · p. 42

G.2 Model scaling laws for surface-level variables

G.1.1 Why increase the data diversity?

  • Diverse datasets contribute simulated weather patterns, plausible forecast evolutions, alternative model parameterizations, and complementary resolution and vertical-structure information.These sources expose Aurora to rare or future atmospheric states and reduce dependence on one numerical weather prediction model’s biases.Source: §G.1.1 · pp. 42–43

G.1.2 Detailed analysis

  • Additional data generally improve individual atmospheric variables, but specific humidity can worsen at some levels and geopotential benefits most consistently.The geopotential result occurs even when the added configuration lacks that variable, indicating useful learning from incomplete variable sets.Source: §G.1.2 · p. 43
  • C5, the superset of the other configurations, often underperforms even the ERA5-only model, potentially because of dataset mismatch, incomplete coverage, or insufficient training.The authors note that the large mixture may require substantially more than 150 k training steps.Source: §G.1.2 · p. 43

H.1 Experimental setup

  • The 0.25° comparison evaluates Aurora and GraphCast on HRES-T0 2022 using operational initialization times, HRES analysis targets, relative RMSE skill scores, and supplementary ACC metrics.This protocol is intended to reflect operational conditions and avoid ERA5’s future-observation contamination and lower fidelity relative to IFS HRES analysis.Source: §H.1 · p. 47

H.2 Aurora versus GraphCast at 0.25°

At 0.25°, Aurora generally outperforms GraphCast, Pangu, and IFS-based references across weather variables and lead times. The gains extend to station observations and extreme-event evaluation, while specific humidity and some upper-atmosphere or seasonal cases remain less favorable.

H.2 Aurora versus GraphCast at 0.25°

  • Aurora matches or outperforms GraphCast on 94% of targets, with gains up to 40% in the upper atmosphere and strongest improvements beyond three days.GraphCast remains more accurate for specific humidity at some short lead times near the surface and lower atmosphere.Source: §H.2 · p. 47, Figure H8 · p. 48
  • Aurora’s comparison with GraphCast is attributed to data diversity, larger scale, flexible encoding, foundation-model pretraining, and fine-tuning procedures.The cited evidence links diverse data to up to 35% extreme-value error reduction and model scaling to systematic validation improvements.Source: §H.2 · p. 48

H.3 Aurora versus Pangu at 0.25°

  • Aurora consistently outperforms Pangu across all variables and pressure levels at 0.25°.The comparison is summarized by the H14 scorecard.Source: §H.3 · p. 54, Figure H14 · p. 54

H.4 Aurora versus IFS HRES at 0.25°

  • Aurora largely outperforms IFS HRES and exceeds the IFS ensemble mean through approximately seven days, but remains worse than HRES at the top of the atmosphere.This upper-atmosphere exception persists despite Aurora’s improvement over GraphCast there.Source: §H.4 · p. 54, Figure H15 · p. 54, Figure H16 · p. 55

H.5 Verification against weather station measurements at 0.25°

  • Against station observations, IFS HRES leads Aurora only within the first 12 hours for temperature, while Aurora improves over GraphCast for wind speed during the first six days.Aurora’s wind-speed RMSE becomes worse than GraphCast after day 7, particularly at 18 UTC timestamps.Source: §H.5 · p. 54, Figure H17 · p. 55

H.6 Extreme event prediction at 0.25°

  • Aurora outperforms GraphCast for surface wind-speed extremes, and AI weather models become markedly better relative to IFS HRES at longer lead times across the distribution.Thresholded RMSE evaluates upper and lower tails by sweeping thresholds across the 2022 HRES dataset.Source: §H.6 · p. 55, Figure H18 · p. 56
  • For atmospheric extremes across four pressure levels, Aurora remains able to outperform GraphCast, although higher-atmosphere prediction is more challenging.Near-surface results are consistent with the surface-variable analysis, while atmospheric temperature is less affected by near-surface bias.Source: §H.6 · p. 55, Figure H19 · p. 57

I.1 Comparison against operational CAMS

This section compares Aurora with operational and other forecasting systems across atmospheric and wave forecasting, including tropical-cyclone tracking and spectral behavior. It examines how pretraining, resolution, and LoRA adaptation affect forecast performance, detail, and consistency with analyses.

I.1 Comparison against operational CAMS

  • Aurora matches or outperforms CAMS on 76% of all targets, remains within 20% RMSE on 96%, and beats persistence across all variables and lead times.For surface-level variables, Aurora matches or outperforms CAMS on 89% of variables; CAMS analysis is treated as ground truth and may favor CAMS because it assimilates observations without real-time emissions.Source: §I.1 · p. 58, Figure I20 · p. 59, Figure I22 · p. 61
  • Fine-tuned Aurora consistently outperforms models trained from scratch on CAMS data, with pretraining improving performance by 54% on average.The scratch model uses the CAMS analysis training dataset, whose limited data and challenging distributional behavior may make atmospheric prediction difficult.Source: §I.1 · p. 58, Figure I26 · p. 64

I.2 Comparison against operational IFS HRES-WAM

  • Aurora is competitive with HRES-WAM on 95% of targets and matches or outperforms it on 85% of all wave and meteorological targets.The comparison uses Aurora and HRES-WAM at 0.25° for the relevant wave and meteorological variables.Source: Figure I27 · p. 65, §I.2 · p. 65

I.3.1 Baselines and ground truth

  • Aurora’s tropical-cyclone comparisons use official and guidance forecasts across multiple agencies, IBTrACS ground truth, and a six-hour delay to ensure Aurora is treated as an early model.Aurora retains predictions even when its simpler tracker fails, slightly disadvantaging it because baselines and IBTrACS use state-of-the-art trackers.Source: §I.3.1 · p. 68, §I.3.2 · p. 68

I.3.3 Tracker

  • The tropical-cyclone tracker follows local minima in MSL, falls back to Z700 when needed, extrapolates from the previous eight positions, and produces forecasts to five days.It searches progressively smaller boxes around the current position and uses 6 h predictions at each update.Source: §I.3.3 · pp. 68–69

I.3.4 Details for Figure 4a

  • For Figure 4a, tropical-cyclone errors are averaged over disturbance or tropical storm times, with 95% uncertainty intervals obtained by resampling entire cyclones.The evaluation includes 1333 storm-time pairs from 76 storms in the North Atlantic and East Pacific, 737 pairs from 47 storms in the West Pacific, and 213 pairs from 35 storms in Australia.Source: §I.3.4 · p. 70

I.5 Comparison against weather station measurements

  • Aurora’s weather-station evaluation outperforms IFS HRES for 10 m wind speed forecasts at every lead time through 10 days.The evaluation uses raw WeatherReal-ISD observations from daily forecasts issued at 00 UTC between 2 January and 19 December 2023; comparison on 2 m temperature gives Aurora an initialization advantage, while its errors remain parallel to HRES as lead time increases.Source: §I.5 · p. 74

I.6 Benefits of modelling at 0.1° resolution

  • Aurora 0.1° improves meso-β-scale predictability by 5.4% in R2 on average, while meso-α and synoptic predictability changes remain 0.3% and 0.03%.The comparison uses Aurora 0.25° predictions regridded to 0.1°, with IFS HRES analysis at 0.1° as ground truth and spectral decomposition by meteorological scale.Source: §I.6 · pp. 74–75, Figure I35 · p. 75

I.7 Storm Ciarán

  • Aurora forecasts Storm Ciarán similarly to IFS analysis, capturing intensification and approximately 7 ms−1 higher peak wind speed than other AI techniques.Aurora’s forecasts are smoother than IFS analysis but less smooth than 0.25° models; at 0.1° resolution, it captures the sharp wind-speed increase more closely.Source: §I.7 · p. 76
  • Aurora variants at 0.1° resolution clearly outperform 0.25° variants when evaluated against IFS analysis for Storm Ciarán.LoRA improves minimum mean sea level pressure, whereas disabling LoRA better captures peak 10 m wind speed; 0.1° Aurora with LoRA nearly matches IFS analysis for pressure.Source: §I.7 · p. 76
  • LoRA provides a controllable trade-off between RMSE performance and high-frequency detail during long-term roll-outs, supporting application-specific operational predictions.LoRA is favored for smooth quantities such as mean sea-level pressure, while disabling it preserves wind-speed extrema; spectral analysis supports this distinction.Source: §I.7 · p. 76

I.8 Power spectra

  • At short wavelengths, Aurora has uniformly higher power than GraphCast, while its mid-range power exceeds GraphCast at 1–2 days but falls below it at 5–10 days.Adding LoRA steps generally reduces forecast power, and 40 steps at 10 days clearly reduce short-wavelength power, producing blurring.Source: §I.8 · p. 76, Figure I37 · p. 77, Figure I39 · p. 78
  • Aurora’s 0.1° forecasts generally have lower power than 0.25° forecasts but extend to shorter wavelengths, reflecting higher resolution.The spectral difference between Aurora with and without LoRA is analogous to the difference between the IFS ensemble mean and IFS HRES.Source: §I.8 · pp. 76–77, Figure I38 · p. 77

Related papers

Loading 2405.13063v3…