PaperlayerView exact revision on arXiv

Source-linked AI summary

EPT-2 Technical Report

Roberto Molinaro, Niall Siegenheim, Niels Poulsen, Jordan Dane Daubinet, Henry Martin, Mark Frey, Kevin Thiart, Alexander Jakob Dautel, Andreas Schlueter, Alex Grigoryev, Bogdan Danciu, Nikoo Ekhtiari, Bas Steunebrink, Leonie Wagner, Marvin Vincent Gabler

arXiv:2507.09703v1cs.LGcs.AI

TL;DR

EPT-2 addresses the need for accurate, efficient Earth-system forecasts across energy-relevant variables and long lead times. It introduces a dynamically conditioned deterministic model and the perturbation-based ensemble EPT-2e. Across the reported evaluations, EPT-2 improves on predecessor and operational baselines, while EPT-2e outperforms ECMWF ENS mean with lower computational cost.

  • Problem

    The paper targets the challenge of producing accurate forecasts for energy-relevant weather variables across long lead times while supporting operational forecasting needs.

    Source: §I · p. 1, §V · p. 5

  • Method

    The paper introduces EPT-2, a dynamically conditioned deterministic model, and EPT-2e, its perturbation-based ensemble variant.

    Source: §I · p. 1, A. Key Features · p. 2

  • Results

    EPT-2 consistently outperforms EPT-1.5 and ECMWF HRES across the evaluated variables and lead times, while EPT-2e outperforms the ECMWF ENS mean across the reported forecast range.

    Source: A. Deterministic Forecasting Results · p. 4, B. Probabilistic Forecasting Results · p. 4, §V · p. 5

  • Takeaways & Limitations

    The EPT-2 family provides hourly global forecasts up to 20 days ahead, including high-resolution ensemble prediction for weather-sensitive applications.

    Source: §V · p. 5

Abstract

from arXiv · show

We present EPT-2, the latest iteration in our Earth Physics Transformer (EPT) family of foundation AI models for Earth system forecasting. EPT-2 delivers substantial improvements over its predecessor, EPT-1.5, and sets a new state of the art in predicting energy-relevant variables-including 10m and 100m wind speed, 2m temperature, and surface solar radiation-across the full 0-240h forecast horizon. It consistently outperforms leading AI weather models such as Microsoft Aurora, as well as the operational numerical forecast system IFS HRES from the European Centre for Medium-Range Weather Forecasts (ECMWF). In parallel, we introduce a perturbation-based ensemble model of EPT-2 for probabilistic forecasting, called EPT-2e. Remarkably, EPT-2e significantly surpasses the ECMWF ENS mean-long considered the gold standard for medium- to longrange forecasting-while operating at a fraction of the computational cost. EPT models, as well as third-party forecasts, are accessible via the app.jua.ai platform.

I. INTRODUCTION

EPT-2 and EPT-2e extend Jua’s AI weather-modeling work with deterministic and ensemble forecasting designed for broad, operational use. The models target improved accuracy, efficiency, and coverage across energy-relevant weather variables and lead times.

  • AI Earth-system models directly map observational data to forecasts using deep learning, complementing numerical weather prediction based on discretized physical equations.Source: §I · p. 1
  • EPT models build on earlier AI weather systems, including Vilhelm, EPT-1, and EPT-1.5, which progressively expanded operational and high-resolution forecasting capabilities.EPT-1.5 had previously outperformed GraphCast, Pangu-Weather, and FuXi-Weather [10].Source: §I · p. 1
  • EPT-2 is introduced as a compute-efficient, dynamically conditioned model delivering state-of-the-art deterministic forecasts.Source: §I · p. 1
  • EPT-2e is an ensemble version that outperforms Aurora and ECMWF ENS across key variables and the full 0–240 h forecast range.It achieves lower RMSE and CRPS throughout that range.Source: §I · p. 1
  • The EPT-2 family offers approximately 25% faster inference, flexible lead-time handling, and broad variable support.Source: §I · p. 1

II. EPT-2: MODEL ARCHITECTURE AND OPERATIONAL SETTING

EPT-2 uses a learned forecasting operator that maps recent atmospheric states to future conditions, with training and evaluation structures designed for gridded weather fields. Its objective applies latitude-aware spatial weighting to account for uneven grid-cell areas.

  • The learned approach replaces hand-coded physical solvers with physical relationships learned from data while retaining the state-to-state forecasting logic of numerical weather prediction.Source: §II · p. 2
  • EPT-2 approximates the mapping from current and previous weather states to a future state with a machine-learning model fθ.Source: §II · p. 2
  • The model is trained by minimizing an objective over weather variables and spatial grid dimensions.C denotes weather variables, while H and W represent latitude and longitude dimensions.Source: §II · p. 2
  • Latitude-based weights adjust each objective term for varying grid-point density across geographic coordinate systems.Source: §II · p. 2

A. Key Features

EPT-2 combines diverse-data training, arbitrary lead-time forecasting, and perturbation-based probabilistic prediction with operational specifications for hourly, global forecasts. The reported configuration uses roughly 9 × 9 km spatial resolution and runs to 20 days ahead.

A. Key Features

  • EPT-2 offers predictions for any lead time ∆t and generates native hourly forecasts in operational use.Source: A. Key Features · p. 2
  • EPT-2e extends the core model with perturbation-based ensembles for uncertainty-aware forecasting and more robust decision-making under uncertainty.Source: A. Key Features · p. 2

B. Operational Specifications •

  • The discussed EPT-2 variant operates at 0.083° resolution, roughly 9 × 9 km at the equator.A higher-resolution variant exists but is reserved for future discussion; the paper uses the 9 × 9 km variant for comparability.Source: B. Operational Specifications • · p. 2
  • EPT-2 runs hourly up to 20 days ahead and executes four times daily at 00:00, 06:00, 12:00, and 18:00 UTC.The standard version typically has a 6-hour delay, while the early version is available approximately 2.5 hours earlier1.Source: B. Operational Specifications • · p. 2

III. BENCHMARKS DETAILS

The benchmark framework uses established gridded and station-based ground truths to evaluate global weather forecasts. Weather stations provide the primary evaluation target because they offer observations less affected by model-training biases.

III. BENCHMARKS DETAILS

A. Ground Truth Datasets

  • ERA5 is an hourly global reanalysis spanning 1940 to the present at approximately 31 x 31 km resolution.Source: A. Ground Truth Datasets · p. 2
  • IFS HRES initial conditions provide spatially complete 9 x 9 km inputs widely used for aligned benchmarking despite not being direct observations.Source: A. Ground Truth Datasets · p. 2
  • WeatherReal-ISD aggregates quality-controlled observations from over 14,000 global stations, which serve as the primary evaluation target.Station observations are spatially sparse and unevenly distributed but are treated as a more neutral ground truth than potentially training-influenced gridded datasets.Source: A. Ground Truth Datasets · pp. 2–3
  • Figure 1 maps the spatial distribution of the weather stations used for benchmarking.Source: Figure 1 · p. 3

B. Models Evaluations

The evaluation separates gridded and point benchmarks and standardizes initialization, references, and metrics. RMSE assesses deterministic error, while CRPS evaluates probabilistic forecast quality through accuracy and ensemble spread.

B. Models Evaluations

C. Methodology

  • IFS HRES is the deterministic reference, while ECMWF ENS mean is additionally reported for probabilistic evaluations.IFS HRES is benchmarked against initial conditions and weather stations; the ENS mean averages forecasts across ensemble members.Source: C. Methodology · p. 3
  • RMSE and CRPS jointly assess deterministic and probabilistic forecast quality.RMSE measures average forecast-error magnitude, whereas CRPS measures the difference between a forecast CDF and the observation step function.Source: C. Methodology · p. 3
  • For ensemble forecasts, CRPS separates accuracy from ensemble spread, and lower CRPS indicates better overall probabilistic performance.A calibrated ensemble balances forecast skill with spread rather than being excessively narrow or broad.Source: C. Methodology · p. 3

IV. RESULTS

The results compare EPT-2-family forecasts with EPT-1.5 and Aurora under deterministic and probabilistic settings. Probabilistic evaluation uses perturbed initial conditions propagated independently to form ensembles.

IV. RESULTS

  • The study compares the EPT-2 family with EPT-1.5 and Aurora across deterministic and probabilistic evaluation settings.Source: §IV · p. 3
  • Deterministic forecasting evaluates skill from unperturbed initial conditions.Source: §IV · p. 3
  • Probabilistic forecasting evaluates uncertainty estimates and distributional outputs using ensembles generated from independently propagated perturbations of the initial conditions.Source: §IV · p. 3

A. Deterministic Forecasting Results

EPT-2 improves on EPT-1.5 and generally outperforms Aurora and HRES across deterministic forecasts, with performance varying by variable, reference, and lead time. The evaluation covers 0–240 h forecasts of wind, temperature, and solar radiation.

A. Deterministic Forecasting Results

  • EPT-2 consistently outperforms EPT-1.5 across all evaluated variables and lead times against initial conditions.The variables include 10 m and 100 m wind speed, 2 m air temperature, and surface solar radiation.Source: A. Deterministic Forecasting Results · p. 4
  • EPT-2 has lower Aurora RMSE for 10 m wind speed across 0–240 h and for 2 m temperature through 130 h.Its relative temperature advantage decreases at longer lead times.Source: A. Deterministic Forecasting Results · p. 4
  • EPT-2 improves over HRES beyond 12 h across all variables, while solar-radiation performance exceeds both EPT-1.5 and HRES throughout 0–240 h.Source: A. Deterministic Forecasting Results · p. 4
  • Against station observations, EPT-2 consistently improves 10 m wind forecasts over EPT-1.5 and becomes more accurate for temperature after approximately 120 h.For temperature, EPT-2 initially has slightly higher RMSE than EPT-1.5 through 100 h.Source: A. Deterministic Forecasting Results · p. 4
  • Against Aurora and HRES at stations, EPT-2 gains lower wind RMSE beyond 6 days and maintains a temperature advantage across almost the full lead-time range, while all EPT variants outperform HRES.Source: A. Deterministic Forecasting Results · p. 4
  • Four-day Zurich time series illustrate that Aurora misses most temperature peaks and overlooks a brief wind-speed rise and fall around 36 h.The wind-pattern difference is identified as potentially important for energy applications.Source: A. Deterministic Forecasting Results · p. 4

B. Probabilistic Forecasting Results.

EPT-2e provides probabilistic forecasts that outperform competing ensembles across the 0–240 h horizon, using only 10 members and substantially lower computational cost.

Probabilistic Forecasting Results

C. Comparison with Aurora

EPT-2 compares favorably with Aurora in forecast accuracy, training efficiency, inference speed, and flexible lead-time handling.

Comparison with Aurora

  • EPT-2 consistently outperforms Aurora across variables and lead times in forecast accuracy.Source: C. Comparison with Aurora · p. 5
  • EPT-2 achieves comparable or superior forecast skill after pretraining on 8 H100 GPUs for 10 days, versus Aurora’s 32 A100 GPUs for 18 days.The report also describes EPT-2 as more compute-efficient and cost-effective than other state-of-the-art models.Source: C. Comparison with Aurora · p. 5
  • EPT-2 enables approximately 25% faster inference than Aurora, supporting time-critical operational deployments.Source: C. Comparison with Aurora · p. 5
  • Dynamic lead-time conditioning lets EPT-2 generalize across arbitrary lead times without specialized output heads or model replications.This reduces training and inference complexity relative to Aurora’s fixed-lead-time forecasts.Source: C. Comparison with Aurora · p. 5
  • Overall, EPT-2 matches or outperforms Aurora in most tasks while using less training compute, lower inference cost, and greater temporal flexibility.Source: C. Comparison with Aurora · p. 5

V. DISCUSSION

EPT-2 improves energy-relevant weather forecasting across deterministic and ensemble evaluations, with hourly global forecasts and operational relevance for weather-sensitive sectors.

V. DISCUSSION

  • EPT-2 improves wind, temperature, and solar-radiation forecasting while providing hourly global forecasts up to 20 days ahead and high-resolution ensemble prediction.Source: §V · p. 5
  • EPT-2 consistently outperforms ECMWF HRES across multiple variables, while EPT-2e exceeds the ECMWF ENS mean across the full time range using fewer computational resources.Source: §V · p. 5
  • The reported accuracy, temporal resolution, and efficiency are positioned as useful for real-time decision-making in weather-sensitive sectors.Source: §V · p. 5
  • The deterministic temperature evaluation uses 2m air-temperature RMSE against in-situ weather-station observations over the 0–240 h horizon.Source: Figure 6 · p. 6
  • The ensemble-mean temperature evaluation likewise measures 2m air-temperature RMSE against in-situ station observations across the 0–240 h horizon.Source: Figure 8 · p. 6

VII. SUPPLEMENTARY MATERIAL

Supplementary figures visualize EPT-2, EPT-1.5, and Aurora forecasts for temperature and wind, including global snapshots and Zurich time series.

VII. SUPPLEMENTARY MATERIAL

  • Figure 12 compares 2m air-temperature predictions from EPT-2, EPT-1.5, and Aurora at 1-day, 2-day, and 10-day lead times alongside the initial condition.Source: Figure 12 · p. 8
  • Figures 13 and 15 show Zurich time series for temperature and wind forecasts, with hourly EPT outputs versus six-hour Aurora outputs.Source: Figure 13 · p. 8, Figure 15 · p. 9
  • Figure 14 compares 10m wind-speed predictions from EPT-2, EPT-1.5, and Aurora at 1-day, 2-day, and 10-day lead times alongside the initial condition.Source: Figure 14 · p. 9
Loading 2507.09703v1…