TLDR: Jua.ai’s new EPT-2 model significantly advances AI-based Earth system forecasting, outperforming previous versions and leading models like Microsoft Aurora and ECMWF’s IFS HRES and ENS. It provides accurate, hourly forecasts for energy-relevant variables up to 20 days, including robust probabilistic predictions, while being highly computationally efficient in both training and inference. This makes EPT-2 a powerful tool for weather-sensitive sectors like energy trading.
Jua.ai has unveiled its latest breakthrough in artificial intelligence for Earth system forecasting: EPT-2. This new model represents a significant leap forward, building upon its predecessors in the Earth Physics Transformer (EPT) family. EPT-2 is designed to predict crucial energy-relevant variables, such as wind speed at 10m and 100m, 2m temperature, and surface solar radiation, across a comprehensive 0-240 hour forecast horizon.
Remarkably, EPT-2 consistently outperforms other prominent AI weather models, including Microsoft Aurora, and even surpasses the highly respected operational numerical forecast system IFS HRES from the European Centre for Medium-Range Weather Forecasts (ECMWF).
What is EPT-2?
At its core, EPT-2 is a sophisticated machine learning engine that learns to map past atmospheric conditions to future states. Unlike traditional numerical weather prediction, which relies on hand-coded physical solvers, EPT-2 learns these solvers directly from vast amounts of data. This approach allows it to process diverse datasets from satellites, radars, and sensors, capturing complex interactions within the Earth’s systems.
Key features of EPT-2 include its foundation as an Earth Systems Model, trained on extensive datasets. It offers ‘Any Lead Time Forecasting,’ meaning it can provide predictions for any desired future time point, generating native hourly forecasts. This is a notable advantage over many current AI models that are limited to fixed forecast intervals. Additionally, EPT-2 supports ‘Probabilistic Forecasting’ through its ensemble model, EPT-2e, which provides uncertainty-aware predictions for more robust decision-making.
Operational Capabilities
EPT-2 operates at a spatial resolution of approximately 9×9 km at the equator, delivering highly detailed forecasts. It currently runs at a temporal resolution of 1 hour, extending up to 20 days into the future. The model executes four times daily at 00:00, 06:00, 12:00, and 18:00 UTC, with both early and standard versions available.
Performance Benchmarks
The performance of EPT-2 was rigorously evaluated against established benchmarks. For deterministic forecasting, EPT-2 showed lower Root Mean Square Error (RMSE) compared to EPT-1.5, Aurora, and ECMWF IFS HRES across various variables and lead times. For instance, it consistently outperformed Aurora in 10m wind speed forecasts and maintained an advantage in 2m temperature predictions for most lead times.
In the realm of probabilistic forecasting, the EPT-2e ensemble model demonstrated exceptional performance. It achieved the lowest RMSE and Continuous Ranked Probability Score (CRPS) for 2m air temperature and 10m wind speed, consistently outperforming the ECMWF ENS mean, which is often considered the gold standard. This is particularly impressive given that EPT-2e uses significantly fewer ensemble members (10 vs. 50 for ECMWF ENS), leading to a fraction of the computational cost.
Efficiency and Flexibility
A significant advantage of EPT-2 is its computational efficiency. While Aurora required 32 A100 GPUs for 18 days of pretraining, EPT-2 was pretrained on just 8 H100 GPUs for 10 days. Despite using considerably less hardware and training time, EPT-2 achieves comparable or superior forecast skill. Furthermore, EPT-2 enables approximately 25% faster inference than Aurora, making it more suitable for time-critical operational deployments.
Its dynamic lead time conditioning also provides greater flexibility, allowing it to generalize across arbitrary lead times without needing multiple specialized output heads or model replications, which simplifies both training and inference processes.
Also Read:
- A New Hybrid AI Model for Smarter Energy Forecasting
- Unveiling the Continuous Nature of Time Series with NeuTSFlow
Impact and Future
EPT-2 represents a substantial advancement in AI-based weather forecasting, particularly for sectors dependent on accurate, high-frequency predictions like energy production, including renewables, gas, and grid management. Its combination of accuracy, high temporal resolution, and efficiency positions it as a powerful tool for real-time operational use, enabling more informed and responsive decision-making in weather-sensitive industries. For more details, you can refer to the full research paper: EPT-2 Technical Report.
