Aurora 1.5: Extending open foundation models for weather and Earth-system applications

2026-07-09 · Source: Microsoft Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Environmental Science & Earth Systems, Data Science & Analytics · Depth: Advanced, medium

Summary

Microsoft has released Aurora 1.5, a significant extension of its open-source Aurora Earth System foundation model. This update introduces 22 new weather variables, bringing the total to 26, alongside hourly temporal resolution and probabilistic ensemble forecasting capabilities. Available on GitHub with model checkpoints on Hugging Face, Aurora 1.5 aims to empower researchers and developers while connecting to Microsoft Weather services for operational use. The model's new ensemble feature runs multiple simulations to quantify forecast uncertainty, outperforming the ECMWF dynamical ensemble on 88.9% of evaluated targets and reducing tropical cyclone track errors by approximately one-third. Aurora 1.5 supports diverse applications, from energy and agriculture to climate risk and carbon dioxide removal, demonstrating its utility as an enterprise-grade weather solution integrated into platforms like Microsoft Foundry.

Key takeaway

For Machine Learning Engineers or Directors of AI/ML evaluating advanced weather forecasting solutions, Aurora 1.5 provides a powerful, open-source option. Its enhanced hourly resolution and probabilistic ensemble forecasting, which outperforms ECMWF ENS on 88.9% of targets, can significantly improve your operational guidance and risk management. You should explore integrating Aurora 1.5 into your workflows for more confident decision-making in weather-dependent sectors like energy, agriculture, or climate resilience planning.

Key insights

Aurora 1.5 extends open Earth-system foundation models with advanced probabilistic ensemble forecasting for practical, scalable applications.

Principles

• Open-source foundation models accelerate research and operational deployment.
• Ensemble forecasting quantifies uncertainty, enhancing decision-making for weather-sensitive operations.
• Integrating research models with robust infrastructure enables enterprise-grade solutions.

Method

Multi-stage fine-tuning involved expanding variables, adding hourly resolution, introducing stochastic perturbations in the latent conditioning pathway, optimizing for probabilistic quality, and auto-regressive fine-tuning on ECMWF HRES data from 2018 to 2023.

In practice

• Apply Aurora 1.5 for precision operational guidance like precipitation onset.
• Utilize ensemble forecasts for power systems, agriculture, and extreme-weather planning.
• Integrate Aurora-derived representations for carbon dioxide removal estimation.

Topics

• Aurora 1.5
• Earth System Models
• Ensemble Forecasting
• Open-Source AI
• Weather Forecasting
• Climate Resilience

Code references

• microsoft/aurora
• microsoft/vibe-kit

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• NVIDIA Launches Earth-2 Family of Open Models — the World’s First Fully Open, Accelerated Set of Models and Tools for AI Weather — NVIDIA's Earth-2 provides an open, accelerated AI software stack for comprehensive weather and climate prediction.
• Time-Series LLMs, Explained with t0-alpha — Small, open-weight time-series foundation models like t0-alpha offer robust probabilistic forecasting, often outperforming classical baselines.
• Ensemble High-Resolution Weather Forecasting on AMD Instinct GPU Accelerators — Diffusion models enable robust ensemble weather forecasting by generating diverse atmospheric states to quantify uncertainty.
• The next chapter in flood resilience: Open sourcing Google’s hydrology framework — Open-sourcing Google's AI hydrology framework democratizes advanced flood forecasting, enabling local adaptation and global resilience.
• Does Aurora Encode Atmospheric Structure? Latent Regime Analysis and Attribution — Aurora implicitly learns complex atmospheric structures and meteorological coherence without explicit instruction.
• Interview with AAAI Fellow Yan Liu: machine learning for time series — Physics-informed foundation models enable accurate, adaptable time series analysis, especially in data-scarce scientific and real-world applications.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Microsoft Research.