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

Concept 1: What is a Foundation Model (and Why It Matters for Weather)?

Before diving into Aurora specifically, you need to understand what a foundation model is.

Traditional Approach vs. Foundation Model

Traditional weather models are:

• Built on physics equations (fluid dynamics, thermodynamics)
• Designed for one specific task (e.g., only ocean waves, or only surface temperature)
• Computationally expensive to run

A foundation model is:

• Trained on massive amounts of data using AI/machine learning
• Designed to be adaptable — one model can be fine-tuned for many different tasks
• Think of it like GPT for language, but for the Earth system

Key Insight

Aurora is a single model that can be adapted to medium-range weather, ocean waves, atmospheric chemistry, and climate applications — rather than needing a separate model for each.

---

Concept 2: What is Aurora (the Original Model)?

Background

• Developed by Microsoft Research AI for Science
• First introduced in 2024, published in Nature in 2025
• It proved that one AI model could handle multiple Earth-system tasks

What Made It Special?

Analogy

Think of Aurora like a Swiss Army knife for Earth-system science — one tool with many attachments, rather than a separate tool for every job.

---

Concept 3: What is Aurora 1.5? (The Major Update)

Aurora 1.5 is an extension of the original Aurora model. It adds three major capabilities:

3 Core Additions

① More Weather Variables (Breadth Update)

• Original Aurora tracked 4 weather variables
• Aurora 1.5 adds 22 more variables (total: 26)
• These include:
  • Surface conditions
  • Wind, temperature, humidity
  • Precipitation
  • Radiation fields
  • Pressure-level data

Why it matters: More variables = more sectors can use it (energy, agriculture, transport, climate risk)

② Hourly Temporal Resolution

• Previous models gave forecasts at coarser time steps
• Aurora 1.5 provides hour-by-hour forecasts

Why it matters: Critical for time-sensitive decisions like:

• When exactly will rain begin?
• When will a hurricane make landfall?
• When should energy traders act?

③ Ensemble Forecasting (Most Requested Feature)

• This is the biggest conceptual addition — explained in detail in Concept 4

---

Concept 4: What is Ensemble Forecasting?

This is the most important new concept in Aurora 1.5, so let's break it down carefully.

The Core Problem: Weather is Uncertain

Weather forecasts are sensitive to:

1. Initial conditions — tiny measurement errors in today's data compound over time
2. Model uncertainty — no model perfectly represents the atmosphere

The Old Approach: Single "Best Guess" Forecast

• Run the model once
• Get one predicted outcome
• Problem: You don't know how confident to be in that prediction

The Ensemble Approach: Multiple Simulations

• Run the model many times with slightly different starting conditions
• Get a range of possible outcomes
• This range tells you the probability of different scenarios

Visual Analogy

Imagine predicting where a hurricane will go:

Single forecast:  →  One track line on a map

Ensemble forecast: →  A "cone" of many possible tracks
                      (some go left, some go right, 
                       most cluster in the middle)

The spread of those tracks tells you how uncertain the forecast is.

How Aurora 1.5 Builds Its Ensemble

The team used a multi-stage process:

1. Started with the original Aurora model
2. Added more variables + hourly resolution
3. Introduced stochastic perturbations (controlled random variations) into the model's internal processing pathway
4. Optimized specifically for probabilistic forecast quality
5. Fine-tuned on ECMWF high-resolution data (2018–2023) for stability

Key term: Stochastic perturbations = intentional small random changes introduced to simulate the natural uncertainty in initial conditions

---

Concept 5: How Do You Measure Ensemble Forecast Quality?

Probabilistic Skill Scores

Unlike single forecasts (where you just check if the prediction was right), ensemble forecasts are judged on:

• Calibration — Does a "70% chance of rain" actually happen 70% of the time?
• Sharpness — Is the uncertainty range narrow (confident) when it should be?
• Spread-skill relationship — Does wider spread actually correspond to harder-to-predict situations?

Aurora 1.5's Results

• Outperforms ECMWF's dynamical ensemble (the current gold standard) on 88.9% of evaluated targets
• For tropical cyclones (all 2024–2025 storms evaluated):
  • ~one-third lower track error compared to original Aurora
  • Demonstrated on Hurricane Helene as a real-world example

ECMWF = European Centre for Medium-Range Weather Forecasts — widely considered the world's best operational weather forecasting system. Beating it is a significant benchmark.

---

Concept 6: Beyond Weather — Earth-System Foundation Models

Aurora 1.5 illustrates a broader concept: Earth-system foundation models can be applied far beyond weather forecasting.

What is an "Earth System"?

The Earth system includes:

• Atmosphere (weather, climate)
• Oceans (waves, currents, temperature)
• Land surface (soil, vegetation, carbon)
• Cryosphere (ice, snow)

All of these interact with each other.

Aurora's Expanding Applications

Key Principle

Foundation models complement physics-based models — they don't replace them. The best approach combines AI flexibility with established scientific understanding.

---

Concept 7: Open Source + Operational Use — The Two-Track Model

Aurora 1.5 demonstrates an important deployment philosophy:

Track 1: Open Research

• Model code released on GitHub
• Model weights (checkpoints) on Hugging Face
• Anyone can download, evaluate, modify, and build on it
• Encourages global scientific collaboration

Track 2: Operational/Enterprise Use

• Microsoft Weather provides managed services on top of the open model
• Adds: data pipelines, cloud infrastructure, reliability guarantees, support
• Connects to Microsoft Foundry, Planetary Computer Pro, Azure
• For organizations that need weather intelligence but can't build everything themselves

Why This Matters

This two-track approach solves a common tension in AI:

• Openness drives scientific progress and trust
• Managed services make the technology accessible to organizations without deep AI expertise

---

Summary: The Conceptual Arc

Here's how all the concepts connect:

Foundation Models (Concept 1)
        ↓
Aurora: One model for many Earth-system tasks (Concept 2)
        ↓
Aurora 1.5: More variables + hourly resolution + ensembles (Concept 3)
        ↓
Ensemble Forecasting: Quantifying uncertainty with multiple runs (Concept 4)
        ↓
Measuring probabilistic skill (Concept 5)
        ↓
Expanding to the full Earth system (Concept 6)
        ↓
Open source + operational deployment (Concept 7)

The Big Takeaway

Aurora 1.5 represents a maturation of AI weather modeling — moving from "can AI do weather?" to "how do we make AI weather models reliable, uncertain-aware, and broadly useful for real decisions?"

The ensemble capability is the philosophical heart of this: acknowledging and quantifying uncertainty is more valuable than false precision.