Aurora

What Aurora is

Aurora is a foundation model for the Earth system, published in Nature in May 2025, volume 641, pages 1180 to 1187^[1][2]. The term foundation model describes a large neural network trained first on broad data and only afterwards adapted to particular jobs — the same pattern that underlies general-purpose language models. Aurora applies that pattern to geophysical data. It is positioned as a model covering several Earth system domains rather than a single forecasting task^[1].

What it was trained on

The model was trained on more than one million hours of diverse geophysical data^[1]. Scale matters here: a network that has seen many kinds of measurement can pick up structure that transfers between them, so adapting it to a new problem does not mean starting over. Aurora can be fine-tuned for diverse applications at modest expense^[1]. Fine-tuning means taking the already trained network and adjusting it on a smaller, task-specific dataset, which is far cheaper than training from scratch.

Results

The authors compare Aurora with operational forecasts — the forecasts that weather and environmental agencies actually issue — across several areas. It outperforms operational forecasts for high-resolution weather^[1], for air quality^[1], for ocean waves^[1] and for tropical cyclone tracks^[1]. These are not variations on one problem: air quality depends on chemistry and emissions, ocean waves on wind and sea state, cyclone tracks on the large-scale flow. Covering all of them with one model is what the "several domains" framing means in practice.

The comparison comes with a second number attached. All of this is done at orders of magnitude lower computational cost than the numerical models it is compared with^[1]. Numerical models are the traditional approach: they encode physical equations and solve them step by step on large computers, and the expense of running them at high resolution limits how often and where such forecasts can be produced.

Authors and stated aim

The authors include Cristian Bodnar, Wessel Bruinsma, Ana Lucic and Megan Stanley^[1]. They present the work as a step towards democratising accurate and efficient Earth system prediction^[1]. The reasoning behind that claim is visible in the results above: if accurate prediction can be had for far less computation, and if a single trained model can be adapted to a new domain cheaply, then the ability to build forecasting tools is less tightly bound to a large computing budget.

Where it sits

Machine-learning approaches to weather and environmental forecasting are usually judged by comparing them directly against operational forecasts, and that is the yardstick used here. Aurora's distinguishing feature is breadth: not one model per variable or per hazard, but a single model trained on broad geophysical data and tested across four different forecast problems^[1]. The paper's contribution is the combination — one model, four domains, and a computational cost orders of magnitude below the numerical models it is measured against^[1].