AI predicts extreme weather 5,000 times faster! Microsoft launches Aurora to predict global storms with the eyes of AI

Since the beginning of human history, we have been obsessed with predicting the weather and deciphering the "language of the sky" in various ways. We slowly discovered that vegetation and clouds seem to be related to the weather. This is not just Because the need of human beings to engage in production is also the need of human beings to sing to the strong wind and recite poems under the moonlight.

The Storm Singer in "A Song of Ice and Fire" predicts weather and storms through singing and chanting, and people also fantasize about having the superpower of "changing the weather."

Recently, weather experts and weather forecasts have made us unable to escape from embodied experience and the physical world, but now, AI has changed the situation.

Fine-tuned content: Microsoft released Aurora, its first large-scale atmospheric basic model, which can learn from data and make predictions, showing amazing accuracy and efficiency.

Changes are not just brought about by one company, but are global.

The European Center for Medium-Range Weather Forecasts, the world's top numerical weather forecasting organization, maintains an extremely rich data set, providing strong data support for AI weather forecasting. This data set contains data information from multiple dimensions such as the atmosphere, ocean, and land in Europe and surrounding countries and regions. These data have been carefully observed, analyzed and modeled to form the

In the future, a computer may be able to capture the global "changes" without the need for physics.

The impact doesn’t stop there. If we can already use AI to predict global weather, will “modeling” the earth be far behind?

Microsoft releases the first large-scale atmospheric basic model

Extreme weather events occur frequently around the world. In the face of sudden storms, human beings appear particularly small. .

Always worrying about extreme weather exposes the limitations of current weather forecast models and highlights the need for more accurate forecasts in the face of climate change.

A pressing question arises: How can we better predict and prepare for such extreme weather events?

A recent study by Charlton Perez and others highlights the challenges even the most advanced artificial intelligence weather prediction models face in capturing the storm’s rapid intensification and peak wind speeds.

To help address these challenges, a Microsoft research team developed Aurora, which means "Aurora", a cutting-edge artificial intelligence-based model that can extract data from large amounts of atmospheric data. extract valuable insights.

Picture

Paper address: https://www.microsoft.com/en-us/research/publication/aurora -a-foundation-model-of-the-atmosphere/

Aurora provides a new approach to weather forecasting that could transform our ability to predict and mitigate the effects of extreme events.

Flexible 3D atmospheric basic model

Pictures

During pre-training , Aurora is optimized to minimize losses on multiple heterogeneous data sets with different resolutions, variables, and stress levels. The model is fine-tuned in two stages: (1) fine-tuning pre-trained weights in a short period of time; (2) long-lead-time (rollout) fine-tuning using low-rank adaptability (LoRA). The fine-tuned model will be used to handle various operational forecast situations at different resolutions

Although the parameter size is only 1.3B, Aurora has experienced various weather and climate conditions for more than one million hours. It is trained in simulations, which gives it a comprehensive understanding of atmospheric dynamics.

Therefore, the model can perform various prediction tasks excellently even in data-scarce areas or extreme weather conditions.

By operating at a high spatial resolution of 0.1° (approximately 11 kilometers at the equator), Aurora is able to capture the intricate details of atmospheric processes, providing more accurate operational forecasts than ever before , while the computational cost is only a fraction of that of traditional numerical weather prediction systems.

According to researchers’ estimates, Aurora’s calculation speed is increased by about 5,000 times compared with the Integrated Forecasting System (IFS), the SOTA in the numerical prediction system world.

In addition to its stunning accuracy and efficiency, Aurora stands out for its versatility.

The model can predict a wide range of atmospheric variables, from temperature and wind speed to air pollution levels and greenhouse gas concentrations.

Aurora's architecture is designed to handle heterogeneous gold standard inputs and generate predictions at varying resolutions and fidelity levels.

The model consists of a flexible 3D Swin Transformer and Perceiver-based encoder and decoder, capable of processing and predicting a range of atmospheric variables across space and pressure levels.

By pre-training on large amounts of diverse data and fine-tuning for specific tasks, Aurora learns to capture the intricate patterns and structures in the atmosphere, allowing it to perform even when fine-tuned for specific tasks. It can perform well with limited training data.

Quick prediction of atmospheric chemistry and air pollution

Pictures

Aurora outperforms running CAMS on a number of objectives: (a) Aurora predicted total NO2 column samples compared to CAMS analysis; (b) Aurora’s latitude-weighted mean relative to CAMS Root Square Error (RMSE), negative values ​​(blue) mean Aurora is better Monitoring services (CAMS) data are highly heterogeneous, which is a notoriously difficult task.

Aurora leverages its flexible encoder-decoder architecture and attention mechanism to effectively process and learn from these challenging data, capturing the unique characteristics of air pollutants and their relationship with meteorology relationship between variables.

This enables Aurora to produce accurate five-day global air pollution forecasts at 0.4° spatial resolution, outperforming state-of-the-art atmospheric chemistry simulations on 74% of all targets, Its excellent adaptability and potential in solving a variety of environmental forecasting problems are demonstrated, even when data are sparse or highly complex.

Data diversity and model scaling improve atmospheric forecasts

The study also found that compared to training on a single dataset, Pre-training on different data sets can significantly improve Aurora's performance.

By integrating data from climate simulations, reanalysis products and operational forecasts, Aurora can learn more powerful and general representations of atmospheric dynamics.

Precisely because of its size and diverse pre-training datasets, Aurora is able to outperform state-of-the-art numerical weather prediction models and specialized deep learning methods across a variety of tasks and resolutions .

Picture

Picture

Picture

Pre-training on different data and scaling up the model improves performance. Each doubling of model size reduces training loss by 5%

## As a direct result of #Aurora's scale, performance is better than the best professional deep learning models, both in terms of architecture design and training data corpus, as well as pre-training and fine-tuning protocols.

To further validate the benefits of fine-tuning large models pre-trained on multiple datasets, the Microsoft team compared Aurora to GraphCast, which was pre-trained only on ERA5 , is currently considered the most proficient artificial intelligence model with a resolution of 0.25 degrees and a prediction time of up to five days.

In addition, the researchers also included IFS HRES (the gold standard for numerical weather prediction) into the comparison.

The results show that Aurora outperforms both GraphCast and IFS HRES when comparing analysis, weather station observations, and extreme values.

Picture

Picture

Picture

Aurora outperforms GraphCast on the vast majority of targets

A paradigm shift in earth system modeling

Aurora’s impact extends far beyond atmospheric forecasting.

By demonstrating the power of fundamental models in Earth science, this research paves the way for the development of comprehensive models that encompass the entire Earth system.

The ability of underlying models to excel at downstream tasks in data-scarce situations will enable access to accurate weather and climate information in data-scarce regions, such as developing countries and polar regions. More democratization.

This will have far-reaching impacts on sectors such as agriculture, transportation, energy harvesting and disaster preparedness, allowing communities to better adapt to the challenges posed by climate change.

No physics required? Huge Progress in AI Weather Forecasting

Changes are coming so fast, like tornadoes, that the weather forecasting community is undergoing major changes.

The ultimate goal is revolutionary: using new AI-based methods, weather forecasts can be run on desktop computers!

Over the past 18 months, weather forecasting has emerged as one of the most promising AI applications, and recent advances have caused a huge stir in the meteorology community.

This is thanks to a secret weapon: an extremely rich data set.

The European Center for Medium-Range Weather Forecasts (ECMWF), the world's leading numerical weather forecasting organization, maintains a set of data sets on atmospheric, land and ocean weather, which are updated every day around the world. Recorded every few hours, with data going back to 1940.

Pictures

Data from the past 50 years is especially abundant after global satellite coverage. This dataset is called ERA5 and is publicly available.

ERA5 was not created specifically for artificial intelligence applications, but ERA5 has played a huge role in the development of artificial intelligence weather applications.

Computer scientists won’t really start seriously using this data to train artificial intelligence models to predict weather until 2022.

Since then, the technology has grown by leaps and bounds. In some cases, the output of these models is already better than the global weather models that scientists have spent decades designing and building, and which require some of the world's most powerful supercomputers to run.

Matthew Chantry, head of artificial intelligence forecasting work at the European Meteorological Center ECMWF, said in an interview, "It is obvious that machine learning is an important part of future weather forecasting."

ECMWF is recruiting talent to develop machine learning-based Earth system simulations

The origin of AI weather forecasting

Some early academic research using deep learning techniques based on neural networks for weather forecasting began about 6 years ago.

At first, computer scientists were not very optimistic about whether this approach would work because it was so different from the science of weather forecasting that had been developed over decades.

When the time comes to 2022, people have slightly let go of their doubts about AI models.

First, physicist and data scientist Ryan Keisler showed some preliminary results using "graph neural network".

Picture

Paper address: https://arxiv.org/abs/2202.07575

Afterwards, The “Pangu-Weather” model proposed by Chinese scientists was directly listed in Nature.

The results show that it even outperforms today's strongest physics-based model - ECMWF - in some cases.

Picture

Paper address: https://www.nature.com/articles/s41586-023-06185-3

This sent shockwaves through the community of scientists who use deep learning techniques and weather modeling.

Soon, European scientists began to create an operating model based on the research results of other deep learning models, which did not take too long.

By the end of last year, the new Artificial Intelligence Integrated Forecast System (AIFS) had produced "very promising" results. This spring, European forecasters began issuing real-time forecasts.

At present, physics-based weather models are still indispensable. They are incredibly powerful tools that significantly improve our ability to produce five-, seven- and occasionally 10-day weather forecasts for major events and are trusted by forecasters around the world.

But what does the future look like? Maybe in ten years, AI will be in charge of everything in the weather field.

Reference:

https://www.php.cn/link/3b2f3a493d32e9aca1df90ef35b587e7

https://www.php.cn/link/3d65824c0a13e8417758ea807a431500

The above is the detailed content of AI predicts extreme weather 5,000 times faster! Microsoft launches Aurora to predict global storms with the eyes of AI. For more information, please follow other related articles on the PHP Chinese website!