No physics? No problem. AI weather forecasting is already making huge strides.

🌈 Abstract

The article discusses the development of smaller atmospheric balloons and the use of AI-based weather forecasting models by the company WindBorne. It also covers the origins and advancements of AI forecasting techniques in the weather modeling community.

🙋 Q&A

[01] Origins of AI Forecasting

1. What are the key points about the origins of AI forecasting for weather models?

• Some early academic work on using deep learning techniques for weather forecasting began about six years ago
• Deep learning is based on neural networks, which can be trained to identify patterns and make predictions from data
• There was initial skepticism about whether AI models could outperform traditional physics-based weather models
• In 2022, promising results were presented by researchers using "graph neural networks" and the "Pangu-Weather" model developed by Huawei, which outperformed the ECMWF physics-based model in certain circumstances
• This prompted the European weather center to start developing their own AI-focused weather model, called AIFS, which began producing very promising results by the end of 2023

[02] WindBorne's Weather Modeling Approach

1. What are the key points about WindBorne's weather modeling approach?

• WindBorne developed much smaller atmospheric balloons, each with a mass of less than 6 pounds, designed to persist in the atmosphere for weeks
• By launching hundreds of these balloons per day, WindBorne can accrue data from around the world, operating the largest atmospheric balloon constellation
• To test assimilating this balloon data into forecast models, WindBorne began developing its own weather model about a year ago
• WindBorne chose to experiment with AI forecasting, as traditional physics-based models require a large amount of computing power
• WindBorne's AI weather model, called WeatherMesh, was soon performing better than traditional models on tasks like hurricane forecasting
• WindBorne now offers two products to customers: balloon data and the WeatherMesh deep-learning weather model