This summer, 38 million farmers in India received AI-powered forecasts of the start of the monsoon season, helping them make more informed decisions about when to plant their crops for the season. These forecasts were powered in part by NeuralGCM, a Google Research model, which combines traditional physics-based modeling with machine learning for improved simulation accuracy and efficiency.

An AI model to predict weather and climate

For years, weather and climate models have been costly and complex, often requiring a supercomputer to run. Our teams at Google Research wanted to see if we could build these models more efficiently and more accurately, leading to the creation of NeuralGCM. Unlike traditional models that rely purely on hard-coded physics, this AI-driven model is trained on decades of historical weather data to infer patterns and learn from past events, while also using physics. Crucially, it’s designed to be flexible and efficient — it can run on a single laptop, making high-quality forecasting more accessible to the scientific community.

A collaboration with University of Chicago

When we open-sourced NeuralGCM, we hoped the community would use this new tool to power their own innovative applications. The University of Chicago’s Human-Centered Weather Forecasts Initiative did just that. They recognized that one of the most impactful, yet increasingly challenging, decisions for Indian farmers is when to put seeds in the ground. Hundreds of millions of smallholder farmers across the tropics depend on information about when the rainy season, known as the monsoon, will come each year. However, accurate forecasting of when the monsoon will begin, especially at long lead times and at local scales, has remained a century-old challenge.

By rigorously testing several AI weather models, the University of Chicago team found that NeuralGCM, when blended with other advanced models like the European Centre for Medium-Range Weather Forecasts (ECMWF)’s Artificial Intelligence/Integrated Forecasting System (AIFS) and historical data, was the right tool for the job. It accurately predicted the onset of the Indian monsoon up to a month in advance, even capturing an unusual dry spell in the progression of the monsoon.