Precipitation is hard for weather models. This is because it depends on small events. Simulating rain is tough for extreme weather and long times. Knowing when to plant crops or how to prepare for big storms helps people. Last year, we shared NeuralGCM. It is an open-source model. It uses AI and physics to run fast and accurate weather simulations. Our 2024 paper shows NeuralGCM makes better 2-15 day weather forecasts. It also shows past temperatures more clearly than old models. This is a big step for better climate models.
Now, in a new paper in Science Advances, we explain how NeuralGCM learned from NASA's rain data. This helps it simulate rain better. At its current size (280 km), it is better than a top weather model for medium-range forecasts (up to 15 days). It is also better than climate models. NeuralGCM better shows average rain, extreme rain (top 0.1% of rainfall), and daily weather changes.
NeuralGCM is part of Earth AI. It uses physics and AI for long-term questions. It works with AI-only models like WeatherNext 2.
NeuralGCM uses physics (gray sphere) and AI (cartoon box and umbrella) for weather. It helps simulate clouds, radiation, and rain.
Clouds Make Rain Hard to Predict
Clouds are the source of rain. They can be very small, much smaller than weather models can see. Clouds change fast. The tiny physics inside them create rain. Big models cannot handle this detail.
Models use guesses called parameterizations. These guesses use other information. NeuralGCM uses AI to learn from weather data instead of guesses.
We made NeuralGCM better at rain. We trained its AI part on satellite rain data. The first NeuralGCM used past weather data from physics models. But these models often get rain wrong. Training on them means repeating their mistakes.
Instead, we trained NeuralGCM's rain part on NASA satellite data from 2001 to 2018. NeuralGCM's special design let us train it on this data. Older hybrid models could only use simulations or reanalysis data. By training on satellite data, we found a better way to show rain using AI.
Weather models divide Earth into large boxes. Rain depends on smaller things. Traditional models guess the impact of small events using math. NeuralGCM uses AI trained on satellite rain data to better show rain.
Predicting Rain for the Next 15 Days
We tested NeuralGCM for two-week forecasts. We used WeatherBench 2. We compared it to a top model from the European Centre for Medium-range Weather Forecasts (ECMWF). NeuralGCM was much better at showing rain than the ECMWF model. This was true for 24-hour and 6-hour rain totals. The advantage was strong over land.
The left graph shows 24-hour rain forecasts from NeuralGCM (blue) and ECMWF (orange). We used satellite data to check forecasts. NeuralGCM was better for all 15 days. The right maps show the forecast scores.
Knowing about rain helps towns manage floods and dry spells. It helps farmers use water. It helps keep people safe. NeuralGCM's current size is too big for exact forecasts. But this method could be used at smaller sizes to make better forecast tools.
Rain Patterns Over Many Years
Over longer times, knowing rain patterns helps with floods, crops, and water. We looked at large areas and long times. NeuralGCM's average rain error was less than half a millimeter per day. This is 40% better than top tools used in the latest IPCC report. The error was even smaller over land.
The top graph shows rain errors. It compares old models to NeuralGCM trained on satellite rain. NeuralGCM had less error over land. The maps show rain bias for NeuralGCM and another model.
NeuralGCM was also much better at showing extreme rain. Extreme events are hard to predict. Old models often show too much light rain and too little heavy rain. This is called the drizzle problem. NeuralGCM showed rain amounts better, especially heavy rain. Predicting these damaging events is key for climate science and safety.
The top graph shows rain amounts. NeuralGCM (blue) matches observed rain well. Old models (orange, green) show too much light rain and too little heavy rain. The maps show rain bias. NeuralGCM is better at showing heavy rain.
We also checked how rain falls in a day. The Amazon rainforest gets heavy rain in the afternoon. Climate models often show rain a few hours too early. NeuralGCM shows the timing and amount of daily rain better. This is important for nature, weather, and water.
NeuralGCM shows the time of day for rain better than other tools. NeuralGCM and satellite data often show rain late in the day. Older models show rain earlier, around midday.
Looking Ahead
This is a step forward for rain forecasts. We already have real-world use. The University of Chicago and the Indian Ministry of Agriculture used NeuralGCM. They used AI forecasts to predict the monsoon season. They chose NeuralGCM after testing. They built a forecast tool used last summer.
We have made NeuralGCM open-source code. We hope people will build on it. This rain model is also free for others to use. We hope this helps make better long-term rain predictions, especially with climate change.
Learn more about Google's Earth AI work: Google Earth, Earth Engine, AlphaEarth Foundations, and Earth AI.