The research, which was published today in the scientific journal Nature, found that 56 government meteorologists considered the DeepMind predictions to be more accurate and useful in 89% of cases when compared to two other short-term forecasting methods.

Near-term weather forecasting—in this case, accurately predicting local rainfall and snow in a specific location within a two-hour window—is an important challenge in meteorology. Accurate predictions, particularly of especially intense rainfall, can save lives from flooding and other storm damage. It can also help emergency services and energy companies decide where to deploy first responders and repair crews to speed rescue and recovery efforts during natural disasters.

Making these short range forecasts, which meteorologists sometimes refer to as “nowcasting,” has long been a challenge in meteorology and it is one that weather forecasters say is becoming more difficult as climate change makes intense local rainfall both more frequent and difficult to predict using standard weather models.

Those traditional models are based on a series of physics equations that try to essentially simulate what is happening in the atmosphere. But the complexity of the equations makes it difficult to continually update the forecasts to take into account new information. In many places around the globe, including the U.K., powerful weather radar can detect precipitation at ground level at a resolution of one square kilometer every five minutes.

DeepMind’s scientists decided to create an A.I. system, based on neural networks, a kind of machine learning loosely based on the human brain, that takes in the radar images from the recent past and then generates a series of projections of what the radar image will look like in the future. The system looks only at the radar image and does not take into account the atmospheric conditions, such as humidity, barometric pressure, temperature and wind speed. The system was trained on three years of historical U.K. weather radar data, looking at this data over a 20-minute period and then trying to predict precipitation over the following 90 minutes.

The company’s scientists said that the Met Office meteorologists preferred its A.I.-generated forecasts to not only the physics-based ones, but also another kind of A.I.-based weather prediction system in which the software is also trained from historical data. But this other A.I. method cannot produce the detailed, fine-grained images that DeepMind’s software does.

Niall Robinson, head of partnerships and product innovation at the Met Office, said the government forecaster was considering how it may use the DeepMind research in its on-going forecasts. He also said the agency was looking at other possible collaborations with DeepMind on the impact of climate change and other aspects of weather forecasting.

More tech coverage from Fortune:

• Europe wants one device charger to rule them all—and it doesn’t come from Apple
• Once an oddity of Japan’s digital culture, VTubers have become a global hit—and brands want in
• Facebook puts Instagram Kids on hold amid growing concerns
• Frustrated carmakers upend industry after chip shortage shatters their faith in suppliers
• Remitly CEO on his money-transfer company’s 10-year journey to an IPO