OTTAWA — Companies in the business of forecasting the weather are using artificial intelligence to make predictions quicker and more accurately, a valuable service for clients whose fortunes depend on if it rains or shines. 

“Almost everything we touch has now gone the way of AI,” said Nana Banerjee, CEO of Pelmorex, the Oakville, Ont.-headquartered firm that runs the Weather Network.

Tens of millions of people each month use the company’s apps, websites and television channels to check the forecast. Over the last two years, Pelmorex has adopted a type of AI called graph neural networks (GNNs), which map out the links in historical weather data to find patterns and make predictions. The firm’s new deep learning models produce forecasts in minutes, and are more accurate than the physics-based systems they replaced, Banerjee claimed. “That’s a huge leap for us in this sector.”

Meteorologists have traditionally used mathematical equations to predict how current conditions could change due to wind, pressure, temperature and other factors. Supercomputers running these physics-based models typically take several hours to produce a forecast, according to Banerjee. 

Pelmorex’s clients are willing to pay for more up-to-date predictions. Wind, precipitation and other weather conditions “change around very rapidly,” said Banerjee. “So the value is immense.” Customers whose businesses depend on forecasts include retailers stocking shelves with umbrellas or air conditioners; traders buying and selling commodities; and airlines plotting flight paths around air pockets and through cloud cover. 

Banerjee claims Pelmorex is the first commercial organization successfully using GNNs for weather forecasting. “We don’t know of our competitors even trying any of it,” he said. Google’s DeepMind has also used GNNs to improve weather predictions, but it doesn’t currently sell the service. 

Pelmorex has filed for patents on the technology to prevent rivals from copying its approach. The company also plans to license its models to broadcasters who run forecasts but don’t generate them in-house.

The Pelmorex team of AI developers and meteorologists working on its weather prediction systems now makes up about a fifth of its over 500-person workforce; the rest work in sales, marketing and operations. The privately held company doesn’t disclose revenues. In addition to selling data to businesses, Pelmorex makes money by running targeted ads on its weather apps, websites and television channels. 

The firm is also using generative AI to communicate with users. Pelmorex has created AI ‘avatars” of its meteorologists, letting it produce personalized video forecasts based on local weather conditions without having to record millions of different versions. It first rolled out the system on televisions in Home Hardware stores. 

While Pelmorex is using AI to produce short-term forecasts more quickly and accurately, San Francisco-based startup ClimateAI is employing the technology to make longer-term predictions. “Most of the weather companies stay within two weeks,” said CEO Himanshu Gupta. By contrast, ClimateAI predicts the likelihood and impact of extreme events like heatwaves, wildfires, droughts and hurricanes up to six months out.

The firm focuses on conditions in the oceans, which have a strong influence on weather in the long term. ClimateAI’s deep-learning models find and monitor patterns in data like the salinity of the water and the temperature at different depths. Then the system connects those trends to information that government agencies publish, resulting in a more accurate forecast.

Still, like ChatGPT, neural networks focused on weather sometimes make things up. “It might tell you, ‘There’s an extreme heatwave and temperatures might cross 70 degrees celsius,’” Gupta said. (The highest temperature on record is 56.7 °C, in California in July 1913). To avoid such hallucinations, ClimateAI introduces limits to what the system can forecast that are based on real-world physics.

The firm uses a second set of deep-learning models to predict how crop yields will respond to extreme weather events. Its clients include agricultural producers like Dole and consumer goods firms like Suntory, which use the firm’s software to monitor risks to their farms or supply chains.  

Take California nut farmers. “Almonds are very sensitive to drought risk, pistachios are not,” said Gupta. The opposite is true for the risk of a warm winter. With its long-term weather forecasts, ClimateAI can advise food firms on the best time to plant and harvest a crop, or the ideal location for their farms.

The startup has about 40 staff, and has raised US$37.5 million to date. It’s backed by Canadian financiers in the Public Sector Pension Investment Board, Toronto-headquartered Radical Ventures, and Vancouver-based Yaletown Partners. ClimateAI is set to expand into manufacturing sectors and supply chains beyond food. “Our models keep on getting better,” Gupta said.