“I always knew there was a big problem,” he said of the climate change-fueled dangers to agricultural commodities like cotton and the resulting impact on livelihoods. “The question was, ‘Do we have the technology to solve this?’”
That eventually led him to co-found ClimateAi, a Bay Area-based startup that aims to help farms and other businesses prepare for a hotter, more disruptive climate using the power of artificial intelligence. By harnessing machine learning models, the company says its customers can anticipate and prepare for climate risks to their supply chains and operations over periods ranging from weeks to seasons. That timeframe, Gupta said, is traditionally a modeling blindspot for climate forecasters without access to AI-powered tools.
Forecasting extreme weather events like hurricanes and heat waves for hyper-specific geographies weeks to seasons ahead is difficult to do accurately. But having sufficient time to prepare for a natural disaster can be the difference between averting catastrophe and not.
If you tell a company’s supply chain manager a heat wave is approaching that may impact their cotton supply chain a week before it arrives, Gupta said, that isn’t enough time to move inventory or prepare in other ways. “But if you were to tell them that there’s a higher risk of a heat wave in the next season in this area where we think that cotton yield might be down, this becomes very actionable for them,” he added.ClimateAi created deep learning models, incorporating oceanic parameters like sea surface temperature and oceanic salinity, to better predict climate risks, even for regions with limited weather data. “The secret sauce is that oceans are the long-term memory of Earth‘s climate that drive medium- to long-term weather of a particular location, and we can tap into that memory using machine learning approaches,” Gupta said. Those approaches include ingesting data from a variety of sources and using it to search for patterns and make predictions. ClimateAi pairs its deep learning models with both government-run forecast models like those from the US National Oceanic and Atmospheric Administration and numerical models like those used by the Intergovernmental Panel on Climate Change to achieve what it claims are higher accuracy and forecasting reliability.
The AI boom has led to a surge of startups using AI tools to help companies and institutions prepare for climate impacts. Jupiter Intelligence, whose customers have included Con Edison and BP, also uses models to analyze climate-related financial risks. Google has a research group that applies AI to weather and climate challenges. In April, the climate intelligence company Tomorrow.io, which also creates AI-powered models, became the first private company to launch its own weather satellites carrying radar that it hopes will help improve its modeling.
But ClimateAi’s differentiator, according to co-founder and former chief technology officer Maximilian Evans, is what it calls “impact modeling.” The company essentially creates two different types of models that can talk to each other. One is climate-focused, using inputs from satellites, radar stations and both public and private weather station data. The other is business-focused and incorporates historical customer data for variables such as crop yields. Combining them can provide a clearer view of, for example, the impact of heat risk on crops.
Evans said the goal isn’t just to provide an accurate climate forecast; it’s to provide “decision accuracy” that allows companies to feel comfortable using the forecast to take action.
Most of ClimateAi’s approximately 40 global customers are in the food and agriculture industries, but Gupta is hoping to expand further into other sectors including manufacturing and energy.
Some of the company’s existing customers are already starting to see the benefits of using the technology. Driscoll’s Inc., a leading supplier of fruit and other produce, has been using ClimateAi’s tools to help its operations. The company is one of the world’s largest berry suppliers, generating $4 billion in annual revenue for its strawberry, blueberry, raspberry and blackberry production.
Though the berries are cultivated across North America, about half of the company’s farms are located in California, a state prone to bouts of extreme heat, drought and wildfires. Berries are difficult to grow in very hot conditions and are sensitive to weather changes. Strawberries, for example, suffer from discoloration and poor root development in very hot weather. Drought and a lack of water result in smaller berry sizes and lower crop yields. That’s why the company is using ClimateAi’s technology to help forecast where to site future berry farms and determine its long-term growing strategy.
Typically, scouting a new location takes many years and involves sending local teams out to collect soil and weather data. But with the help of predictive tools like ClimateAi, companies can find new locations “in a matter of minutes,” Evans said.
“We’re seeing the impact of climate change. In California, we have all this volatility,” said Soren Bjorn, Driscoll’s president of the Americas. “This is like an insurance policy.”
Despite the promise, there are challenges AI models face when predicting future local climate shifts.
Data in some parts of the world remains scarce, which can be a problem for companies based outside of the US and Europe, said David John Gagne, a machine learning scientist at the National Center for Atmospheric Research. Some locations have more observations from surface instruments and weather balloons than others, so training a model on mostly US data and applying it to make decisions for a company operating in South America, for example, could make for a less accurate model given the different climate.
“There have been definitely forecast improvements, but there’s still a lot of uncertainty in the process,” Gagne said. “The uncertainty, in some cases, may be more than what the people who have to make decisions would like.”
Gupta acknowledged that this is a problem for ClimateAi. The company lost a Brazilian customer because ClimateAi’s predictive modeling could not be applied on its farm, an issue Gupta attributed to Brazil’s low density of weather stations.
“Probabilistically, there’s always going to be a chance that we’re wrong,” Evans said. But the status quo and alternative is relying on historical averages, which are increasingly worse predictors of what will happen in a climate-change impacted future.
“So it’s not about ‘what’s the flip side,’ but ‘what is the best alternative?’” he said.
