Last spring, when the Department of Energy released a report outlining the key areas to prioritize artificial intelligence deployment in the energy sector, weather forecasting was front and center.
Of the four topic areas the agency highlighted, AI for weather forecasting was deemed essential for both grid operations and reliability, as well as grid resilience. In the months since that report — and the years since the release of ChatGPT — AI weather forecasting has garnered growing interest and funding from around the energy ecosystem. Major companies and startups alike have gotten in on the space — even as cuts at the federal government level make certain kinds of forecasts more challenging.
A year ago, the former Google X executive Julian Green launched his company Brightband out of stealth with a $10 million Series A under its belt. And last month at New York Climate Week, Brightband convened a panel of experts to assess the state of the sector, and discuss the future of weather forecasting. I moderated the conversation with leaders in the space from tech giants Google and Nvidia, as well as startups Sorcerer and Earthmover.
Below is an excerpt of our conversation, edited for brevity and clarity.
What is the current state of AI weather forecasting?
Julian Greene, Brightband: Right now, AI forecasting is better [than traditional forecasting] at two-week global weather, and it is maybe 10% or 20% more accurate. That translates into half a degree of temperature, or a few degrees of temperature over maybe a seven-day timeline. It translates into a few tenths of wind speed accuracy over that same time.
So it is more accurate on global averages…I think we have yet to see all the other ways that it is amazing and useful. It’s already made 70 years of progress in two years, and we’re just beginning.
Mike Pritchard, Nvidia: It’s amazing. I’m an atmospheric scientist by training. I spent a dozen years as a climate science professor — and [then with AI] a whole new way to simulate the atmosphere was born and no one expected it, and it has completely new properties and is fascinating. I think we’re still in our infancy of understanding the implications.
AI weather systems should be able to use more observations much more flexibly, [but] we don’t know the limits of the predictability of the system.
As to whether weather forecasting will shift to being entirely AI-based, I think you should look to the most credible, most experienced practitioners of operational weather prediction for answers. The European Center for Medium-Range Weather Forecasts is a great example. They’re intentionally maintaining a physics model and are building an ML model. The cross-pollination there is fascinating. One of the things I’ve been most impressed with is they take their high-resolution physics models…and they’re marrying them with the AI model [to create this] cyborg physics-AI model.
But it’s not a coincidence — and [in fact is] probably logical right now — that most meteorological agencies, serious ones, are maintaining both physics models and AI models. Then again, the scores are unequivocal. It’s not that hard to score an AI weather model and to convince yourself that there is real skill beyond the capacity of today’s best physics models for the medium range.
Denise Pearl, Google: I think there is this shift from traditional ways of doing weather forecasting to AI weather forecasting. [Within Google,] what’s been interesting is you build these amazing new models like Graphcast or Gencast, but once you start using them, you start to figure out that you do need the plumbing. And you do need the additional data, and you need all of these amazing companies that can help you process that and integrate it into your decision support.
Where are AI weather forecasting models being adopted most rapidly?
Austin Tindle, Sorcerer: From my perspective, the biggest places that you’re seeing widespread adoption is in those places where it’s been typically too expensive to run numerical weather prediction models. So there are a lot of smaller scale use cases. There’s a lot of research use cases. There are a lot of people that can use or develop on or play with these models that wouldn’t have been able to otherwise. There’s a lot of experiments that you could run that would have been too expensive.
Ryan Abernathy, Earthmover: With AI weather models, very, very small teams of smart people — especially with some capital to deploy — can run those models, and can get very unique and special fine-tuned value out of them. And this is driving a real gold rush right now around looking for ways to leverage those, specifically in energy markets and commodities. And that’s super exciting from a scientific point of view, and it’s doing things in the market as well.
Denise Pearl, Google: We are afforded the opportunity to talk to customers across all industries — including healthcare, retail, utilities, all of them — and when you sit down and ask them about how they’re thinking about sustaining their business in the future, weather is often a consideration.
I think that in the past, they were restricted or limited by how they could insert weather visibility into their decision-making process. But now, because there have been efficiencies gained on the infrastructure, what took days and millions and millions of dollars now you can do in minutes. And this idea of having maybe the opportunity of not just one prediction, but 70 potential predictions about what could happen, just gives people a lot of opportunity to imagine how they can do new things.
Are we going to see the spread of super-customized AI weather models?
Mike Pritchard, NVIDIA: I think there’s going to be a biodiverse universe of forecast models, because you mentioned it’s so easy now for small teams to create new ones and novel ones.
From an atmospheric science perspective, as long as multi-day prediction is the time horizon, global scale is a necessary condition. So it’s sort of a fool’s errand to try to rigidly fine-tune an AI weather model for one region and one metric…but there’s all kinds of other things that could be fine-tuned. Down-scaling can be fine-tuned, generative sampling can be fine-tuned.
When people condition weather models for samples that are guided, their conditions are totally different. You know, the truckers care about one thing, and the energy traders care about a completely different thing. But I think interactive, generative, conditional weather models are going to be especially biodiverse. So, yeah, I don’t see one giant model. I see a growing ecosystem.
Denise Pearl, Google: It’s [already] happening. It’s absolutely happening. I don’t see people asking for something in a specific place at a specific time, but what I see them saying is, how do we incorporate this new ability to have better visibility into weather outcomes across our global supply chain, and how do we make that part of our decisionmaking? It’s not the only decision-making point, but because of the cost-effectiveness and the accuracy, they can lean on it more than what I’ve seen before.
So that’s the go-to-market that customers are asking for now. Is it custom? In a lot of cases — we work with a really large ecosystem of partners — absolutely. Should it get a little bit more streamlined and repeatable? I think it will over time. But this is happening today, we are getting these requests.
What are the biggest forces impacting the future of AI weather forecasting?
Austin Tindle, Sorcerer: The biggest thing from my perspective is that we need more people working on these things. There are a handful of other AI and weather companies that are at the earlier stage, but there could be 10 to 100 times more. I think we’d be able to dramatically increase the pace at which we’re building better tools and better models.
Moving the industry forward [will require] a really interesting and hard intersection of skills, which we haven’t caught up to quite yet. You need the machine learning skill set. You need the atmospheric science skill set. In our case you need aerospace and a bunch of others as well. It’s an intersection of very hard skills.
Ryan Abernathy, Earthmover: I don’t think we’ve talked enough about data. There’s a lot more that we can still get out of data drive weather modeling, that can only happen with more data. That’s what we learned from LLMs, and it’s no different in AI weather predictions. Data collection, at least in the U.S., is actually at risk. We’re at risk of having fewer weather observations, whether that’s through canceling key satellite missions or reducing the National Weather Service capability to do soundings.
What I want to know is, can the AI models tell us where we need more observations in order to increase skill for some particular metric, and how do we get them? Ultimately, I think this is what’s going to drive the industry. The technology is moving so fast, but it’s ultimately limited by the data that we have about the Earth, and we’ve got to keep that growing.
