Jun 26, 2013
Dark Sky’s creators Adam Grossman and Jason LaPorte on how they developed their breakthrough meteorological technology.
Carolyn Marks Blackwood
Adam Grossman, Jack Turner, and Jason LaPorte had an idea for a simple weather app, with one major twist. Instead of predicting weather by the day or week, their app would predict rainfall minute by minute-one hour into the future.
Since launching Dark Sky, their Troy, New York-based company, in 2012, the co-founders have plenty of satisfied (and dry) customers. The $4 app has been downloaded more than 100,000 times. Here’s how it all started.
Getting Drenched on I-90
Grossman: I was driving to Cleveland with my girlfriend to visit my family. We were on I-90 and stopped at a rest area to get some gas and food. When we were inside, it started pouring–just one of those summer thunderstorms that sneak up on you and start pouring buckets. We didn’t want to go back to the car, because we would have gotten drenched. The question was, “When is this rain going to let up? Do we have to wait 10 minutes or an hour?”
I got out my iPhone and pulled up the weather. It didn’t say when the rain would stop. It just said, “Forty percent chance of rain.”I spent the next two weeks on my laptop trying to figure out how to pull in National Weather Service radar data to see if there was a way to solve this problem.
Laporte: When Adam told me about the idea, my first reaction was, “Wow, why isn’t anyone else doing that?” It’s something people wanted–even if they didn’t know they wanted it. In the popular consciousness, it was still very much a science-fiction sort of thing. You think of Back to the Future; they have this funny watch that tells you it’s going to stop raining in five seconds.
We’re Not Meteorologists. So, How Do We Do This?
Grossman: In 2011, we did a Kickstarter and raised $40,000–enough to create the app. None of our family or friends thought it would actually work. When you tell someone that you’ve created something that can predict the rain, they’ll smile and nod. If it’s your mom, she’ll say, “That’s great.” But I don’t think they thought it would work.
None of us are schooled in meteorology. We’re computer nerds: I’m a Web developer, Jason studied computer science, and Jack makes weird computer art.
So, the challenge was, OK, how do we do this? Conventional weather forecasting often involves physical models simulated using supercomputers.
The atmosphere is modeled using fluid dynamics. Running these simulations is a slow process, which is fine for seven-day forecasts. But to make our short-term forecasts in real time, we had to take a different approach.
Laporte: We may not be well versed in meteorological methods, but we are well versed in computer methods. There are a lot of computer-vision algorithms and statistical algorithms you can use to predict how a storm is moving. I guess I would call them glorified cheats.
Grossman: We rely on data from Doppler radar stations–there are more than 150 in the U.S. The raw radar data is from the National Weather Service. It’s free.
The U.S. kind of gleefully gives it out to everybody to see what they can do with it. We convert the Doppler data into images and then explore it in the same way a human meteorologist might.
A trained person can look at a series of radar images and see how the storm is moving and evolving, which areas are increasing or decreasing in intensity, and where the rain is headed. We basically simulate this process, using algorithms. Because it’s automated, we can watch millions of square miles simultaneously and in real time.
People Love Us–the Weathermen, Not So Much
Laporte: Meteorologists? Yeah, they don’t like us very much.
Grossman: Well, they’re sort of split into two camps. There are the meteorologists that think Dark Sky’s really cool and want to understand how it works.
Then there’s the other camp, which really doesn’t like what we’re doing. Meteorologists know a lot more about the weather than we do, so it’s useful to hear what they have to say.
When we’re not in foul moods, we take the time to explain how our technology works. We say, “We take a statistical approach, not a physical-modeling approach–and here are the upsides and the downsides.”
Our algorithms really work only up to an hour in advance. Beyond that, you really need a physical model to figure out what’s happening. The downside to a physical model is that it’s impractical. You’d need a supercomputer cluster to run that.
Building a Weather Brand
Grossman: Recently, we launched a full-featured online weather service, Forecast.io. It’s our answer to Weather.com, AccuWeather, and Weather Underground.
Laporte: People are happy to switch to it. It doesn’t smother you with ads or news stories. It’s getting 70,000 new users a day. You can tell how many intense weather systems there are because it mirrors how many users we have on our site.
Grossman: Right now, we’re still paying the bills with sales of Dark Sky. The reaction from customers has been phenomenal.
Last summer, I was sitting on the banks of Lake Erie next to a little kid and his dad. I saw on my phone that it would rain. I told the kid, “It’s going to rain in exactly six minutes.” He laughed. Exactly six minutes later, the drops started falling. The kid’s eyes lit up. He asked, “How did you do that?” I looked at him and said, “I’m from the future.”