AI Case Study

"After surviving the devastating flood in Kashmir, Wani returned to Stanford, where he was studying structural engineering. He began contemplating how to predict a disaster’s damage. The idea was that if city officials could anticipate which areas would be most harmed, they would be able to deploy resources faster and more efficiently throughout the disaster zone. But he had a problem: analyzing a single building using traditional structural engineering software took seven days on Stanford’s supercomputer. “We had to recreate that for the entire city” for the idea to work, Wani says. “We didn’t have seven days or seven years. We wanted to do it in three to five minutes.”

He decided to focus first on earthquakes, which are more of a threat than floods in California. Wani teamed up with fellow Stanford students Nicole Hu, a computer scientist who focuses on machine learning, and Tim Frank, an earthquake engineer, to build an algorithm that can digest data about how a building was built and how it’s been retrofitted over time. This data is combined with information on the building’s materials and surrounding soil properties to extrapolate what happens to this system when shaking occurs. Then, when a quake hits, the model absorbs new information coming from on-the-ground emergency responders, 911 calls, or even Twitter to make its predictions of the damage more accurate.

Because the model identifies patterns by looking through large amounts of data, it needs less computing power than the previous method of asking a computer to perform complex physics equations to understand how shaking will impact a structure. The trade-off is accuracy: Hu estimates that the algorithm is only about 85% accurate. With more data over time, that number will improve, but the team believes that it’s good enough to paint a broad picture of damage immediately after a quake. (Of course, they won’t know for sure until a major earthquake hits.)

Wani, Hu, and Frank started One Concern in 2015 and then released its earthquake platform, called Seismic Concern, in 2016. Seismic Concern predicts the damage caused by earthquakes on a block-by-block level and is now used by eight different municipalities, including the cities of San Francisco, Los Angeles, and Cupertino.

Seismic Concern can start generating predictions 15 minutes after an earthquake hits with about 85% accuracy, and constantly updates as on-the-ground data about the actual state of damage comes in from 911 calls and emergency response teams.

Now, the company is launching Flood Concern, a constantly evolving risk map that crunches huge amounts of data based on the physics of how water flows, information about previous floods, and even satellite imagery to approximate the depth, direction, and speed of the water–and determine which areas of a city are most at risk. Layered on top of the damage prediction is demographic data, so that emergency planners can see what areas of a city might have particularly vulnerable populations, like a significant percentage of seniors or disabled citizens. With that kind of information, planners can figure out which areas should be evacuated, where to put shelters, and what critical infrastructure–like schools or hospitals–needs the most help when flooding begins.

Even though there hasn’t been a big quake since San Francisco started using Seismic Concern, Dayton uses it for other types of emergencies as well. Earlier this year, when a high-pressure gas line ruptured during construction, Dayton used Seismic Concern’s demographic layer to look up the types of residents in the area the city planned to evacuate. He said it helped the city decide if they needed a shelter and how big it should be. “It’s not just something that we’ll trigger after the major earthquake,” he says. “It’s a tool we’re using every time we have a shelter or evacuation order.” The city has been using the platform for about two and a half years and has a multiyear contract with the startup at a rate of $100,000 a year."