How AI Is Revealing the Secrets of Iran’s Nascent Centrifuge Factory
Satellites can’t directly observe the underground facility, but analysis of its surroundings yields a progress report.
Iran is perhaps 18 to 24 months from completing an underground centrifuge assembly hall at its Natanz nuclear facility, according to a recently released analysis from Center for Security and International Cooperation.
The analysis shows that Iran will be able to rebuild and extend its ability to enrich uranium despite several high-profile setbacks that Iranian officials have blamed on sabotage.
The new facility, which CISAC analysts first described to the New York Times in December, is just south of existing facilities at Natanz. It is being built deep within a mountain, where it is less vulnerable to air strikes — and hidden from imaging satellites. So CISAC analysts used AI tools from Orbital Insight to help them track construction workers at the site. They found that they could track labor fluctuations related to last year’s explosion at Natanz, and others related to the excavation and construction of the new assembly hall.
“AI- and machine learning-driven analysis is helping us to better understand where the laborers were at what time,” CISAC affiliate Allison Puccioni said.
The researchers determined the Iranians began construction on the new site last year between Aug. 30 and Sept. 14. They observed a nine-fold increase in vehicles at the site over the next three months, “indicating a significant growth in activity,” they wrote in June in the journal Janes Intelligence Review. . The paper describes the essential role that artificial intelligence played in the analysis. “Orbital Insight’s object detection algorithm counted vehicles in 84 satellite images collected between May 2018 and May 2021, offering insight into activity at the existing and future facilities at Natanz. The vehicle activity was tracked specifically in the parking lot outside the main Natanz site, as well as at the construction support facility at Natanz South, suggesting that the vehicles at each site were directly related to operations or construction activity.”
The vehicle activity dropped off during the spring, the researchers wrote. Based on that and other factors, such as the hardening of the roads and the construction of a new parking lot, they determined that “the facility at this point has, for the most part, been completed. They’re going to assemble the infrastructure, reinforce the infrastructure, and possibly commence operations of centrifuge assembly, I want to say in 18 months to two years, depending on how much infrastructure they are going to put in there,” Puccioni said.
The construction of the new assembly hall shows that Iran is “working very hard to maintain their nuclear enrichment capability” and is “reinforcing their nuclear weapons capability,” Puccioni said.
As the volume of satellite imagery explodes, along with other potentially useful data sources such as social media posts, telephonic data, etc, Puccioni said, AI will play a growing role in helping analysts reach conclusions including some that they may not have been able to reach before. “In short, I think it’s a great tool,” she said.
James Crawford, who founded Orbital Insight in 2014, said he was motivated by the sudden emergence of new satellite imagery companies such as Planet (formerly Planet Labs) and the exponential growth in new data sources.
“I knew that Planet Labs was coming, knew that Skybox was coming [and others] DigitalGlobe was doing... I could see it coming. I said look, this is going to overwhelm the human analysts. There just aren’t that many analysts… I saw a unique opportunity here to take what AI was doing in computer vision and bring it to bear on a rapidly emerging problem, basically too many satellite images.”
A few years ago, he tried to informally calculate how many analysts would be needed to analyze pictures covering the entire globe every day. His conclusion: “If you want to look at the whole Earth every day you would need 8 million analysts.”
The next challenge is finding new data science techniques that researchers might apply across different types of analysis, to see, for instance, if some aspect of machine vision could would be relevant to the problem.
“Our approaches to integrating these disparate sources, like integrating GIS data with cell phone data with satellite imagery of parking, our approaches to integrating that are fairly problem specific. So this analysis in Iran was done by the domain experts. So one of the things our data science team is working on is how do we build general tools that will allow people to do data synthesis across these different types of data,” Crawford said.