Can artificial intelligence help contain contagion in the sky?
CU Boulder researcher receives NSF funding to study COVID-19 spread in airplane cabins
At least two companies, AstraZeneca and Moderna, are preparing to enter “phase three” human trials of a vaccine for the SARS-CoV-2 virus this summer, with other companies not far behind, Soumya Swaminathan, chief scientist for the World Health Organization, announced on June 26.
But even if everything goes right, the U.S. Department of Health and Human Services doesn’t expect to have enough doses available for the American population until at least January 2021.
In the meantime, people will continue to be exposed, perhaps especially those who travel by air.
“Until we have a vaccine or a cure, this virus is not going to go away. But at some point, all of us are going to have to start traveling,” says Maziar Raissi, assistant professor of applied mathematics at the University of Colorado Boulder, a specialist in machine-learning and deep-learning algorithms.
With that conundrum in mind, Raissi submitted a research proposal to the Division of Mathematical Science at the National Science Foundation to use “computational mathematics and state-of-the-art physics-informed deep learning techniques” to model, analyze and predict how air flow in an aircraft cabin will influence contagion.
“As we know, this virus is airborne,” he says. “Studying how it travels in the circulating air flow inside an airplane is very important.”
Raissi, who joined the CU Boulder faculty in January, pulled his initial proposal together in just four days this spring. On June 4, the NSF notified him that the project had been funded.
“I have not seen this kind of turn around ever in my career,” says Keith Julien, chair of the Department of Applied Mathematics.
Because of the high cost and difficulty of studying air flow in an actual airplane, Raissi will use physics-based models to mathematically describe how air moves inside the “complicated geometry of an airplane cabin.”
The research will “efficiently combine” physics-based models for “fluid dynamics, scalar transport, epidemiology and airborne infection to analyze the spread of COVID-19 within a closed system such as an airplane,” he wrote in his proposal.
“We’ll analyze the data and use it for prediction of what will happen if somebody coughs next to us, for example,” Raissi says. “Is it going to be filtered out or not? Will it make a difference if the (coughing) person is sitting next to us, or five rows in front of us?”
Raissi says the data will be useful in determining aspects like how to seat passengers, passenger density, and the rate of flow from aircraft ventilation systems. For example, he says, the data might indicate how to better direct air flow to reduce the risk of infection.
He plans to produce open-source software that is “agnostic to geometry,” he says. In other words, by plugging in different variables, the software should be able to predict how air flow affects the virus in myriad different spaces, from supermarkets to college campuses to trains.
Raissi came to CU Boulder after working as a senior software engineer with the Silicon Valley firm NVIDIA. Prior to that, he received his PhD from the University of Maryland College Park and carried out postdoctoral research at Brown University.
How moral do we want to be, versus how much we maximize profits—even humans have difficulty with such choices"
“I received the offer from CU Boulder before getting the offer from NVIDIA,” he says, but deferred coming to Boulder for a semester in order to gain experience in private industry.
The field of machine-learning went into hyperdrive following the 2012 publication of work on artificial neural networks by British-Canadian computer scientist and cognitive psychologist Geoffrey Hinton, known as the “godfather of deep learning.”.
“That has led to a lot of cool innovations,” Raissi says.
“Physics-informed deep learning is the idea that you replace data with physics, the laws and equations discovered by Newton, Einstein and other super geniuses. … This means the machine doesn’t have to learn everything from scratch. Rather, we can transfer human knowledge gathered through centuries of scientific discovery to a deep neural network.”
With the continuing proliferation of self-driving cars, drones, robots and other artificial-intelligence machines, it has become critical to ensure that deep neural networks will be safe, secure and essentially unhackable. As machines become more and more autonomous, it may even become necessary to program a kind of “morality” into their learning and decision-making. For example, a self-driving car may have to decide whether to hit a person who has jumped into the street or risk an accident that could injure or kill a passenger.
“We want to create a balance between maximizing reward or minimizing the loss. But at the same time, we have to be morally correct,” Raissi says, describing such situations as “a multi-objective optimization problem.”
“How moral do we want to be, versus how much we maximize profits—even humans have difficulty with such choices.”
With COVID-19, Raissi says that, “we have to carefully balance the tradeoff between saving people’s lives and their livelihoods.”