Each year, Penn Engineering, The Mack Institute at the Wharton School, The Penn Venture Lab, and the Penn Center for Innovation host the Y-Prize competition. Starting with technologies developed by Penn Engineering researchers, contestants are charged with finding significant real-world applications and building business plans around them.
This year’s technologies are “Solar Powered Aerial Vehicles,” which use only sunlight to create lift force, and “Physics-Informed Neural Networks,” algorithms that can utilize physical principles to infer unknowns.
Teams enter the competition by pitching a real-world problem and product based on the capabilities of one of these technologies. One winning team will receive $10,000 to kick start the development of their prototype and help get their ideas to the market. For example, last year’s Y-Prize winners, Ossum Technologies, used the “steerable needle” technology developed by Mark Yim, Asa Whitney Professor of Mechanical Engineering and Applied Mechanics, to create a tool for stabilizing bone fractures.
Visit the Y-Prize competition site to learn more.
Solar Powered Aerial Vehicles
The first technology for this year’s competition is “Solar Powered Aerial Vehicles,” ultra-thin structures that levitate using the power of photophoresis. A physical phenomenon wherein microscopic particles (dependent on their properties) either move toward or away from a given light source, photophoresis allows the aerial vehicles to levitate without any moving parts.
These vehicles, developed by Igor Bargatin, Associate Professor of Mechanical Engineering and Applied Mechanics and colleagues, have been composed using two different materials. One model is made using two thin plates of aluminum oxides connected by channels known as nanocardboard. The other model, made from mylar that is coated on one side with carbon nanotubes, allows for a difference in properties between the top and bottom. The vehicles are roughly eight centimeters in diameter.
In terms of operation, Bargatin compares the aerial vehicle to a helicopter. The blades of a helicopter spin and push the air down to create a downward jet, an equal and opposite reaction occurs, and the helicopter propels upward. In the case of the aerial vehicle, the downward jet is the result of temperature differences: when the temperature is higher on the bottom of the plate than on the top, a process called thermal creep occurs and creates the jet.
“Just like in a helicopter, you get the lift force, and potentially you can also make it maneuver, although we’re not there yet,” says Bargatin.
While they haven’t begun experimentation with the structures as payload-bearing vehicles, Bargatin and team imagine that this technology will be used for carrying small cameras or small sensors in the future, using only the power of light.
Physics-Informed Neural Networks
The second technology for this year’s competition is “Physics-Informed Neural Networks,” a computational technique that aims to blend physics with artificial intelligence to open the path to learning functional relationships between the inputs (excitations) and the outputs (responses) of a system, even in cases where some variables can’t be directly measured. This technology was developed by Paris Perdikaris, Assistant Professor of Mechanical Engineering and Applied Mechanics, and colleagues in his Predictive Intelligence Lab.
A neural network is an algorithm that represents a functional relationship. It receives inputs, it has tunable parameters and, according to those parameters, it can make predictions.
“The goal,” says Perdikaris, “is to take observed data and underlying physical principles and come up with ways of tuning those parameters accordingly so the neural network predicts what we’re interested in, and does so in a reliable and accurate fashion.”
The user would provide two inputs: one, any observed data collected of the system being studied and two, a set of constraints, or equations, that determine how the interactions in that system are governed. Ideally, the system provides an AI educated guess about what it does not know, based on what it does.
Perdikaris and team hope the neural networks will someday be used to advance medical diagnostics. Currently, predicting biomarkers such as blood pressure in human arteries is difficult, as collecting the data in a non-invasive way is inaccessible.
“We’re trying to gather measurements that are easy to gather in a clinic, such as measurements of blood velocity, and then compliment that with the underlying knowledge of fluid mechanics — these are the physical principles that come into play – so that our system is able to infer the blood pressure from the measurements of velocity,” says Perdikaris.