Heard on Campus: PhysicsX founder on deep learning in physics and engineering
“Deep learning is driving a revolution in physics, science, and engineering,” said Robin Tuluie, founder and chairman of PhysicsX, during the Alumni and Friends Reunion of the Penn State Institute for Gravitation and the Cosmos (IGC) on May 30. Tuluie, who was a postdoctoral scholar with the IGC from 1993 to 1995, presented a talk about the use of deep learning to push the boundaries of engineering.
Deep learning is a type of artificial intelligence machine learning inspired by neural networks in the human brain, using several layers of artificial neurons to process data in different ways. Much like large language models (LLMs) — a type of deep learning model— like OpenAI’s ChatGPT and Microsoft Copilot “train” on existing texts and language to create logical sentences, large physics models and large geometry models train on data from physics, geometry, and spatial relationships in order to generate predictions for variables like air or fluid flow across arbitrary geometries. PhysicsX has created large physics models and large geometry models to address engineering challenges in optimizing structures, from aircraft wings and hydro-turbine blades to the tires of Formula 1 cars and the heat-exchangers used to keep batteries in electric vehicles cool.
Traditional analysis of structural optimization, Tuluie explained, might involve running many simulations to explore, for example, how adjusting aspects of an airplane wing’s geometry might reduce drag. But even on the fastest supercomputers, the hundreds of simulations required might take hours, days, or even months to run. But after a day or two of training, the deep geometry models can accomplish the same results in less than a second.
“These deep learning models can do this in a second or less, and you can exploit that speed,” he said. “You have real time models with the fidelity of very high performing numerical simulations.”
Deep geometry models, he said, use geometry to inform physics. In the example of airplane optimization, instead of describing a wing with parameters like the length and angle of the wing that are input into a model, the model uses a geometrical “mesh,” as if a piece of mesh was wrapped around the entire wing and each of the intersections of fabric were included as an input. This might yield 10 or 100 million mesh points, or several hundred million for more complex shapes like a car. The researchers at PhysicsX trained their airplane optimization model on the mesh from a wide variety of aircrafts—from drones to airplanes and even birds and insects.
“When you start to learn these geometries, you start to see these clusters. We go from things we know—birds, drones, or airplanes—but we cover stuff in between,” Tuluie said. “When we're in the white space, in the space between a bird and an airplane, we can explore a space without prescription. [We can] discover entirely new geometries, geometries beyond human dimensions. And now we're exploring all sorts of geometries that make the airplane safe, not just the aerodynamic performance.”
Before founding PhysicsX, Tuluie was previously the head of research and development at Renault (Alpine) F1, where his innovations helped the Formula 1 team to win back-to-back double world championships, and served as the vehicle technology director at Bentley Motors. His expertise in numerical modeling traces back to his postdoctoral research at Penn State, where he used numerical simulations to explore the cosmic microwave background—radiation that fills all space in the observable universe.
The IGC reunion event welcomed back alumni and friends of the institute as well as current members for a series of presentations, networking opportunities, panel discussions, and a showcase of new developments and initiatives at the IGC. Other speakers included current members Tetyana Pitik, N3AS Postdoctoral Fellow at the University of California, Berkeley and visiting scholar at Penn State; and Bingjie Wang, assistant research professor of astronomy and astrophysics at Penn State; as well as former postdoctoral scholars Tristan McLoughlin, professor of pure and applied mathematics at Trinity College Dublin; Gordana Tešic, data scientist at Meta, Ottawa; Jonathan Trump, associate professor of physics at the University of Connecticut; and Surabhi Sachdev, assistant professor of physics at Georgia Tech.
The IGC is a research community that seeks to collectively push the boundaries of understanding of the fundamental forces of nature and how they shape the evolution of the universe. Meeting this challenge requires a joint effort between many converging areas of expertise and an enthusiasm for new ideas and new perspectives. The IGC strives to provide a fertile environment for collaboration between the domains of particle physics, gravitational physics, mathematics, cosmology, astrophysics, statistics, and computation.