Q&A: How Can Deep Learning Improve Physics and Engineering?
Deep learning is a form of artificial intelligence (AI) inspired by the neural networks of the human brain. The most familiar examples are large language models (LLMs), which train on text in order to predict and generate coherent sentences. Robin Tuluie, who was a postdoctoral scholar in the Penn State Institute for Gravitation and the Cosmos from 1993 to 1995, however, works with large physics models (LPMs) and large geometry models (LGMs)—systems trained on high-quality physics, geometry, and spatial data in order to optimize designs in engineering and manufacturing.
As the founder and chairman of PhysicsX, Tuluie and his team use these models to improve everything from aircraft wings and wind-turbine blades to Formula 1 tires and the heat exchangers that keep electric vehicle batteries cool.
We sat down with Tuluie to discuss his career path, how physics and deep learning can work together to accelerate innovation, and the opportunities and challenges that come with artificial intelligence.
Q: How did you end up moving from astrophysics to improving Formula 1 race cars?
Tuluie: As a graduate student, I worked with Richard Matzner at the University of Texas at Austin—one of the leaders of the project simulating the collision of two black holes, referred to as a “Grand Challenge” because it was probably the most complex numerical simulation in science at the time. My focus was using numerical physics simulations to study tiny variations, called anisotropies, in the cosmic microwave background (CMB)—the faint radiation left over from the big bang. Later, as a postdoc at Penn State with Pablo Laguna, then an assistant professor of astronomy and astrophysics, I continued that work and discovered a new physical effect in the CMB—a faint small-scale dipole signature that could, for the first time, allow measurement of the transverse velocity of superclusters of galaxies. It was fascinating work, but it might have taken 20 years before satellites could confirm it. I was too impatient for that! So I went into racing and engineering—fields where you learn fast and see results right away.
Q: What sparked your interest in racing, and how did your physics background help you advance in that field?
Tuluie: Growing up in Germany, I loved tinkering with my moped to make it faster. On a family trip to the US, I saw all these amazing motorcycles you could legally modify—which you couldn’t do back home—and I thought, “I need to move here.” I started motorcycle racing at 18 while studying physics at the University of California, Berkeley—secretly, since my dad would never have approved. I’d ride my bike to school during the week, then on weekends strip off the street gear, bolt on the race kit, head to Sears Point to compete, and switch it all back before Monday morning.
When I got to Penn State, I built my first race bike from scratch—won two national championships on it, as well as Daytona—and that set the stage for my engineering path. After my postdoctoral work, I joined a motorcycle company, then an engineering firm, and finally made it to Formula 1, where I’d always wanted to be.
As a physicist, you learn to break down problems into fundamentals: the forces, the governing equations, and how they interact. When I joined Renault F1, using numerical simulation for design and development was still in its infancy; the wind tunnel and track were dominant. One day, the car’s half-scale model was bouncing in the wind tunnel—I recognized it as a structural resonance problem, the same physics that affects skyscrapers in high winds. I suggested adding a tuned mass damper, just as architects do for tall buildings. We simulated, tested, and adapted it for the actual car. The result was a major performance gain, and we won the world championship on the back of that insight.
Later, at Mercedes F1, we built the first fully simulated car from the ground up. Numerical simulation drove every aspect of design, and the team went on to win eight world championships in a row. I then joined Bentley as vehicle technology director, where we built an end-to-end numerical simulation capability for one of the world’s leading luxury car manufacturers. Eventually, I founded my own engineering consultancy there, but when we hit the limits of scale, I left to start PhysicsX with Nicolas Haag, a brilliant young engineer who became my cofounder.
The common thread through all of it—from racing as a hobby to Formula 1 to global automotive and now PhysicsX—is physics. Using numerical physics and now deep learning, we can solve problems faster than ever before and unlock entirely new frontiers in engineering.
Q: What was your goal for PhysicsX and how has it grown?
Tuluie: From the start, our belief was simple: With the best numerical simulation and machine learning, you can develop things at unprecedented speed. A typical car takes five to six years to go from concept to production, while a Formula 1 car takes just one. You have to identify early which parts take the longest to build and ensure they’re informed by simulation from day one.
When geometric deep learning, pioneered by Michael Bronstein at Oxford, began showing real-world potential, we knew we were perfectly positioned to take it forward. Since then, PhysicsX has grown to over 200 people.
Our mission is to help customers build the best products in the shortest possible time, applying physics and AI to make a positive impact in the world. As physicists, our purpose is to serve humanity: to understand the world around us, make life better, and explore the unknown. Using hard science to tackle tough engineering challenges and make a real dent in the universe—that’s what drives us.
Q: Can you elaborate on the deep learning models that you use?
Tuluie: Deep learning uses artificial neural networks with many layers of computation to recognize patterns and make predictions. Most people are familiar with large language models, which are trained on text to generate answers. We build large physics models, trained on numerical physics data to answer physics questions, with data, rather than text. Once trained, we can show the model a geometry—say, a shape—and it instantly predicts its physical properties: how aerodynamic it is, how strong, or how it radiates electromagnetically. These models deliver answers in under a second, compared to hours or days with traditional simulation, and, because they’re trained on high-precision, engineering-grade data, they’re remarkably accurate.
We can then tweak the geometry, feed it back into the model, and repeat the loop—from geometry to physics prediction to optimized geometry. Automating this process lets us perform around 100,000 optimizations a day. In a couple of weeks, that’s millions of designs tested, ensuring we’ve found the best solutions for any given physics problem.
Q: What is the impact of running these models? What opportunities and challenges do we face with artificial intelligence?
Tuluie: LLMs like ChatGPT have drawn criticism for the immense resources required to train them: huge compute power, water for cooling, and millions of dollars in cost. Our models are very different. They’re specialized, trained on smaller but far higher-quality datasets, so training costs and resource use are just a fraction of what LLMs require. And the benefits are tangible: A small aerodynamic improvement on an aircraft can translate into major fuel savings—cutting carbon emissions far beyond the footprint of training the model itself.
At PhysicsX, one of our founding principles is that our work must make a measurable, positive difference in the world. Reducing CO₂ emissions through better design is core to that mission.
AI is an extraordinary gift—and a profound responsibility. Humanity has faced such moments before: nuclear energy, space exploration, the internet. Each has brought both progress and some peril. How we handle this one, whether we can focus our collective creativity on using AI for good, will define us. I hope we can continue developing it with care, integrity, and purpose.
Q: Do you have any advice for current students?
Tuluie: Whether you’re an undergraduate or graduate student, stay curious—and go deep. Learn how the scientific process works and apply it everywhere, not just in science. Seek clarity and insight in everything you do. Develop your mind as the analytical engine it is but also cultivate empathy and integrity. We’ll need both to navigate the challenges ahead. If you can do that, the world is your oyster. You’ll be ready to tackle the hardest problems and make a real, lasting dent in the universe.