AI Models Trained on Physics Are Reshaping Engineering

Large language models have already made waves in software development. Now, a new breed of AI—large physics models—is beginning to make its mark on design engineering. These systems are starting to supplement or even replace traditional physics simulations in fields such as automotive, aerospace, and semiconductor manufacturing.

At Nvidia’s GTC conference in March, industry leaders discussed how these AI models are speeding up workflows. Jaguar Land Rover, for instance, is using technology from Neural Concept to streamline vehicle design. General Motors has integrated physics-based AI into its design process, cutting the time to calculate a car’s drag coefficient from two weeks to just minutes. ‘Experts from aerodynamics and the creative studio can now sit together and iterate instantly,’ says Rene Strauss, GM’s director of virtual integration engineering.

The speed gains are staggering. Running an AI model trained on simulation data can be 10,000 to nearly a million times faster than running the actual simulation, depending on complexity and resolution, notes Jacomo Corbo, CEO of PhysicsX. While accuracy remains a concern later in the design cycle, early-stage AI predictions are considered sufficiently reliable for iterative exploration.

Training these models varies: some use transformer architectures, others rely on geometric deep learning or neural operators. Most companies currently train their own models on proprietary simulation data, but efforts are underway—like PhysicsX’s collaboration with Nvidia—to develop open standards and more generalizable ‘foundational’ models. As these models grow, they exhibit scaling laws similar to LLMs, improving performance and adaptability.

Despite the advances, the role of simulation and human engineers remains vital. Some experts, like Neural Concept’s Thomas von Tschammer, believe simulation will never be fully replaced; instead, AI will make its use smarter and faster. Others, like Corbo, envision inference eventually taking over traditional numerical simulation. Whichever path prevails, all agree that engineers will remain essential, empowered by these tools to focus on higher-value design decisions. As von Tschammer puts it, ‘We still need engineers more than ever.’