Physics AI could turn days of simulation into seconds

For most of the last century, a new wing, battery or car frame has been checked the same way: engineers build a virtual copy of it and solve the equations of physics, point by point, across that copy. The answers are trusted, but a single simulation can take hours or days on a roomful of computers, and preparing the model can take days more. This is the bottleneck Luminary's CEO, Pete Schlampp, puts at the centre of a contributed piece published on Engineering.com. His claim is that AI models can now predict the same behaviour much faster.

The mechanism is close to what made large language models work. An LLM learns the patterns of language from billions of documents and then writes new text without looking anything up. A Large Physics Model, or LPM, learns the patterns of physics from thousands of physics-based simulations: how air flows over a wing, how heat moves through a battery, how metal bends in a crash. Once trained, it predicts how a new shape will behave without solving the equations from scratch. According to the author, it then gives the pressure, stress and heat at every point of the surface in seconds, where the traditional method needed days.

The idea is not new on its own, and the author places it in a wider run of results. In 2024, Demis Hassabis and John Jumper of Google DeepMind shared the Nobel Prize in Chemistry for AlphaFold, which predicted the structures of virtually all 200 million proteins known to science. A year earlier, DeepMind's GraphCast produced 10-day weather forecasts more accurate than the leading operational system on roughly 90 percent of the measures tested, in under a minute instead of hours on a supercomputer. In February 2025, the European Centre for Medium-Range Weather Forecasts began running a machine-learning model operationally alongside its existing system, reporting gains of up to 20 percent on tropical cyclone predictions and a thousandfold reduction in the energy needed to produce a forecast. These are weather and biology systems, not LPMs for industrial parts, but the author argues they share the same trait: predicting how a physical system behaves before anyone runs an experiment.

The author attributes three developments to the rise of this technology. GPUs make it fast and affordable to generate the thousands of training runs an LPM needs, cutting a single simulation from days to minutes. Attention-based transformer architectures give predictions at the scale and speed engineering requires. And AI agents that explore thousands of designs need physics feedback, since without it they produce random designs that may not work in the real world. The author also says published results show LPM predictions within a few percent of the slower methods. Those studies are not cited in the piece, and the author runs a company that sells this technology, so the figure is his claim until independent tests confirm it.

If the claims hold, the gain goes beyond the design office. The same model that predicts how a wing performs before it is built can predict how consistent a batch of manufactured parts will be, and catch a defect that a traditional check would miss. Once a plane, engine or vehicle is in service, it can process sensor data and flag a part close to failure, on every unit and continuously, instead of only at scheduled inspections. Pushed further, a fast enough model could sit inside a control system and adjust a machine's trajectory in real time. For aircraft makers, carmakers and battery engineers, that means fewer physical prototypes and quicker iterations. The piece gives no date for when this reaches most firms, and no cost figure, so the timing for smaller engineering teams remains open.