New AI Trick Makes Physics Simulations Up to 14x More Accurate

Engineers and scientists rely on computer simulations to answer hard questions: Will this wing stall? How does blood flow through an artery? What happens when a wave slams into a sea wall? The maths behind those answers — partial differential equations — is famously slow to solve. A newer approach called physics-informed neural networks, or PINNs, uses AI that learns the rules of physics instead of just memorising data. The promise is fast, flexible simulation without needing a giant grid of points covering every inch of the problem.

The trouble is that PINNs often fail exactly where accuracy matters most. They check their work at evenly spread points, which is fine for smooth problems but misses sudden, violent changes — shock waves, sharp spikes, unstable bursts. Two researchers, Yousen Wang and Wei Zhao, fixed this by letting the AI watch its own gradients (the steepness of change in the solution) and drop more checkpoints wherever things are getting rough, while keeping some spread evenly so it doesn't lose sight of the big picture. They call the method 3GAS-PINNs.

They tested it on three classic nonlinear equations used to describe shock waves, water waves and light pulses. On those problems, the new method captured the sharp features the standard version blurred over — and improved accuracy by up to a factor of 14.

So what? Simulations are how planes, cars, bridges, weather forecasts and medical devices get designed and tested before anyone builds or risks anything real. Making them dramatically more accurate with less computing time means faster design cycles, cheaper experiments, and better predictions of dangerous extremes. The catch: this is a mathematical proof of concept, tested on three textbook problems, not on real aircraft or real oceans. Turning it into everyday engineering software will take more work — but it points in a genuinely useful direction.