New AI Finds Hidden Science Formulas 10x Faster

AI is getting better at doing the work of a scientist: looking at raw data and guessing the math formula underneath. This is called symbolic regression. Instead of just making predictions, it produces a real equation you can write down and use. For example, given measurements of a swinging pendulum, it might discover the formula for how long each swing takes.

There's a bottleneck in this process, though. When the AI evolves candidate formulas, it needs to fine-tune the numbers in the formulas — like adjusting coefficients so the equation fits the data well. That fine-tuning step has always been slow. So slow that most modern AI systems simply skipped it, which meant they missed many correct equations.

Now, researchers have built a new solver that does that fine-tuning efficiently on a GPU, the same kind of chips that power video games and AI chatbots. By grouping many formulas together and processing them in parallel, the system can check over half a million formulas per second. Compared to a leading older tool running on 64 top-end CPU cores, the new approach is about ten times faster — while producing equally accurate results.

In tests, this solver was plugged into an existing symbolic-regression engine called EvoGP. With the upgrade, the engine found the correct governing equation on 10 out of 18 made-up physics problems. Without it, it found zero. This isn't just about being faster — it's about uncovering answers that were previously out of reach. That could mean sooner breakthroughs in materials science, climate modeling, biology, and anything where scientists are searching for hidden patterns in massive data.