Nine Numbers Beat Fifty Thousand

A scientist runs the simulation. The numbers come back, the curve bends exactly where the experiment said it should, and still, staring at the screen, she cannot tell you why. This is the bargain of modern computational science. You can have the answer, or you can understand it, and more and more often you are asked to choose.

The math underneath describes how things spread and change: heat creeping through metal, a disease moving through a population, a system tipping toward collapse. For a lucky few, a person once derived a clean formula, the kind a textbook prints in a box. For the rest, which is nearly all of them, there are two options. You can grind out a table of numbers that tells you what happens but never why. Or you can train a neural network, which predicts beautifully and explains nothing — a student who aces every exam by memorizing the answer key and can't show a line of work.

A new system called ASYS goes after the option everyone had given up on. Instead of a table or a black box, it hunts for the readable formula itself, even for problems where none had ever been found. It works in two nested loops. In the outer, a large language model proposes shapes for the formula — something like "e to the minus r-squared, times a wave, with nine adjustable knobs" — drawing on the physics literature. In the inner, the system tunes those knobs to fit the physics. It never peeks at the real solution. It only checks whether a candidate obeys the equation's own rules, at the boundary as well as the interior.

The proving ground was a notoriously hard equation describing how living cells release a chemical that draws other cells toward them, until they pile into a single catastrophic clump. Specialists had built a neural network just for this problem, tuned to its geometry, with 49,921 internal weights. They even had to hand it the right coordinate system first, a trick for reframing its geometry. ASYS got none of that help. It found the shape on its own and wrote it down in nine numbers. By the standard measure of error, where lower is better, the nine-number formula scored 0.188 against the giant's 0.258. Nine moving parts beat fifty thousand.

That gap is the whole point. The formula is not simple — nine knobs is real machinery, not a slogan — but a physicist can read it, see which variables dominate, which vanish, and ask the next question. The network can't offer that, and on most problems it is the only tool there is.

Because it doesn't always win. On the cruelest stress test, ASYS scaled itself up — not nine knobs now but more than thirteen thousand — and still lost badly, 0.465 to the network's 0.196. The extra machinery bought nothing: the AI never imagined the right family of shapes, never found the coordinate trick the specialists had been handed for free. The bottleneck was never size. It was imagination. If the model never proposes the right kind of formula, no amount of fitting saves it. So the question stands. If what we want from science is to understand, and a formula you can read sometimes beats a machine you can't, what is the black box for?