Chemists Just Got a Faster AI That Admits When It's Unsure

A pair of researchers has published a new way to make one of science's most trusted prediction tools — the Gaussian process — fast enough to use on large chemistry problems. Gaussian processes are prized because they don't just give you an answer; they also tell you how confident they are. That honesty is exactly what you want when screening molecules for a possible new drug. The catch has always been speed: the math balloons as the problem grows.

Their trick is symmetry. A molecule doesn't change when you rotate it in space, so a good model shouldn't have to relearn that fact from scratch. By baking this property — called "equivariance," meaning rotating the input rotates the output and nothing else changes — into the model, the researchers cut down both the data and the computing power needed. They call their method "triply scalable" because it scales up in three directions at once: bigger inputs, larger symmetry groups, and more data.

They tested it on a real quantum-chemistry task: predicting the electric dipole moment of N-methylformamide, a small molecule. The method delivered accurate predictions with honest error bars, at a fraction of the computing cost of classical Gaussian process inference. The team also built a faster exact version that avoids storing enormous tables of numbers, making it practical on ordinary hardware rather than a supercomputer.

Why does this matter outside a lab? Drug and materials discovery runs on guesses that must be tested, and every test costs time and money. An AI that says "I'm 90% sure" lets chemists skip the shaky candidates and spend lab time on the promising ones. Faster, uncertainty-aware prediction could shorten the path from idea to medicine, battery, or catalyst.