New AI Method Avoids Dead-End Designs, Cutting Costly Tests

Imagine you're designing a pressure vessel — a tank built to hold gas under extreme force. You can't just do the math and know the best shape and thickness. You have to try options and measure the results, and each measurement is slow and expensive. Engineers call this "black-box optimization": finding the best answer when you can only test it, not calculate it.

Today's tools handle this by building a cheap stand-in guess — a "surrogate model" — that predicts how good each option is, so you run fewer real tests. But real designs come with rules: that valve must be paired with this pipe, this material can't be used at that temperature. Standard stand-in guesses ignore those rules and waste expensive tests on designs that could never be built.

Keisuke Onoue and Ryosuke Kojima's method, presented at the ECML PKDD 2026 conference, bakes the rules directly into the guess. Their technique, based on "tensor decomposition" — a way of spotting patterns in big tables of options — learns which combinations are allowed at the same time as it learns which ones are good. In experiments including a pressure vessel design task, it steered away from impossible designs and reached strong answers with fewer attempts.

The catch: this is academic research, not software you can buy today. It works best when the rules are known ahead of time and written down clearly. If the rules are hidden, fuzzy, or keep changing, the advantage shrinks. Still, the underlying idea matters. Every wasted trial costs real money, and fewer wasted trials means cheaper, faster design of tanks, engines, chips and even drug molecules — savings that eventually get passed along.