AI found a cheaper way to 3D-print NASA's rocket engine alloy

Researchers at Washington State University used an AI system to search more than 100 million possible configurations for 3D-printing GRCop-42, a copper alloy NASA developed for rocket engine parts, and found six settings that work on ordinary commercial machines — including, for the first time, at 500 watts of laser power.

GRCop-42, an alloy of copper, chromium and niobium, was designed by NASA for the parts of a rocket engine that have to do two contradictory things at once: survive extreme heat and conduct it away quickly. Combustion chamber liners and fuel injector face plates are the typical uses, and the alloy family has already flown — Relativity Space's 3D-printed engines used NASA's GRCop alloys. The catch has always been manufacturing. Printing it well normally demands very high laser power, which is why, as WSU computer science professor Jana Doppa put it, "ninety percent of commercial printers cannot print this metal alloy."

What makes this an AI result rather than a metallurgy one is the shape of the search problem. The combinations of laser power, scan speed, layer thickness and the other settings run past 100 million, and each candidate has to be physically printed before anyone knows whether it worked — a slow and expensive test. The team started from 37 configurations that had already failed, built a model that estimated the odds of success for untested ones, then had it propose small batches that deliberately balanced two goals: trying settings that looked promising, and probing uncertain regions to make the model itself sharper.

The budget was 40 experiments over roughly three months, and six workable configurations emerged from it at several laser power levels. The payoff is not a better alloy but a cheaper route to the same one: lower power means less energy, less wear on equipment and lower post-processing costs — and, the part Doppa emphasizes, access for universities, small labs and companies that never had the specialized high-power machines.

The work was published in the Proceedings of the AAAI Conference on Artificial Intelligence and received that conference's Innovative Deployed Application Award. Doppa led the team with Azza Fadhel, Nathaniel Zuckschwerdt, Susmita Bose and Amit Bandyopadhyay at WSU, alongside Aryan Deshwal at the University of Minnesota.