TL;DR: An AI-guided method called BEAM found six working 3D-printing recipes for NASA's GRCop-42 copper alloy in 40 tests over three months, including the first successful prints at 500 watts.
AI narrows 100 million options to 40 tests
Researchers at Washington State University reported an artificial intelligence system that identified workable 3D-printing settings for GRCop-42, a NASA-developed copper alloy used in rocket engine combustion chambers, after earlier laboratory work had not produced a usable low-power result, according to 3D Printing. The system started from data on 37 failed configurations and then guided 40 additional experiments over three months to find six successful configurations at different laser powers, according to the same report. One of those successes was the first recorded printing of the alloy at 500 watts, a power level available on many commercial machines, according to Advanced Manufacturing.
The peer-reviewed description of the work was published on Sept. 1, 2026, in AI Magazine, volume 47, issue 3, as article e70069, according to Experts@Minnesota. The article lists authors Azza Fadhel, Nathaniel W. Zuckschwerdt, Aryan Deshwal, Susmita Bose, Amit Bandyopadhyay and Jana Doppa, according to the same university record. The publication carries DOI https://doi.org/10.1002/aaai.70069 and was issued by John Wiley & Sons Ltd on behalf of the Association for the Advancement of Artificial Intelligence, according to that record.
The method is named Bayesian Experimental design for AM, abbreviated BEAM, and it combines AI-driven adaptive experimental design with domain knowledge to accelerate discovery of feasible process configurations, according to Experts@Minnesota. BEAM treats process development as a closed-loop learning problem in which a probabilistic surrogate model learns from prior experiments and iteratively recommends promising configurations for laboratory validation, according to the same abstract. The project used BEAM on a directed energy deposition process, known as DED, to print GRCop-42, according to that abstract.
The result matters for access to the alloy because about 90% of commercial printers cannot currently print GRCop-42 at the higher powers it has historically required, according to 3D Printing. Research leader Jana Doppa, Huie-Rogers Endowed Chair Professor of Computer Science, said the newly identified parameters allow use of those commercial printers, according to the same report. Doppa described that outcome as essentially democratizing printing of the alloy, according to Advanced Manufacturing.
Why GRCop-42 resists ordinary printing
GRCop-42 is a copper-chromium-niobium alloy developed by NASA for use in rocket engine combustion chambers, according to 3D Printing. The material retains strength at high heat and conducts heat well, characteristics valued for combustion-chamber service, according to the same report. Those same thermal properties make layer-by-layer fabrication difficult because the printing step has historically required substantial laser power on specialized machines, according to 3D Printing.
The detailed journal abstract describes GRCop-42 as a NASA-developed copper alloy that is difficult to process using conventional infrared laser systems because of low laser absorptivity and high thermal conductivity, according to Experts@Minnesota. Low absorptivity means much of the infrared energy is not taken up by the metal, while high conductivity moves heat away from the intended melt zone, according to that description of the processing challenge. Earlier attempts at lower wattages available on common printers had not succeeded, according to 3D Printing.
In some of those early trials, a selected configuration produced only a melted mass rather than a usable build, according to first author Azza Fadhel, a doctoral student in computer science, as reported by 3D Printing. Fadhel said the material in those cases was not really printable in the tested condition, according to the same report. That failure history shaped the decision to treat the problem as a search for rare feasible settings rather than incremental tuning of one setting, according to Advanced Manufacturing.
The alloy application remains extreme-temperature environments such as liquid rocket engine combustion chambers, where the part must carry load while moving heat efficiently, according to Advanced Manufacturing. The journal record says the new work enabled high-quality GRCop-42 fabrication on widely available lower-power infrared laser platforms for the first time, according to Experts@Minnesota. That statement refers to the directed energy deposition demonstration reported in the paper, not to every commercial printer model, according to the same abstract.
A search space too large to test by hand
Printed output quality depends on complex interactions among process parameters such as laser power, scan speed and feed rate, according to Experts@Minnesota. Conventional trial-and-error approaches are inefficient for that task because experiments are costly and the process parameter space is extremely large, according to the same abstract. The broader paper frames that situation as part of a growing class of scientific discovery problems in which experiments are expensive and experimental resources must be allocated intelligently, according to that record.
For GRCop-42 on the studied process, the team faced more than 100 million possible configurations, according to 3D Printing. A single print can cost hundreds of dollars, and checking quality can take days, according to the same report. Manual testing of all combinations was therefore not realistic even with substantial time and funding, according to Fadhel as reported by 3D Printing.
The starting point was 37 failed configurations previously tested in the School of Mechanical and Materials Engineering at Washington State University, according to 3D Printing. Those failures provided the initial training signal for the model rather than being discarded, according to Advanced Manufacturing. Researchers trained an AI model using data from those 37 previously unsuccessful experiments before recommending new settings, according to that outlet.
Doppa described the learning challenge in practical terms because each test returns essentially a binary success-or-failure signal, according to Advanced Manufacturing. The objective was to minimize the number of attempts needed to reach successful settings quickly, according to the same report. Doppa called the task a very challenging case for AI for that reason, according to that outlet.
How BEAM selects the next experiment
BEAM creates a human-AI collaborative workflow in which domain experts define physical constraints, AI prioritizes experiments, and laboratory feedback improves future recommendations, according to Experts@Minnesota. The probabilistic surrogate model updates its expectations after each laboratory result and then proposes the next configurations for validation, according to the same abstract. That closed loop continued through the three-month deployment described in the paper, according to that record.
In practice, the AI model estimated how likely each untested setting was to succeed and then recommended small batches for printing, according to 3D Printing. The batches deliberately mixed likely winners with uncertain cases intended to improve the model, according to the same report. Advanced Manufacturing described the same balance as weighing combinations that appeared promising against others that could provide valuable new information.
New results were fed back into the model after each batch, refining later recommendations, according to reporting on the study design. Fadhel said she welcomed every returned result, including failures, because each outcome improved the model, as reported in coverage of the project. That feedback step is central to the closed-loop learning description in the journal abstract, according to Experts@Minnesota.
The physical printing work was carried out by Nathaniel Zuckschwerdt, Susmita Bose and Amit Bandyopadhyay in mechanical and materials engineering, with Aryan Deshwal of the University of Minnesota also collaborating, according to 3D Printing. Deshwal is listed with Computer Science and Engineering in the university publication record, according to Experts@Minnesota. Doppa, a professor of computer science and engineering at Washington State University, led the work, according to Advanced Manufacturing and 3D Printing.
The paper also appeared in the Proceedings of the AAAI Conference on Artificial Intelligence and received that group’s Innovative Deployed Application Award, according to 3D Printing. The journal version in AI Magazine documents the same BEAM deployment on directed energy deposition for GRCop-42, according to Experts@Minnesota.
Six working recipes in three months
Within three months, BEAM discovered multiple defect-free process configurations across laser power levels from 950 to 500 watts, according to Experts@Minnesota. That outcome followed several months of unsuccessful manual experimentation, according to the same abstract. The abstract characterizes the change as a dramatic reduction in time and resource expenditure compared with that earlier manual effort, according to that record.
News reports quantify the same deployment as 40 total experiments yielding six successful settings at different laser powers, according to 3D Printing. The 40-test count includes the AI-guided trials conducted after the initial 37 failures, according to Advanced Manufacturing. The six successes spanned laser powers down to the 500-watt level, according to 3D Printing.
The 500-watt result is reported as the first successful printing of GRCop-42 using a 500-watt laser, a power level available on many commercial machines, according to Advanced Manufacturing. The journal abstract places the low end of the discovered defect-free range at 500 watts and the high end at 950 watts, according to Experts@Minnesota. By covering that interval, the team showed options beyond a single narrow recipe, according to that abstract.
Lower laser power could reduce energy consumption, decrease wear on printing equipment and lower the costs of post-print sample processing, according to Advanced Manufacturing. Those potential benefits were presented as expectations associated with lower-power operation, not as measured savings in the reported deployment, according to the same outlet. The team also said the approach could make the alloy more accessible to universities, smaller laboratories and manufacturers that lack specialized high-power equipment, according to that report.
What lower power changes for users
The access argument rests on installed hardware. About 90% of commercial printers cannot supply the power historically needed for the alloy, according to 3D Printing. Finding feasible parameters at 500 to 950 watts therefore brings a larger share of existing infrared laser platforms into consideration for GRCop-42 work, according to Experts@Minnesota and 3D Printing. Doppa said the parameters allow use of those commercial printers, according to 3D Printing.
For university and small-laboratory users, the practical difference is the ability to attempt the alloy without acquiring a specialized high-power system, according to Advanced Manufacturing. Researchers said the method could lower energy consumption and production costs while expanding access for those groups, according to the same outlet. The reports do not list purchase prices, operating costs or specific printer models that were validated outside the study equipment, so those purchasing details remain unreported in the cited sources.
Quality assessment remains a time-consuming step because thoroughly analyzing a finished sample can require several days, according to reporting on the project. That inspection burden is part of why reducing the number of physical trials from millions of possibilities to 40 mattered, according to Advanced Manufacturing. Each avoided print also avoids hundreds of dollars in direct print cost, according to 3D Printing.
The team also argues the same AI-guided approach could be adapted to other metal alloys and additive manufacturing systems, according to Advanced Manufacturing. That statement is a claim about transferability, not a report of additional alloys already completed in the cited sources. The journal abstract similarly presents the GRCop-42 deployment as a demonstration of how AI-guided experimental design can accelerate scientific and engineering discovery in resource-constrained settings, according to Experts@Minnesota.
Team, roles and publication record
The collaboration joined computer science researchers with specialists in mechanical and materials engineering at Washington State University, plus a University of Minnesota collaborator, according to 3D Printing. Fadhel served as first author and handled the modeling side as a computer science doctoral student, according to the same report. Zuckschwerdt, Bose and Bandyopadhyay handled the printing work in the School of Mechanical and Materials Engineering, according to that outlet.
Deshwal contributed from the University of Minnesota and appears in the publication record under Computer Science and Engineering, according to Experts@Minnesota and 3D Printing. Doppa led the research and holds the Huie-Rogers Endowed Chair in Computer Science, according to 3D Printing. Advanced Manufacturing identifies Doppa as a professor of computer science and engineering at Washington State University.
The journal record confirms publication in AI Magazine on Sept. 1, 2026, with the citation Fadhel, Zuckschwerdt, Deshwal, Bose, Bandyopadhyay and Doppa, according to Experts@Minnesota. The article number, volume and issue provide a stable reference for readers seeking the full methods description, according to the same record. The DOI link offers direct access to the publisher page, according to that university entry.
Reporting on the conference version says the paper appeared in the Proceedings of the AAAI Conference on Artificial Intelligence, where it received the Innovative Deployed Application Award, according to 3D Printing. That award recognizes deployed use, which aligns with the paper’s emphasis on laboratory validation rather than simulation alone, according to the same report and the journal abstract. The combination of peer-reviewed publication and a deployment award gives readers two independent markers of review, according to those sources.
Limits of what the sources establish
The cited sources establish a specific directed energy deposition demonstration with GRCop-42, not a universal recipe for all printers or alloys, according to Experts@Minnesota. The defect-free configurations were found across 950 to 500 watts in the reported setup, according to that abstract. Performance on different machines, powders or part geometries was not detailed in the cited reports.
The cost discussion is framed in general terms: prints cost hundreds of dollars and analysis can take days, according to 3D Printing. Lower-power operation is expected to reduce energy use, equipment wear and post-processing costs, according to Advanced Manufacturing. Exact dollar savings, energy measurements and maintenance intervals were not provided in the cited sources.
Future use for other alloys, manufacturing systems, materials research and drug development is presented as a direction the researchers believe is possible, according to Advanced Manufacturing. The sources do not report completed drug-discovery experiments or additional alloy qualifications by the Sept. 1, 2026 publication date. Readers evaluating the method for another material would need the full paper methods and independent testing on their own equipment, according to the publication record and the reported 40-experiment scope.