A Washington State University team used machine-learning-guided experiments to identify six process settings for depositing NASA's GRCop-42 copper alloy by powder-fed directed energy deposition. The lowest successful laser power was 500 watts. The important result is not a speed or power record: the researchers used an uncertainty-aware model to choose expensive physical trials and searched for several feasible process windows rather than one nominal optimum.

WSU announced the work on August 24, 2026, while a detailed description of the method had been posted to the arXiv preprint server in January. That status matters. The paper is publicly inspectable, but the available version is not a peer-reviewed journal article. Samples were produced on one five-axis DED platform, so the outcome is not a transferable recipe for every machine, powder batch or geometry.

Why GRCop-42 is difficult to process

GRCop-42 is a copper-chromium-niobium alloy developed by NASA for high-heat-flux rocket propulsion hardware. NASA technical documents describe a copper matrix strengthened by Cr2Nb precipitates, offering a useful balance of thermal conductivity and strength for combustion-chamber liners and similar parts. NASA has previously produced near-fully-dense GRCop-42 by laser powder bed fusion and advanced printed chamber hardware through hot-fire testing.

Copper nevertheless reflects much of a near-infrared laser's energy and rapidly conducts absorbed heat away from the melt zone. Too little energy can create lack of fusion; too much or poorly distributed energy can destabilise the melt pool, increase evaporation and distort the deposit. The WSU team targeted a 500–700 W range that may be accessible to more laboratory and smaller industrial DED systems. According to the preprint, published process knowledge for GRCop-42 below 900 W was limited.

The target was a set of feasible settings, not one optimum

Additive process development is often framed as optimisation: minimise porosity or maximise density. Production does not always benefit from one isolated peak. Several settings at different laser powers and feed rates can reveal operating margin when geometry, throughput, heat input or machine constraints change.

The researchers designed BEAM — Bayesian Experimental design for Additive Manufacturing — to discover distinct configurations that meet stated feasibility criteria. Four inputs were varied: laser power, travel speed, powder feed rate and carrier-gas flow. The authors estimate that the discretised search space contained more than 100 million combinations, far too many for an economical full-factorial campaign.

How the AI-guided experimental loop worked

BEAM did not print material virtually and did not replace measurement. After each batch, it learned a probabilistic surrogate from the physical samples, then proposed a new batch expected to expose another feasible region or reduce uncertainty. Researchers deposited real specimens with those settings, evaluated them and returned the observations to the model.

This active-learning or adaptive-design approach differs from a one-off prediction model because the decision system also recommends where to collect the next data. Engineers remain responsible for defining variables, safe machine boundaries, acceptance criteria and measurement. If those definitions are incomplete or the observations are noisy, an efficient algorithm can still search for the wrong target.

The machine, materials and reported result

The paper describes a five-axis powder-fed DED system with a 1,000 W fibre laser, two powder hoppers and an argon environment held below 20 ppm oxygen. GRCop-42 and Inconel 718 powders with a 15–53 micrometre size range were deposited on 316L stainless-steel substrates. Inconel trials supported development and validation of the method; the newsworthy challenge was the more difficult copper alloy.

WSU says the team began with 37 unsuccessful historical attempts and completed 40 new experiments over three months. Six configurations met the stated feasibility conditions at different powers, including the 500 W setting. Those six points are not six certified end-use components. They are successful parameter configurations for the studied specimens and evaluation procedure.

What does a successful or defect-free specimen prove?

The phrase defect-free can become much broader in news coverage than in a laboratory. An absence of detected defects in a metallographic section or at a chosen measurement resolution does not prove that a large complex component is pore-free, crack-free, chemically uniform or mechanically compliant in every direction. The result supports more efficient process screening, not qualification of flight hardware.

Earlier NASA DED research provides a useful comparison. It reported less than 0.1 per cent porosity in deposited GRCop-42 while room-temperature tensile strength remained below that of laser-powder-bed material. Density alone is therefore insufficient. Grain structure, precipitates, heat treatment, anisotropy, surface condition and service thermal cycles all help determine whether the material is fit for an application.

Why lower laser power could matter

WSU describes the work as a step toward democratising GRCop-42 printing. The technical basis is that a 500–700 W window may be reachable by more systems than a kilowatt-class process. Lower nominal power does not make metal DED inexpensive by itself. Controlled powder delivery, inert shielding, optics, extraction, material characterisation, machine safeguards and qualified staff are still required.

The preprint uses illustrative costs of about $300 per kilogram for GRCop-42 powder, $500–$1,000 for a DED run and $500 plus one to four weeks for post-build analysis of a sample. These are the authors' estimates, not universal market prices, but they explain why experiment count matters. Replacing dozens of low-value trials with targeted builds saves development time, machine capacity and material before it saves production energy.

What cannot be transferred from this result

The 500 W point is not a universal GRCop-42 recipe. A different spot size, wavelength, nozzle, feeder, path strategy, layer height, powder chemistry or substrate changes the energy and mass balance. A DED result also cannot be copied directly into laser powder bed fusion: powder presentation, heat flow, melt-pool behaviour and geometric resolution are different.

Nor does this study prove that BEAM will deliver the same advantage for every alloy. Its promise is to choose informative measurements when physical data are scarce and expensive. Performance depends on problem dimension, measurement uncertainty, the chosen feasibility threshold and whether the search domain contains a workable region at all. Independent groups and additional platforms still need to reproduce the outcome.

What industrial validation would require

The next step is to repeat the proposed settings across powder and substrate lots, days and machine states. Larger geometries should then test sensitivity to heat accumulation. Volumetric inspection, chemistry and microstructure analysis, tensile, fatigue and thermal-conductivity data, and a documented heat-treatment and finishing route would be required before an engineering material specification could be established.

The framework's broader value would be stronger if it found stable windows on different machines with fewer trials than a fixed design of experiments or expert-led parameter campaign. A fair comparison must report cost, failed builds, uncertainty and reproducibility, not merely the best specimen from each route.

Verified facts, research claims and our conclusion

It is verified that the WSU team documented BEAM in a public preprint, explored four DED inputs and reports six feasible GRCop-42 settings from 40 experiments, including one at 500 W. NASA sources independently establish the alloy's origin, high-heat-flux purpose and earlier processing by both powder-bed and directed-energy methods.

It is the researchers' claim that the approach could democratise processing of this alloy and extend to other materials. No public independent industrial benchmark, end-use part qualification or peer-reviewed final paper currently establishes those broader conclusions. LAYERTIMES' editorial assessment is that the genuine advance lies in changing the aim of parameter development: search for several usable windows rather than one attractive optimum, while closing every algorithmic decision with a physical experiment. That is a promising direction for costly metal-AM development, but it is not yet a production recipe.