printed in metal,

unproven in flight

Problem statement

Metal additive manufacturing (AM) can produce rocket engine components and structural parts with geometries impossible to achieve through traditional machining or casting — internal cooling channels, topology-optimized structures, consolidated assemblies. However, qualifying AM parts for human-rated or high-reliability spaceflight remains a largely unsolved problem. Unlike wrought or cast metals with decades of statistical property databases, AM parts exhibit location-dependent microstructure (varying with build orientation, distance from the build plate, and local thermal history), stochastic defects (porosity, lack-of-fusion voids), and residual stresses that are difficult to predict or inspect non-destructively.

Why this matters

Metal AM promises to reduce rocket manufacturing costs by 50–80% and lead times from months to weeks. Relativity Space built an entire rocket (Terran 1) primarily through AM; SpaceX prints SuperDraco chambers; Rocket Lab, Aerojet Rocketdyne, and others use AM for injectors and turbopumps. But qualification currently requires extensive test campaigns for each new part — sometimes destroying dozens of expensive articles to build statistical confidence. This negates AM's cost advantage and prevents rapid design iteration. The qualification bottleneck is the primary barrier to AM becoming the default manufacturing approach for space hardware.

What’s been tried and why it hasn’t worked

Traditional "test-like-you-fly" qualification (building and destructively testing representative articles) is prohibitively expensive when every part has unique thermal history. NASA and ESA have attempted to develop AM-specific standards (NASA-STD-6030, ESA ECSS-Q-ST-70-80C), but these are largely prescriptive rather than performance-based — they specify process parameters rather than guaranteeing part properties. In-situ monitoring (melt pool cameras, acoustic emission during printing) can detect gross defects but cannot yet predict mechanical properties from process signatures. Machine learning models correlating process parameters to part performance show promise but require training data from destructive tests, creating a chicken-and-egg problem.

What would unlock progress

A physics-based or hybrid model that can predict local material properties (fatigue life, fracture toughness) from build parameters and in-situ monitoring data — without requiring destructive testing of each new geometry — would transform qualification. Alternatively, a "digital twin" approach where the build simulation is validated once and then trusted for new geometries could replace per-part testing with per-process qualification. Non-destructive evaluation methods sensitive to the specific defect types in AM (sub-100-µm porosity, lack-of-fusion defects) would also help, as current CT scanning is too slow for production rates and too resolution-limited for critical defect sizes.

Entry points for student teams

A team could correlate in-situ monitoring signals (melt pool thermal imaging, layer-by-layer photography) with post-build CT scan defect data on standard test geometries to build predictive models. Open datasets from NIST AM benchmarks provide a starting point. A design-focused team could develop a qualification framework that maps required confidence levels to inspection and testing requirements, identifying where current NDE methods are sufficient and where gaps remain.