Relativity Space: the printer has to prove its work

First, make a bad print

One of Relativity Space’s more revealing patent applications describes a machine that deliberately interferes with its own printing process. It disturbs the gas protecting the weld, records the electrical signals, and helps the operator find what went wrong inside the metal. The failure is the experiment.⁴

That is an unexpectedly good place to start a story about rockets. A printer can follow a path and produce the right shape while leaving material that behaves differently from what the designer expected. Relativity’s filings examine several ways to close that gap: create identifiable defects, detect them, adjust the process, and test a smaller piece that experienced the same build.⁴³¹⁵

Read together, these five inventions suggest a more interesting question than how much of a rocket can be printed: how much can the factory know about the metal it has just made? That connection is our reading of separate patent families, not a claim that they form one deployed production system.

What is still worth printing?

Terran R’s manufacturing story already combines different methods. Relativity’s March 2025 update describes friction-stir-welded aluminum primary structures alongside printed engine components. Powder bed fusion and wire arc additive manufacturing both appear in the engine work; the company highlights a full-length printed, regeneratively cooled nozzle.¹⁶

Those choices make engineering sense as a division of labor. A large structure and a geometrically intricate engine component ask different things of a manufacturing process. The company’s current vehicle description is a reusable, two-stage rocket; the patents below help us inspect particular manufacturing problems without treating them as a parts list for that vehicle.¹⁷¹⁶

Give the defect a timestamp

In US20260124693A1, the shielding gas protects the region where wire is being melted and deposited. Disrupting that protection can create porosity. The disclosed experiment makes the disturbance intentional and records when it happened. A shared clock ties the disturbance to the voltage and current measurements; cameras can observe the weld at the same time.⁴⁶

The next step is the clever one. Radiographic inspection locates a defect inside the finished material. The disclosure translates that location back into the robot’s coordinates, then joins it to the sensor record from when the robot was printing there. Instead of a folder containing a bad X-ray and another folder containing a million electrical measurements, the experiment can produce a connection between a particular defect and the conditions surrounding its formation.⁴

That makes the data more useful for studying warning signs. An unusual voltage trace alone does not tell you what happened inside the part. A pore alone does not tell you what the welder was doing when it formed. The proposed arrangement links the two.⁴

Three ways to claim the experiment

The published claims draw several boundaries around that idea. Claim 1 specifies the robotic welder, stored voltage/current time series, a shielding-gas disruption mechanism and a common clock. Claim 9 adds video frames and correlates them with electrical measurements. Claim 13 takes a broader method approach: timestamp process information and the interrupt signal, then correlate them.⁴

Those differences matter. The detailed gas-disruption apparatus is one claimed combination, while the application also seeks protection for a more general correlation method. As of this review, these are published application claims; their final scope remains open.⁴

Finding the hole is only half the job

US12682449B2 moves from making a known defect to deciding what a radiograph shows. The disclosed system normalizes image brightness, divides the image into overlapping patches, and classifies pixels as pore or background. It then assembles detections into a calibrated pore map.³

Why overlap the patches? A feature close to the edge of one patch can look different when it sits inside another. The specification discusses reconciling those classifications. Why calibrate? A region of dark pixels is not yet a pore of a known physical size. Reference features give the image a scale.³⁸

The final question concerns acceptance. Individual pores can be grouped into clusters, measured, and evaluated against predetermined criteria. The issued claim ties those operations to identifying a rejected region. The system is doing more than putting a circle around a suspicious spot: it connects image analysis to a decision about a region of the manufactured article.³

The decision at the end of the image

The pore-inspection allowance is useful because it identifies this same distinction. Examiner David F. Dunphy discusses earlier detection, segmentation and clustering techniques, then explains what the cited record did not supply in the claimed combination: the calibrated pore map and the determination that an area should be rejected under predetermined criteria.¹⁴

That is the examiner’s reasoning on the record before him. The issued claim also contains the robotic inspection hardware, localization, brightness normalization and overlapping-image requirements. Reading only the title could leave the impression of a patent on any AI that spots a hole; reading claim 1 reveals a particular inspection workflow.³¹⁴

Let the observation change the next instruction

The earlier control family supplies another piece of the story. In US12547152B2, issued claim 1 feeds sensor data into a classifier, uses its real-time defect classification to generate manufacturing instructions, and controls the deposition robot with those instructions. Independent claim 14 follows a model-based route to predicted process-control parameters.¹

The drawing puts simulation, machine vision, process monitoring and post-build inspection around the deposition apparatus. The specification also describes sharing data between equipment at different locations. Its ambition is a process that can learn from more than the geometry it was originally told to print.¹⁷

The grant’s abstract emphasizes training data and optimization. The issued independent claims deserve a separate read: the consequential link is from an observed or predicted condition to instructions that actually control the machine. The allowance spells out those two combinations.¹¹⁰

Digital Alloys enters the story

The quality-prediction family adds a different perspective. US12591220B2 repeatedly compares measured information with printing parameters, predicts intermediate states of the object, and outputs a quality measure while printing proceeds. Its examples include dimensions, surface finish, strength and warping. These are proposed predictions about the part; the patent does not supply independent evidence that any particular strength estimate is accurate in production.²¹²

This family began with Digital Alloys. The recorded chain for the parent application runs through Digital Alloys (ABC), LLC and Octavian Capital before the assignment to Relativity, recorded March 25, 2023. The intervening conveyances matter: finding Relativity on a later grant does not mean its employees originated every invention in the portfolio.¹⁸²

The technical distinction matters too. The disclosure includes resistive, or Joule, heating of wire. Ownership of this family is not evidence that this printing method is the same process as Stargate WAAM, or that it has been installed on the Terran R line.²

Break the sample, keep the engine part

At some point, predictions need a physical check. A witness coupon is a sample manufactured to represent a larger part. You can pull the sample until it fails, examine its material, or test other properties without sacrificing the actual component.⁵

The difficulty is making it a convincing witness. Relativity’s US20260233323A1 explains that different regions of a large WAAM build can experience different thermal cycles. Printing a small specimen separately, even from the same feedstock, may not reproduce what happened in the part. A miniature shape does not automatically reproduce its residual stresses or cooling history.⁵

The application proposes printing the coupon alongside the part, layer by layer, with an orientation and build volume selected to represent a relevant property. Fixtures provide mechanical and electrical connections; the disclosure also describes cooling arrangements. Fig.5A shows one of the fixtures used to attach the coupon support.⁵⁹

Published claim 1 requires concurrent layer-by-layer manufacture and includes the fixture’s electrical coupling and grounding. Claim 15 addresses a more specific arrangement in which the coupon fixture is mounted to the part and the electrical current path runs through those connected pieces. Again, these are pending claims. The engineering question remains whether the resulting specimen sufficiently represents the property being tested.⁵

The factory needs a memory

A rocket engine part leaves the printer with a history: where the robot moved, what the electrical signals did, how the material cooled, which defects inspection found, and how a representative sample performed. These filings attack different gaps in that history.⁴³¹²⁵

Our inference is that the durable advantage could lie in making those records useful across repeated builds. A factory that can connect a material problem to its cause has a route to improving the next part. Whether Relativity has achieved that at production scale requires evidence beyond these patents: validation results, repeatability, throughput and component performance.

For now, the records give us something concrete to watch. The printer makes the shape. The surrounding measurement and testing systems have to make the case for trusting it.