I spoke with Leap71’s cofounder Lin Kayser about Noyron last October, when he told me that he and his partner Josefine Lissner’s ultimate goal was to create a real world “Jarvis,” the fictional all-purpose engineering AI that works with Tony Stark in the Iron Man movies. Last fall, they had just successfully tested the TKL-5, a 3D-printed, 5-kilonewton rocket engine the company generated using the first version of Noyron. They used all the data they got during the development of the TKL-5 to feed it back to develop Noyron 2.0.

“Most companies would have focused on improving the existing engine, but since our goal is to perfect a computational AI model, we decided on a strategy to broaden the amount of data we would get,” Kayser tells me over email. “If you just test similar designs, similar thrust levels, and the same material, the data and experience you create is relatively narrow.”

To avoid this trap, Lissner suggested they should focus on creating a radically different—and extremely challenging—engine like the aerospike. Unlike the conventional bell-shaped nozzle we are all familiar with, the aerospike channels supersonic exhaust along a cone-like spike that extends outward. The shape tapers toward the back, ensuring that the exhaust gases flow along its surface, expanding naturally. This outer contour adjusts to atmospheric pressure changes and provides superior performance during a spacecraft’s ascent, from sea level to the vacuum of space.

The aerospike dream

“Traditional engines need a different nozzle length, depending on what altitude they fly,” Kayser tells me. On the ground, engines can make do with a short nozzle, but in the vacuum of space, nozzles must extends very long. “Otherwise the gas would go sideways and not produce thrust.” This makes vacuum nozzles for upper stages of space flight long and heavy, increasing the cost of launch of every rocket humanity has ever used, from the Saturn V to the SpaceX Starship. And because you can only optimize the nozzle for one atmospheric regime, a sea-level engine starts very powerful, but then gets less efficient as the exhaust starts to go sideways in the higher, much lighter atmosphere.

“You can actually see this very well in any rocket launch [see the video below], first there is a defined column of hot gas, and then, shortly before staging, the exhaust expands to a very wide cone,” Kayser points out. “Every gas molecule that doesn’t go straight down wastes a significant part of the energy which should result in forward motion.” That’s why current rockets have two or three stages, which complicates their design and cost. This is all avoided by the aerospike.

But despite the clear advantages shown in extensive studies and tests, the aerospike has not yet been used in real operational missions because its technical challenges have prevented it from reaching the level of reliability needed for active deployment in space missions. “Only a few teams have succeeded in hot firing in the last four decades,” Kayser points out. The main problem is the cooling.

“Since the spike sits in the middle of the 5,430°F hot exhaust gas, it is very challenging to cool it properly. Even extremely capable high-temperature metal alloys have melting points around 2,732°F and get weak way before that,” he tells me. It’s complicated to design an aerospike because it requires intricate internal cooling channels and precise geometry, making it difficult to fabricate using traditional methods.

In the ’60s, rocket engine design and manufacturing company Rocketdyne had tried to make the J-2T Aerospike, a toroidal aerospike concept intended as an improvement over the Rocketdyne J-2 engine for the Saturn V. Although promising in tests, the J-2T was never selected for operational use due to the complexity of implementing the technology.

NASA and Lockheed Martin pursued the idea again with the XRS-2200 linear aerospike for the now-defunct X-33 suborbital spaceplane program, designed to demonstrate single-stage-to-orbit (SSTO) technology. It was canceled in 2001 due to technical and budgetary challenges. Many others tried and failed, notably Firefly, a startup that went bankrupt in 2017 and recently came back to life after scrapping its aerospike design.

Some have succeeded, though, including Spanish company Pangea Aerospace, which successfully tested its methane-oxygen aerospike engine DemoP1 in November 2021. That company plans to scale up to the 300 kN ARCOS engine. The latest effort before the Noyron engine came from Polaris Raumflugzeuge, a German company that successfully designed and ignited an aerospike engine during a flight test of the MIRA-II demonstrator, over the Baltic sea.

Artificial intelligence to the rescue

According to Lissner, they were able to extend Noyron’s understanding of physics to address the unique complexity of this type of engine. Leap71’s AI autonomously created a design in which the spike is cooled by intricate channels flooded with cryogenic oxygen, while the outside of the chamber is cooled with kerosene fuel. “One of the results from the previous hot fire campaign was that our copper engine performed better than expected in terms of cooling,” Kayser says. They used the kerosene to cool down the outside of the combustion chamber, which prevented the copper from breaking apart (copper melts at approximately 1,984ºF, but begins to weaken at around 1,300ºF). The chamber pressure in the engines is really high, so the wall has to be able to withstand that.

“What we found in our tests in June is that the cryogenic liquid oxygen cools down the injector of our engine so much that the heat sink effect can be measured all around the chamber,” Kayser says. So the temperatures stayed very low around 284ºF). With all their test data, they updated Noyron’s thermal models. Lissner got to work adding the design logic for the Aerospike.

After the new version was done, it generated the design for the aerospike in a CAD file, ready for 3D printing in one single solid piece of copper. This is something that has never been done before, Kayser says. For 3D printing, they worked with German company Aconity3D, using a technique known as Laser Powder Bed Fusion.

This method melts ultra-fine metal powder, a special alloy made of Copper (the base metal, known for its excellent thermal and electrical conductivity), Chromium (added to improve strength and wear resistance without significantly reducing conductivity), and Zirconium (which Enhances grain structure stability and resistance to high temperatures). The laser melts and solidifies the powder layer by layer, enabling the creation of those complex cooling channels inside the spike and chamber.

Leap71 also worked with the Fraunhofer Institute for Laser Technology, which performed heat treatment to strengthen the metal, while a third company—Solukon, also based in Germany—was responsible for completely cleaning residual powder from the engine’s internal cavities. This is a vital step because any impurities or debris in the engine could cause it to fail or explode.

It worked

After manufacturing, Leap71 sent the aerospike to the Race 2 Space team at the University of Sheffield, which helped prepare the thruster for testing at the Airborne Engineering facilities in Westcott, U.K., a World War II-era concrete bunker that could contain any explosions if something went wrong. There was no way for the team to inspect it because it’s one monolithic solid object that you can’t open and look inside.

They basically had to test blind. “We were relatively sure about the theory behind it, but so many things can only be found out during testing, such as the resistance in the cooling channels, due to surface roughness of the 3D print,” Kayser says. Virtually everything in this engine was novel and untested. “We had data from our last fire, but this engine had much shallower angles, which usually result in surface quality. These angles are dictated by physics, so you have to live with what you get.”

Despite potential risks, the testing team successfully fired the engine on the first attempt on December 18, 2024. “It performed well on the first run, something we did not expect,” Kayser confesses. “We are very happy with the result. We are now cutting the engine in half to inspect it.”

The future is coming fast

Kayser says they did not get the oxygen flow exactly right, which is something you have to fine tune at the test stand. As a result, the engine ran a bit hotter than the team wanted, as the oxygen is the main coolant for the spike. “Instead of risking additional runs, we decided to analyze it and will feed the results of the analysis back into Noyron for further refinement,” he tells me.

Lissner views the achievement as “a great validation of our physics-based AI approach.” And indeed, it’s notable that a two-person AI software company spent about three weeks in total to design one of the most difficult rocket engines imaginable. Lissner says the team is now evaluating test data and feeding it back into Noyron to prepare for another round of tests. “Noyron allows us to drastically reduce the time needed to redesign and iterate after a test, and it helps us rapidly converge on an optimal design,” he says.