Exclusive: Behind the scenes at world first hot fire test of computationally engineered rocket engine

Westcott Venture Park, Aylesbury, Buckinghamshire. June 14th, 2024. 

For Airborne Engineering, this was a routine day. For engineering, it could be one for the history books.

LEAP 71, a Dubai-based company founded by Josefine Lissner and Lin Kayser, had developed a computationally engineered 5 kilonewton (kN) rocket engine with a thrust of 500kg and around 20,000 horsepower. It is a compact engine, but one that would be suitable for a final kick stage of an orbital rocket. Today, that claim was backed up.

Just before 6pm, the engine became what is believed to be the first computationally engineered engine to be successfully test fired, performing as expected and achieving ‘steady-state’, meaning it can be operated for as long as needed.

The engine has since been disassembled at the University of Sheffield, a collaborator on the project, and carefully inspected. The engine was undamaged by two hot fire tests (one lasting 3.5 seconds and the other lasting 12 seconds) with the thruster remaining in the UK for future tests.

TCT Magazine was given exclusive media access by LEAP 71 and Airborne Engineering to report on the very first hot fire tests of an AI-developed rocket engine. Here is how the day played out.

---

Ten of us gather around the rear end of a car. Two pieces of copper are in the hands of one. Twenty eyes stare down.

“It looks amazing,” a voice says.

“We need to drill out those holes as a first priority,” says another.

“We’ve been assured it’s been depowdered.”

From here, a debate commences. It’s 9.30am. Which gives this group of engineers eight and a half hours to put these two pieces of copper through multiple hot fire tests before a curfew designed to limit the volume of noise generated is put in place. If successful, it would be the first time – to the knowledge of anybody here – that a rocket engine designed by an algorithm went through such an examination.

LEAP 71 co-founders Josefine Lissner and Lin Kayser, supported by the University of Sheffield Race 2 Space team members Max Crawford-Collins, Henry Saunders, Alistair John and Oliver Dew, got the engine to this point. Airborne Engineering, the location of this get-together and a company specialising in the analysis of rocket propulsion systems, will lead the efforts today. Gravity Chief Design Officer Sam Rogers is credited with bringing the three parties together and is on-site today to chip in where needed.

All but Dew and John are present as we stand between workshop and testing bays. The rain begins to spit, and by the time a decision is made to head indoors, it’s lashing down. It’s 9.45am as we reclaim mugs of coffee and tune into Airborne Engineering Managing Director James Macfarlane delivering a lengthy oration of the health and safety protocols. All of which can be boiled down to, as a couple quip, ‘don’t be on fire’ and ‘the less screaming the better.’ 

But the tension amongst those involved in this project suggests, should there be fire, there will be screaming.

Fire is what everyone wants to see today. There’s a DSLR camera, a drone, and a journalist all here for that reason alone. To see flames erupt from the engine, perhaps for 15 seconds, but at least for two or three.

As Macfarlane finishes pointing people to the assembly points, warning people of the risks of an 80-year-old building and explaining the site’s flag system – red symbolises propellant on-site (liquid oxygen and kerosene today), yellow indicates a pressure test is ongoing, and both means a test fire is about to commence – engineers take up their stations. Outside, the two topics of conversation centred around a fixturing that needs fixing – in both senses of the word – and the potential for unwanted particles to be occupying the very thin channels of the engine.

The engine was designed with LEAP 71’s Noyron large computational engineering model via PicoGK, before a 3D printable file was transferred to Munich, where AMCM printed the two-part engine (combustion chamber and injector) on an EOS M 290 platform in EOS CopperAlloy CuCuZr. It was then shipped to the University of Sheffield for final machining and, in between, it is said to have been depowdered with Solukon de-powdering technology.

But here, at Airborne Engineering’s facility in Aylesbury, UK, with just eight hours left to conduct the hot fire testing of the engine, doubts begin to creep in. Some are confident the de-powdering process will have done its job, but others are asking the question, ‘what if it hasn’t?’ No scan has taken place since the depowdering process, so there is no way of knowing if all the powder has been pulled out of, or if swarf has found its way into, the channels. Channels that, in swirling around the engine jacket to cool the engine down, measure just 0.8mm at their narrowest points.

The solution is multi-faceted.

It’s 10.15am. At the back of the workshop comes a very dull, very loud, humming noise as the combustion chamber is submerged in a digital ultrasonic cleaning system. The machine has been set to run for seven minutes, but after just five the clear IPA solution has already turned a murky yellow colour, suggesting something was in the manifold.

Macfarlane suspects it is machine cutting fluid which, “wouldn’t make any difference to the kerosene side – kerosene would just clean it out immediately – however, on the oxygen side, it could cause ignition to happen inside the injector, so it could combust with oxygen when you turn on the engine. It’s a small chance but it’s best not to [risk it].”

The other concern is that any unwanted particles in the engine – swarf for example – could block a channel and cause foreign object damage. After a couple more minutes, some specs of swarf begin to rise to the surface.

Moments later, at the other side of the workshop, Rogers is on the search for some tape so he can cover the face of the manifold and blast through the channels on the other side with an air compressor. As this improvisation takes place, it is revealed that the engine being passed around the workshop is technically not completely finished. There are inner bores that were supposed to be machined down by 2.3mm, but the machinists ran out of time, though this may have actually helped to reduce the amount of cutting fluid and swarf in there.

After the air compressor is blown through the manifold, it is due to return to the IPA bath.

‘Is there anything I’m delaying by playing with this?’

‘We want to fire it later today.’

10.45. One by one, probes and sensors are being plugged into the ports of the engine; the engine that now resembles the antagonist in the Hellraiser movies. This is among the first of a series of fiddly, yet very important, jobs that will be undertaken prior to any hot fire tests. These sensors will transmit valuable data back to the LEAP 71 team, helping to inform its computational models for future rocket engine development.

“The tip of this probe touches the copper wall,” Lissner explains, “and some of them actually poke into the fluid, so it conducts the heat out of the engine and then it gets measured. This will track the temperature change along the channels. Even if all of this doesn’t work, we can extract the data at least and build on it.”

This doesn’t stop the nerves, however.

Above the engine, there are discussions taking place about which sensors go where, which will be checking the temperature of the liquid and which will be checking the temperature of the surfaces. Due to time constraints, half of the ports are being filled with blanks.

‘This is scary. Oh Gosh. Will everything be ready in time?’

‘They're about to plumb it up and then it will happen pretty fast. The unknown bits are slowly filtering away.’

11.15am. The nerves come from a lack of control. For a couple of weeks now, the LEAP 71 team has had to sit back and watch others take the computationally engineered 3D print and turn it into a functional, test-ready engine. It was developed with a new ‘Large Computational Engineering Model’ – Noyron – with Lissner inputting a base set of parameters, some of her own aerospace engineering knowledge and information from 120 academic journals. “It goes directly from theory into practice,” Kayser notes, as he explains Noyron is capable not only of generating manufacturable geometry, but also predicts performance and outputs procedures for production steps and technical documentation.  

Today, then, is potentially a breakthrough for computational engineering. Through LEAP 71, Lissner and Kayser are hoping to demonstrate an advancement in design practice by leaning on the power of AI. And today, they’re hoping to give AI a helping hand by test firing a rocket and feeding the findings back into their models. The only way you can build an AI for engineering, Kayser says, is by translating knowledge manually, validating it, and making the platform coherent. As a starting point Lissner and Kayser simply entered a few input parameters, including propellant type, engine thrust, and chamber pressure.

The resulting engine being passed around the workshop certainly looks the part, and its design method has granted LEAP 71 such flexibility that they were still making significant changes to its functionality just weeks out from test day.

“What we are trying to do is create a robust AI-built functioning rocket over a broad range of thrust-level propellant types,” Kayser says. “We literally changed the propellant type, which is usually the first thing you agree on, two weeks out. We originally designed it for isopropanol alcohol and nitrous oxide, which are relatively benign propellants, and two weeks before we went to manufacturing, we said: ‘what are we doing? We have a test site that can do kerosene and liquid oxygen (LOX), let’s change it.’ And the rocket redesigned itself for the change, we printed it, that’s where we are today.”

Lissner and Kayser think they’ve put everything you need to build a rocket engine into one coherent algorithm. But, Kayser admits, whether it will work is another question.

It’s a question that will hopefully be answered today. The search for test site was an arduous one. Such is the rich activity in rocketry, test stand waiting lists are growing larger and larger. Outside, there are multiple anecdotes about engines – increasingly being developed with 3D printing technology – having a development cycle of two months but having to wait up to 18 months to be tested.

It's a source of frustration for everyone here. There are engineers on-site decades into their career, and some just a few years. Those who were inspired by Delta Clipper, and some inspired by Tintin. Kayser’s uncle was, in fact, the founder of the first private space company back in the 1970s. He grew up with rocket engines on the walls but has never been near a rocket test stand before and doesn’t love how long it can take to get here.

“These cycles will accelerate,” Kayser says. “I mean, we have literally changed, two weeks before manufacturing, the propellant and the thrust.”

The propellant is currently filling the tanks at the back of test bay J1 and a poster with all the relevant branding is being positioned on the side wall, so that when the hot fire test happens – if it happens – all those due their credit receive it.

Lissner is pulled away from that job to discuss the priority order of tests. A cold flow test has been agreed upon to help clean out any residue from the channels and to assess how accurately the computational model has predicted the engine’s performance. The next test will be a three-second hot fire test, which won’t see the engine reach complete equilibrium because the cooling channels will still be warming up, but will inform everyone here if something in the engine isn’t working as it should. Once a three second test is out the way – and time willing – a five second test is recommended (the engine will reach complete equilibrium here), before moving on to the long duration test fire of about 15 seconds.

Because of how long it has taken to get to this point – it’s 11.45am already – and how much more there still is to do, a high throttle test looks likely to get bumped up the order. A higher throttle test will generate more heat, but it also gives ample time for the engine’s cooling mechanisms to kick in, whereas a lower throttle test won’t heat up the engine as much, but its cooling mechanisms also won’t perform to their maximum capability.

Lissner explains: “The risk of a nominal design: so, you say five kilonewtons mass flow, that is what I designed the engine to, but then when you throttle, you actually change it to higher or lower. If you do lower, you have less cooling, so it tends to get hotter, and if you go higher throttle, you have higher pressure so it tends to burst. These are the things we’re trying to figure out, which one’s lower risk? We’ve tried to arrange the testing such that the risk of making a full explosion was delayed.”

‘A short run, we’ll learn everything.’

‘I guess three seconds is a long time when you’re watching.’

‘It’s enough time to see if a channel is blocked or any melting happens.’

12pm. Documenting the progress, Rogers saddles up next to the engine on the test stand to deliver a piece to camera, pointing out where the LOX will come from, where the kerosene will come from, and the anatomy of the engine itself. “And we’re almost ready to fire. Well, we’re not really almost ready to fire, are we?”

No. More plumbing needs to be done, lunch needs to be eaten, a cold flow test needs to be completed and then the test bay needs to be reset. Between each test, checks need to be made to ensure the safety of the subsequent hot fire, though quick runs shouldn’t mean the tanks need refilling. A long duration test, however, will use up all 14 litres of the kerosene tank, with this taking at least 15 minutes to refill. Estimations are we’re about three hours out, at best.

‘These channels are very fine.’

‘That’s why I wanted to cold flow it.’

‘I can blow through it, so there’s air coming through it.’

‘Do you reckon you could do 30 bar?’

12.30pm. Out the way of the tinkering, the conversation returns to the LEAP 71 PicoGK computational engineering framework. Lissner explains to those within earshot how the framework is like object-oriented programming, whereby a user defines the property of a component – for a cog, let’s say you need a thickness, inner radius, outer radius and a number of ‘teeth’ – and then the computational engineering model will take over. It will provide methods on how to construct the part in a given position or angle, it will take into consideration the transmission ratio to another gear – this can be provided to the model as an input – and it will build the ‘network’ so that all the components that need to talk to each other can.

Kayser contributes the company’s motivation, explaining how he and Lissner believe engineering is ‘unbelievably slow because it has never been computerised.’ Lissner works off a MacBook Air – a computer without a fan – and can generate models in an hour, though computers with higher specs could generate models in as little as 15 minutes. The LEAP 71 co-founders have also been known to work off a second computer while their primary workstations ‘are busy’ developing those models.

The dynamic is working for them. It has got them here, after all.

‘You’re essentially writing the maths?’

‘It’s not that different to using a CAD system but you’re doing it computationally.’

‘I can’t even really define what I’m asking.’

1pm. Seven people are now huddled around the engine. It’s akin to an operating theatre, with tools being passed back and forth to those in charge. Each bolt is being tightened and each hose is being double checked that it’s plugging the right port.

“This is the igniter,” Lissner points out. “That blows a small torch light into the injector and then helps ignite the propellants once they arrive in the chamber.”

‘It feels like a level of magic that the fluid can go in one end and be out as hot gas. It doesn't feel like it has enough time to do any vaporising, let alone combusting.’

‘It’s insane. And you think, this has 22,000 horsepower. Really? This thing.’

’22,000?’

1.15pm. That is all set to happen at some point today in this test stand – test stand J1.

For safety reasons, the majority of those in attendance will be watching via camera feed in the neighbouring J2 test stand. We’ll get a pretty good view of the hot fire tests from that live transmission, according to the Airborne Engineering team, but earphones won’t be required. We’ll hear it, alright.

‘And if we were to go behind the brow of that hill and peek over the edge?’

‘If you have a line of sight to the engine, the engine has line of sight to you.’

‘Safety goggles?’

‘I’m not concerned about your eyesight.’

1.30pm. Breaking from the camaraderie, the hard work continues. For each hose that’s plugged into the engine, cursory checks need to take place to ensure the data will be transmitted back to an Airborne Engineering workstation on-site. Engineers are reminding themselves to at some point fit the blanks, they’re checking the engine is bolted firmly to the test stand fixturing, and they’re often coming back to the same question: Will this 3D printed engine, with its thin channels, with the potential for interior surfaces to be rough, withstand the tests?

As the day goes on, the main concern appears to be the surface roughness inside the engine. Because liquid oxygen is highly reactive, any loose chips of copper in the channels could get caught in the liquid oxygen stream, hit the wall and create an unwanted spark. You don’t have to be a rocket scientist to guess what’s next, but as it happens there are a few onsite: ‘There goes the engine.’

A cold flow test is now imminent. With the mass flow set to 0.67kg per second, this procedure will blow nitrogen gas through the engine to check the structural integrity of the engine, confirm the mass flow rate, and validate the pressure flow versus the prediction of LEAP 71’s model.

For this test and all subsequent tests, it is explained that an Airborne Engineering team member will count down on the radio, from ten to six, before falling quiet to allow others to warn of any danger in the remaining five seconds before the test firing.

‘I don’t trust channels below about a millimetre and a half to be honest. Mostly because, if it's for any great length, depowdering it will be difficult.’

‘What pressure drop are you expecting down those channels?’

‘Me or the model? The surface roughness compared to the channel dimensions is significant.’

1.45pm. The first countdown commences. Eight of us stand outside. Two more are in the back of J1.

There’s a screech. A gush. And then footsteps. At the back of J1 are two monitors; one showing a graph and another showing a spreadsheet. To the untrained eye, the lines and numbers don’t mean much, but there is one data point on the graph that stands out. Mass flow is 0.583 kg/s and pressure is registered at 24 bar.

The pressure drop was much higher than anticipated, meaning the throttle valve had to open fully because there wasn’t enough pressure in the tank to get the mass flow through the engine. LEAP 71 later confirms it will modify Noyron to output slightly different cooling channel geometry for future engines.

A second cold flow is conducted minutes later. This one is more promising. Mass flow was registered at 0.654 kg/s – just shy of the targeted 0.67 – at 15 bar.

First impressions are that the initial cold flow test may have cleaned out the engine, helping it to perform better the second time around, while the surface roughness around the nozzle is mooted as the most likely reason for the surprise pressure drop.

‘We’ve done some experiments where you’re looking at the pressure drop as it goes along the channel, the straight bits are usually not too bad.’

‘We’re not only going around the nozzle, we’re going around it sideways.’

‘It’s a longer flow.’

‘As long as we get flames…’

It's 2pm as we break for lunch. What’s about to happen next is the LOX and kerosene tanks will be refilled. The hosing couples will be checked to make sure they’re still where they need to be. And another check will be done to ensure they’re still set to transmission the data when the real tests happen.

Lissner and Crawford-Collins are already looking beyond that. There has been an acceptance all day that things might not go to plan here, and bringing together such a variety of perspectives has only compounded that discourse. Some are more confident than others in 3D printing technology, about the size of the channels, even about the ability of a computational model to design an engine and accurately predict today’s outcome. It’s the first time this has ever been done, remember.

She notes how much of a difference there can be in the pressure drops when the small channels are rougher than anticipated, while he suggests placing a sensor at the injector next time to monitor the pressure drop there before reaching the channels.

“It’s designed to have minimal pressure drop on these small inlets,” Lissner explains, “so it’s not designed to have a pressure drop going into the channels.”

“A few of the angles on that outlet, for example, are 105 degrees, but if they're perfectly at 90 degrees, it halves the amount of [machining] time and it doesn’t affect [the performance] as much,” Crawford-Collins says. “Things like that would be interesting to build in, designing for machining as well.”

“We were told to reduce every hole by 10% and that meant everything has to be machined,” Lissner adds. “If I had known how you do it, these ones I don’t need to reduce because you can just tap them.”

“You can spend five years reading articles,” Crawford-Collins sums up, “or you could just do [a test] and find out.”

2.30pm. Everyone has dispersed. Some are queueing for the microwave, some are heading down the road to the nearest café, and some are still loitering around the engine.

A simple question is asked. Since this is an experimental test that will be used to inform future computational engineering models, do you want this to run perfectly? That is to say, will more be learned if things don’t go exactly to plan.

Kayser appreciates the question: “If it just works, we don’t have to change the algorithm. Ideally, it doesn’t blow up. You saw all of these sensors. We have predictions from the model. So, we know what pressure to expect where, we know what temperature to expect where. If the model is right. Now, we will get the actual temperature measurements, which will be off, of course. Will they be 1% off? 50% off? Will it just blow up? We hope it will be a few percent off. Once we have that, we can calibrate our model. And then, if we get the data from the very different engine [to be tested in South Korea], which is much smaller, different materials, etc, to get more data points, we can make the model way more robust. If you test enough rockets, you’ll have a model that works.”

The engine to be tested in South Korea later this summer is developed from the same algorithm, but has been designed for satellite propulsion and has a size of 1 kilonewton (kN) compared to the 5kN engine we’re testing today. It will go ahead only if things go to plan here.

‘Hopefully not too much green stuff.’

‘Green?’

‘When copper burns. Green.’

3.15pm. In the distance the red flag is coming down. But only to allow a yellow flag to go back up with it. It’s a promising sign.

Inside J1, more tinkering. But potentially the last of it. The LOX line is being cooled down, which will take 15 minutes. Then, it’s a quick cold flow through the injector. And finally, one more ignitor test.

4.30pm. Like most things today, it takes longer than expected. But we really are ready as those in attendance head into test stand J2.

The camera feed transmits, from right to left, the backdrop of a neighbouring facility, the doorway of test stand J1, a poster adorning the branding of Leap 71, University of Sheffield, AMCM, PicoGK and Noyron, and finally a computationally engineered two-part engine 3D printed in copper.

Slowly, there is smoke creeping out of the engine’s front. Over the radio, Mcfarlane begins the count at 13 as ten others wait with bated breath. He reaches six. The radio falls silent. Rogers continues the count from five. There’s more smoke now. But only smoke. For a couple of seconds more, it’s still just smoke. And then it comes. A bright burst of fire. The engine’s noise is palpable. And as the flame disappears, so are the screams inside test stand J2. It worked.  

An initial 3.5 second hot fire test is completed successfully. It used an oxidiser-to-fuel ratio of 1.8, lower than the nominal 2.3, but the engine withstood the test, paving the way for a long duration run. If they can fit it in before the 6pm curfew.

Already, though, it is the first rocket engine built through a computational model without human intervention to be hot fire tested. Because of that, it is quite likely that the two weeks between the model being finished and printing commencing was the shortest time from spec to manufacturing for a new rocket engine. It was already the first liquid rocket engine believed to be developed in the United Arab Emirates. And, most importantly, it worked on the first attempt.

‘It’s been quite a morning.’

‘I wouldn’t say it’s our record in getting to firing.’

5.55pm. It is now or never. After allowing those on site to run across from test stand J2 to test stand J1 to get a visual on the cooled-down engine, the routine checks were made to ensure it was safe to do another test. Plumbing, sensors, timing. By now, you know the drill.

And so here we are. Just five minutes to throw the remaining volume of propellant through the engine, this time at the nominal oxidiser-to-fuel ratio, and see if the engine can withstand it. Luckily, it will only take 12 seconds.

Everyone heads back to test stand J2. It has taken all day, and it really has felt like it.

But history will be made in ten, nine, eight, seven, six…

---

Read more: 

“3D printing is a science experiment right now. It's not a manufacturing process.” – TCT speaks to Leap 71

The economics of additive manufacturing are broken — here's how to fix it

Students at University of Sheffield break record with 3D printed rocket engine

---

Watch the hot fire test below.