Can AI build a jet engine? The JARVIS Challenge examines the role of AI clones in hard technology engineering | MIT News

Artificial intelligence has revolutionized software engineering rapidly. Generative AI and large-scale linguistic models (LLMs) can create large volumes of code and documentation; machine learning algorithms can monitor performance and detect security vulnerabilities. But if the job is to conceive, design, and manufacture a complex physical system like an airplane engine, are those AI tools equally flexible?

This past semester, the JARVIS Challenge (Jet-engine AI Research and Validation Intensive Sprint) began testing whether AI can compress the construction test cycle, asking MIT undergraduates to find out if AI can help them build faster and better.

“The JARVIS challenge showed that AI can greatly accelerate the engineering of the most important security hardware, but engineering judgment remains an important distinction. An AI-native engineer is not defined by using AI, but by leading it – knowing when to trust it, when to challenge it, and how to translate the results of AI into functional hardware. Productivity – not limiting the engineering step or analysis – said Professor. Spakovszky, director of the MIT Gas Turbine Laboratory.

Teams, tools, work

The challenge gave the students four weeks to design, build, assemble, and test a small gas turbine aero engine, using AI as a key engineering partner. The goal: to build a “JARVIS-class” single-spool jet engine that produces 50–100 pounds of thrust, runs on Jet-A, and completes five 60-second runs. The teams had complete freedom over design, materials, and construction.

Representing almost every department in the School of Engineering, the 31 students were organized into seven teams, from all freshmen to senior teams. Many early competitors had little experience with turbomachinery, compressive flow, or, in the case of young students, even thermodynamics. Many had never seen the inside of a gas turbine before signing up to build it.

What they have: MIT machine shops and manufacturing vendors; commercial software including Concepts NREC, SolidWorks, and ABAQUS; and various test instruments for separating and assembling individual parts.

Teams also have access to MIT Parley, a newly launched platform that integrates major language models through a single interface. Through Parley, JARVIS administrators can see exactly how students have been using AI tools, including their content, cost per content, specific LLMs being used, and other important information. JARVIS is leading the way by providing early access to Parley for all participants, and with funding from MIT Lincoln Laboratory, the Department of Mechanical Engineering, and corporate sponsors Safran, Voyager Technologies, and Beehive Industries, students had access to unlimited AI implementations.

Funders were drawn by recruiting interest and genuine curiosity about how AI could reshape engineering workflows.

“We see this as the future of engineering,” Ryan (Hal) Hefron of Voyager Technologies told the students. “You’re honing skills that aren’t just fun to have — they’ll be the foundation of the future of the engineering workforce.”

Vincent Garnier, managing director of Safran Tech, watched the competition unfold with excitement. “JARVIS was a real experiment, a learning effort. Frankly we didn’t know what to expect, from the students or from the AI ​​models. What touched me from the students was: first, the enthusiasm for the experiment; then, as the project grew, they all came to a cool-headed realization of what the AI ​​could or could not help them with,” he said approximately that. “It makes me certain that this generation of leading engineers will probably not fall victim to using AI too easily and transparently, and will do so by staying connected to research – physical or mental experiments.”

Faculty leadership – professors Zachary Cordero, Zolti Spakovszky, Masha Folk, and Andreea Bobu from the Department of Aeronautics and Astronautics, as well as Lincoln Laboratory engineers and a team of teaching assistants – were on hand to ensure safety. In weekly progress reviews, they will critically assess student progress and evaluate how well students are using AI.

Spakovszky developed a careful approach to guiding groups in the right direction without giving answers or offering help. After the group presentation, he might ask: “Do you know what rabbet fit is?

Where AI helps and hurts

At the end of week 1, one team withdrew from the tournament; others, with varying degrees of success, have developed their own original gas engine designs. Different groups use AI to shorten textbooks, teach themselves how to use design software, source vendors, create Excel sheets, answer specific questions, find references, and create comparative analysis between design decisions. One team created an agent in Parley and assigned it the role of project manager.

In week 2, teams had to start working on detailed CAD designs, order parts, and prototyping their combus. This is where teams start to hit the limits in their use of AI. While Claude and ChatGPT have been good at providing design alternatives and filling knowledge gaps, teams have found that the stereotypes, sycophancy, and lack of physical understanding that have become popular features of AI production are undermining their confidence and holding them back.

“AI is a useful tool, good at finding information, helping to organize things, and it can write well, but it can’t design,” said Elizabeth Tupaj, a member of the 811 Crew. “When the engineer doesn’t know what’s going on and the AI ​​is in charge then the design is unreliable, at least with AI at its current capabilities.”

Teaching assistant John Zhang comments, “seeing this in person with the students reminded me how important the first impression is. If students can’t get answers from AI early, they quickly get frustrated and form a permanent impression that prevents them from using it later.”

In the final weeks, the finalists hit another hurdle that no AI could solve: working with vendors. “The AI ​​search found vendors we didn’t have a relationship with, who weren’t interested in our tight timeline,” the students reported. “The vendors they come to are the ones our team has a personal relationship with.”

Of the three finalists, only Fast and Fractured won the first attempt to ignite their mini-combustor. The team had used AI extensively in commercial studies and architectural comparisons, arriving at a viable design despite none of them having prior gas turbine experience.

“The JARVIS Challenge demonstrated what can happen when you combine AI-enabled design with motivated students and a culture of rapid exploration,” said Masha Folk, Charles Stark Draper Career Development Professor of Aeronautics and Astronautics. “The highlight moment was when the first student-designed combustor was installed in the lab. It burned flawlessly, ramped up to full power, switched to dual-fuel operation, and ended up burning continuously on 100% Jet-A fuel. This was proof that we could accelerate, build an engineering cycle while giving the students a real design challenge.”

In the early stages of AI engineering

By the end of May, the other two top teams – Fast and Fractured and the 811 Crew – had completed a full engine test. Fast and Fractured, with its AI-assisted design, was delayed by marketer headaches week after week, but was finally able to test. Unfortunately, their hot fire is cut short when the rotor scrapes and catches the stationary housing. The 811 Crew, however, more adept at turbomachinery and competition propulsion concepts, prevailed. Their engine started, successfully converted to Jet-A, and produced net thrust.

“As we stood there with the air-starter, hearing their engines explode and watching them spit fire, it felt like my heart was racing out of my chest. There are so many ways that could go wrong! What these students accomplished in such a short amount of time is amazing,” said PhD student Joe Chiapperi.

Team 811 has been reluctant to use AI throughout the tournament, relying on the basics and teamwork. “We had people who were at least familiar with design software, mechanical engineers who could build anything, and aerospace engineers who had learned to design engines specifically for gas engines,” said Tupaj.

Since the beginning of the JARVIS Challenge, younger students have been using Parley more often and more intelligently, while younger and older students have been using the knowledge in depth.

“JARRIS taught me that finding value in AI takes two things: enough expertise to judge what it’s telling you and catch it when it’s wrong, and enough curiosity to lean on it where it can help,” said Professor Andreea Bobu. “The team that went the fastest in the sprint had knowledge and relied heavily on AI to get there. The team that eventually won was able to resist AI; they had knowledge, but that hesitation slowed them down. The sweet spot seems to be knowing enough to always have a tool with them, and eager enough to pick it up in the first place. To me, that’s who directs the real engines: who will direct the real engines. The tools and the instinct to reach them.”

The clearest takeaway from the competition: engineering knowledge multiplies, and the human factor remains an important part. Knowing first principles and basic concepts breeds good engineering judgment and the ability to navigate difficult decisions when faced with incomplete information. And when it comes to building critical safety systems, there is no substitute for human hands and human accountability.

“JARVIS has shown that AI pilots can have a reproducible effect on engineering productivity, with judgment and first-order thinking that serve as great dividers between teams,” added teaching assistant Kyle Woody.

But the implications of AI in aerospace are significant. If small teams using well-managed AI pilots can compress build test cycles from years to weeks, the implications for workforce structure, R&D timelines, and competitive dynamics can be huge. The students who tackled the JARVIS Challenge are among the first engineers to tackle those calculations not as a thought experiment, but in a machine shop, with a jet engine in the test area.

“JARVIS highlighted the power of AI in the design of physical systems,” said Cordero, associate director of the MIT Gas Turbine Laboratory. “But it also showed that the key to unlocking that power is education, through courses, internships, and extracurriculars like MIT Motorsports and the Rocket team. Working on JARVIS is very much tied to a year at school. My takeaway is that in the age of AI, education is more important than ever.”