Tackling the challenges of AI in space
An AI engineer explains how a new breed of tiny, hyper-efficient ML models is revolutionizing space, extending satellite life and unlocking the next great era of autonomous exploration.
A multimillion-dollar satellite is alone in the void, hurtling through the most hostile environment humans have ever sent technology into. Deep within its memory, a single transistor—weakened by years of relentless radiation bombardment—is about to fail. The resulting “bit flip” could corrupt a critical flight command, triggering a "one-hit kill": a catastrophic failure that turns a marvel of engineering into space junk.
Once a satellite is launched, there are no service calls.
“The thing to remember about satellites is that once they're up there, that's it,” says James Murphy, AI Engineering Lead at Réaltra Space Systems. “Every other satellite has to be able to survive its mission duration, which might be 5, 10, 15, or even 20 years.”
While AI on Earth consumes megawatts in vast data centers, a new breed of machine learning is being built to solve this problem. These are tiny, hyper-efficient models designed to act as an onboard mechanic, constantly monitoring a spacecraft's health and predicting disaster before it strikes. It is AI designed for the ultimate edge: deep space.
Meet the architect: from dark matter to deep space AI
To build AI for space, it helps to have a background in both. James Murphy started his career in physics, working in a German lab on a dark matter detector. This experience spun his interest toward space, leading him to a master's degree in Space Science and Technology in Dublin ("the only degree of its type in the country," he notes).
The program's industry connections led him directly to Réaltra Space Systems, where his first job was designing the command and control software for camera systems on the European Ariane launchers. When the European Space Agency (ESA) later put out a tender for a project on spacecraft anomaly detection, Murphy saw an opportunity.
He arranged a unique, industry-based PhD to tackle the problem. The project was so well-aligned with ESA's goals that it received significant backing. “I essentially had one of the best-funded PhDs in the world,” Murphy recalls. “I had a half-a-million-euro project to fund my PhD work because it aligned identically with what the European Space Agency wanted the company to do.”
Why space is AI's hardest test
Deploying machine learning in space involves more than just creating a rugged computer. The environment itself is actively hostile to electronics. Besides the violent vibrations of launch, a satellite must endure huge thermal swings. The difference in temperature between its sun-facing and shadow-facing sides can be 200 degrees Celsius.
The worst offender, however, is radiation. “Electronics tend to react very strangely to radiation,” Murphy explains. “If radiation hits memory, you can end up with bit flips, which can destroy whatever your memory is holding. It can hit transistors and completely blow them up.”
This environment is paired with a severe communication bottleneck. Operators on the ground cannot simply stream all of a satellite's diagnostic data. “You can't send all of that telemetry back; you have quite narrow telemetry bands,” Murphy says. “Most of the time, you're getting about 10% to 20% of the available telemetry on board back to the ground.” This means any meaningful analysis has to happen on the satellite itself.
For decades, this onboard analysis has been handled by a system called FDIR, or Failure Detection, Isolation, and Recovery. It relies on basic statistical methods like thresholding—if a voltage exceeds X, raise an alarm. If an alarm is triggered, the system follows a rigid, pre-programmed "big decision tree diagram" to isolate the fault and attempt a recovery. This approach is purely reactive and cannot spot the subtle, complex signs of an impending failure.
Building models for extreme constraints
The primary constraint for on-board AI is power. A small satellite might generate only 20 to 30 watts in total. “Edge AI devices, while low-power for terrestrial applications, might use 10 to 15 watts, which is way too much for a satellite power budget of that size,” Murphy notes. This restriction fundamentally changes how AI models are developed.
You can't take a large, powerful model and simply shrink it. “We have to flip the table on how to develop these models,” Murphy says. “Rather than saying, ‘Okay, we've got the best model possible, now let's deploy it,’ we have to build our models from the ground up.”
This means designing for the hardware first. The team works with models that are orders of magnitude smaller than what most AI practitioners are used to. “For context, our models are about 50,000 parameters,” he says. This is a stark contrast to the billions of parameters found in today's large language models.
These tiny models are often a type of neural network called an autoencoder. In simple terms, the model learns the signature of "normal" operational data from the satellite's telemetry. It learns to compress this data down to its essential features and then reconstruct it. When new, live data comes in, the model attempts the same compression-reconstruction process. If the reconstructed data doesn't closely match the original input, it means the model has seen something it doesn't recognize as normal: an anomaly.
This approach offers two key advantages over old statistical methods. First, it captures the "inter-channel relationships" (the complex ways different systems affect each other). Second, it understands the "temporal aspect" (how a satellite's telemetry naturally changes across time), allowing the model to adapt as the spacecraft ages.
Protecting Ireland's first satellite
The perfect testbed for this technology was EIRSAT-1, the first fully Irish satellite, which launched in December 2023. Before launch, every satellite undergoes a grueling qualification campaign to ensure it can survive in space. This process generates a wealth of data, including data from real system failures.
“This meant we had a lot of data available from the EIRSAT-1 team where they suffered some real anomalies in their testing,” Murphy says. “It meant we had a labeled dataset that was 100% representative of the mission that was going to fly.”
The results were compelling. The team developed a model that achieved a 95% F1 score in anomaly detection while running on a 2-watt device. More importantly, its value was proven six months into the EIRSAT-1 mission when the satellite experienced a real anomaly.
“What we've shown is that our system would have been able to detect that with very high confidence if it were running,” Murphy says. The AI essentially acts as a "free satellite operator," providing tireless, around-the-clock monitoring that a human team cannot match.
From anomaly detection to autonomous exploration
The final challenge isn't technical; it's human. Getting experienced satellite operators to trust a new system takes time. “To them, the satellite is their baby, and you're trying to take control away from them, so they're not always happy about this,” Murphy acknowledges.
The immediate impact, however, is economic. For services like Starlink, autonomous monitoring is a necessity to manage thousands of satellites. For large Earth observation missions, where downtime can cost around $20,000 per hour, the AI promises to make services both cheaper and more reliable by extending mission life and increasing uptime.
This shift in operational dynamics from costly, manual oversight to cheaper, automated reliability is a pattern the space industry has seen before. In the last decade, a dramatic reduction in launch costs, pioneered by companies like SpaceX, didn't just make existing missions cheaper; it created the economic space for entirely new types of businesses to emerge, from satellite imaging startups to mega-constellations of space-beamed internet.
The autonomous monitoring technology developed by Murphy and his team might represent a similar inflection point, this time focused on the operations and longevity of assets already in orbit. By driving down the cost and risk of keeping satellites functional, it paves the way for applications that are currently not economically viable.
Looking forward, Murphy sees this work as a foundational step toward that future of true space autonomy. He draws a parallel to NASA's Ingenuity helicopter on Mars.
“If you look at the longest-serving rover on Mars, Opportunity, it covered about a marathon's distance—26 miles—in its 16-year life,” he says. Ingenuity, a small drone designed for Mars, was planned for five flights but completed 72. “It's done more exploration on the Martian surface than all of the rovers combined, from that one little innovation.”
That is the true promise of AI in space. It is a small, crucial innovation in autonomy that could unlock the next quantum leap in our exploration of the solar system. “There's this one little innovation right now,” Murphy concludes, “but we don't know what it's going to unlock in another 10 years.”