AI can help us explore new planets and build better satellites, but the hazards of space demand specialized hardware—and right now, we’re lagging behind.
From ChatGPT to deepfakes to data mining, artificial intelligence (AI) is everywhere these days. But its spread toward the final frontier has barely taken its first baby steps. Space is unfriendly to computers, and getting AI-capable chips up for the task takes time and money.
“There is frequently a 20-year technology gap when it comes to the computers you can send to space,” Ken O’Neill, space systems architect at AMD, told engineering.com. “In space, you have a lot of particle radiation, electrons, protons, heavy ions, as well as gamma rays and X-rays, and all of these things have a particularly harsh effect on microelectronics.”
A handful of companies, however, are working to send AI ad astra—and their work will open up new frontiers for the digitally transforming aerospace industry.
Making AI chips fit for space
Computers in space die faster and don’t work as reliably as those on Earth, unless they are designed to survive in the unforgiving environment. So-called single event effects triggered by strikes from energetic particles or cosmic rays can cause software errors and, in some cases, even destroy circuitry. Add to that the sweeping temperature differences that satellites are exposed to as they zoom in and out of sunshine multiple times a day. When not illuminated, spacecraft orbiting Earth face a cold of minus tens of degrees Celsius. When they cross into the sunshine, the temperature soars to over a hundred degrees Celsius. The rattle of the rocket launch adds an extra challenge for the electronics.
To be able to sell their chips as “fit for space,” manufacturers have to prove the devices are compliant with specifications outlined in a document called MIL-PRF-38535, issued by the U.S. military. The qualification process, O’Neill said, takes up to three years and involves rounds of testing including temperature cycling that simulates the frequent temperature changes in orbit, shaker tables that recreate the uproar of the launch and irradiation chambers that allow the engineers to verify the devices’ ability to withstand radiation. The testing itself is preceded by years of development and engineering that ensures the chip is fit for the ride.
In November 2022, AMD released what they describe as the first commercially available space-qualified chip capable of running complex AI and machine learning algorithms in real time.
The journey to the final product, the radiation-tolerant XQR Versal AI Core XQRVC1902, produced 50 patents, O’Neill said, focusing mostly on improving the chip’s ability to survive the high levels of radiation present in space around Earth. The company followed up on their first space-grade product in September 2023 with the Versal AI Core XQRVE2302, which is significantly smaller and about 75% less power hungry.
“We choose our materials and construction techniques based on our previous experience, knowing what would work and what wouldn’t,” O’Neill says. “The choice of materials has to be such that [the chip] will survive the temperature cycling and the high temperature during the operating life test. We have to be careful about the choice of materials for the integrated circuit and for the packaging of the integrated circuit.”
O’Neill added that the space-qualified chips are “substantially” more expensive than their equivalent non-space qualified counterparts, but still considerably cheaper than alternative bespoke solutions that had been available before.
“It’s a leading edge technology,” O’Neill said. “It’s on a 7-nanometer manufacturing technology and it has the same performance specs as the commercial Versal parts.”
The importance of the edge in space
Edge computing can be a huge boon for space technology, explains Henry Zhong, co-founder and head of AI at Spiral Blue, a Sydney-based start-up developing edge-computing platforms and applications that can be used in orbit.
“Edge computing is very useful in space because it’s very difficult to send the data that is collected in space to Earth,” Zhong told engineering.com. “It’s a lot of data. The data is recorded on the satellite, but we don’t have enough bandwidth to send all of it back to Earth.”
Earth-based computers, for example, are used to run AI algorithms that scour images of our planet captured by remote-sensing satellites and look for patterns based on the training they received. But because of the limited downlink availability, satellites only take images when tasked. In many instances, important information doesn’t reach Earth at all or only with a significant delay.
O’Neill cites wildfire detection as an example of an application where AI in space could make a huge difference—limit damage to the environment, save cost and possibly human lives.
“Currently, you may not know about a wildfire in a remote area until it spreads into a large size,” O’Neill says. “With Earth-observation satellites, if you have the AI on board, you can detect the presence of the wildfire earlier. Maybe a few hours earlier, maybe a few days. That not only gives you more time to evacuate people, but it would also help contain the wildfire before it spreads too much.”
Similarly, intelligent algorithms could sift through images and delete those covered in clouds or alert operators to any unusual and unexpected events that might be taking place too far from humans’ immediate reach.
“Only about 10% of the raw data currently gets sent down compared to what the satellite is actually capable of capturing,” says Zhong. “But when it comes to disaster monitoring, you want to be able to assess the situation in a timely manner and make decisions that are frequent and up to date.”
Taking a risk on AI hardware
Spiral Blue has a less perfectionist approach to AI in space than AMD. Embodying the new space ethos of flying cheap things fast and fixing them later, the company has launched four of its custom-designed Space Edge One computers fitted with AI capable chips since January 2023. Only one of these computers is still operational, Zhong said.
“Our goal is to get cheap commodity off the shelf hardware, integrate it into a satellite and then launch as many of them as possible because some are bound to fail,” Zhong says. “They are not designed specifically to be operated in extremely harsh environments, but so far most of them seem quite robust.”
One of Spiral Blue’s computers is circling Earth aboard a satellite of the U.S.-based Earth observation company Satellogic. The AI chip at the computer’s heart is the Nvidia Jetson Xavier NX, a mass-produced, widely available system-on-module (SoM) released in 2019. The chip sits on a custom-designed carrier board, which is then integrated into the customer satellite.
Zhong said the company took their chances with the kind of radiation hazards that AMD so meticulously works to mitigate. Spiral Blue’s engineers designed a copper plate that covers the hardware to dissipate heat and reduce the influx of charged particles. The resulting computer was put through radiation testing, the company said, and did well, giving Spiral Blue sufficient confidence that their Space Edge Computers could survive in space for up to five years. And that might be enough for makers of small satellites and cubesats that are designed and launched with much shorter intended mission times than old-school agency-built satellites.
AMD, on the other hand, guarantees the users of their radiation tolerant XQR Versal AI Core chips up to seven years of flawless operations.
Early days for space-based AI
According to O’Neill, it will take some time for AI to make a major mark on how things are done in space. Right now, he says, academics on all fronts research and speculate about the possible uses of AI in space, but not many practical applications have been developed.
“We’re kind of right at the beginning of the use of AI in space,” O’Neill says. “The demand at the moment is quite low but growing very strongly. Most of the interest comes from research institutions and not much of it has translated into operational systems yet, but it’s certainly happening.”
The possible applications go beyond analyzing images of Earth in space. Maintenance software could monitor and evaluate satellites’ health in real time based on measurements of on-board sensors. Fully autonomous algorithms could help probes visiting distant planets to select the most suitable landing sites and bring the spacecraft to a safe touch down. Clever programs could help satellites dodge space debris.
O’Neill says scientists are intrigued by the possibility of reprogramming chips and adjusting their functionality remotely from Earth, allowing them to finetune algorithms as new discoveries are made and new questions arise.
“There’s so much flexibility to modify and evolve the application even after it’s already been deployed in space,” O’Neill says. “That tends to be very beneficial to our space customers.”
