Space loves AI, AI doesn’t love Space – SpaceNews

author
4 minutes, 29 seconds Read

Space-related applications of artificial intelligence and machine learning are often confined to the ground because moving AI onboard satellites, while promising, is significantly more difficult.

“Running AI in space is like running a marathon on the moon — impressive if achieved, but limited by the environment,” said Sylvester Kaczmarek, chief technology officer at OrbiSky Systems, a London startup focused on AI edge operations.

Advanced processors are power hungry, meaning satellites with onboard AI require large solar panels and extra batteries.

Plus, “radiation in space can fry electronics,” Kaczmarek said.

Power management is another problem. Many AI devices require very high currents at low voltages.

“The demands that AI devices place on power management are new to the space industry,” said Bert Vermeire, Voyager Space chief technology officer. “There are no good solutions with space heritage and it is difficult to identify efficient and small form factor power and management parts that can supply power to the AI devices” in space.

Software must be modified to run on satellites as well.

“Space missions demand AI techniques that can crunch data with limited power and memory,” Kaczmarek said.

Simply put, “Space loves AI, but AI doesn’t seem to love space,” Ralph Grundler, Aitech Systems director of space business development, said in February at the SmallSat Symposium in Mountain View, California. “The electronic companies producing these AI chips haven’t really taken space into account.”

AI-optimized chips are not designed for electronics that “may have to sit on the launch pad in Florida in 100-degree weather for two or three weeks and still work when it gets up to zero atmosphere and minus 40 degrees,” Grundler said.

Upgraded brains

Still for certain missions, the potential benefits of onboard AI are too impressive to ignore.

Instead of downlinking enormous ocean scenes from an Earth-observation satellite, an algorithm trained to spot ships could downlink ship locations, sizes and headings directly to the Coast Guard, said Luis Gomes, AAC Clyde Space CEO.

The Hewlett Packard Enterprise Spaceborne Computer 2, installed on the International Space Station in 2021, has supported dozens of experiments. One experiment applied artificial intelligence to analyze astronaut gloves. Credit: NASA

Onboard AI also could improve spacecraft performance. If a satellite experienced a latchup, a type of short circuit, the onboard AI could identify the problem and remedy it through power cycling or other means, said Andrew Haslehurst, Surrey Satellite Technology chief technology officer, said at the SmallSat Symposium.

Further autonomy supported by AI will be particularly important “as we get into more complicated long-duration platforms or long-distance missions where human interaction will be limited,” David Barnhart, director of the University of Southern California Space Engineering Research Center, said by email.

Even in Earth orbit, the latest remote-sensing satellites gather far more data than they can quickly downlink. Devices optimized for AI and machine learning can help by processing or pre-processing data.

Then, satellites can transmit the most important datasets first and compress the remaining data for onboard storage, said Ken Hermanny, vice president and general manager for Mercury Systems Microelectronics Division Mixed Signal Business in Massachusetts.

“The satellite’s processing brain is being upgraded from the technology of the past,” Hermanny said. “We can customize the functions while it’s in orbit.”

Creative solutions

To help satellites take advantage of AI, companies are shielding terrestrial components and space-qualifying AI-optimized chips and circuit boards.

Mercury Systems, for example, co-developed with Ball Aerospace a space-qualified processing board for field programmable gate arrays, integrated circuits that can be reconfigured for countless applications.

OrbiSky, meanwhile, is developing new components for AI processing on spacecraft and drones. “We call them high-performance, secure AI brains for machines,” Kaczmarek said.

Singapore-based Zero Error Systems produces hardware and software for space-based electronics. ZES’ Latchup Detection and Protection monitors commercial-off-the-shelf (COTS) integrated circuits. When a latchup is detected, it resets the power.

Another ZES product combines hardware and software to detect and correct errors in commercial memory devices.

“We believe COTS is the way to go,” said Wei Shu, Zero Error Systems chief technology officer. “It allows customers to use the most advanced commercial devices in space. But they need to be protected.”

Weeding out devices

Voyager Space Technology Systems, the former Space Micro, works with universities to investigate the impact of radiation on AI-optimized chips or tiles, which combine different type of processors.

With the latest devices, “it’s very difficult to characterize exactly how things will go wrong in a radiation environment and how to mitigate that,” Vermeire said.

To find out, Voyager begins by weeding out devices that exhibit “destructive single event effects, meaning a single particle going through them in the wrong place at the wrong time destroys them,” Vermeire said.

Additional testing focuses on understanding how to mitigate “soft errors,” problems that occur, for example, when high-energy protons strike spacecraft. “We have to do something like shut down and come back up cleanly,” Vermeire said.

Mission dependent

What’s the best approach to bringing AI onboard satellites? It depends on the mission.

Some satellites in low-Earth orbit could take advantage of the latest commercial-off-the-shelf electronics.

Satellites in geostationary or cislunar orbit may require space-qualified parts, shielding or other creative solutions to mitigating radiation’s impact.

In some cases, satellites will be redesigned to make room for the extra power and advanced thermal control systems AI processors demand.


This article first appeared in the March 2024 issue of SpaceNews magazine.

This post was originally published on this site

Similar Posts