With a recent upgrade, HPE’s Spaceborne Computer-2 on the International Space Station is tackling increasingly complex artificial intelligence and machine learning assignments.
“It’s quite a capable little system flying around in space right now,” Mark Fernandez, HPE Space Technologies and Solutions chief scientist and Spaceborne Computer-2 principal investigator, told SpaceNews. “I haven’t declined any experiment requests because I lacked the physical hardware.”
The initial Spaceborne Computer, installed on ISS in 2017, was a test to see how a commercial off-the-shelf HPE processor would function in orbit. Its successor, Spaceborne Computer-2, pulled from HPE’s more rugged Edgeline computers, has performed dozens of experiments since it reached ISS in 2021.
About half of Spaceborne Computer-2 returned to Earth in 2023 for an update. The latest version, officially called Spaceborne Computer-2 but informally known as Spaceborne-2.5, traveled to the space station Jan. 30 on a Northrop Grumman Commercial Resupply flight.
Spaceborne-2.5 – with 130 terabytes of Kioxia storage, a modern operating system, new software and additional security features – is demonstrating some of the ways artificial intelligence is likely to be employed on future lunar and planetary missions.
SPACEBORNE-2.5
HPE is inviting researchers to propose scientific experiments and technology demonstrations for Spaceborne-2.5.
“I’ve got state of the art CPUs. I’ve got state-of-the-art GPUs,” Fernandez said. “I’ve got a modern operating system that supports Docker, and all the artificial intelligence frameworks and machine learning aspects that you may be interested in.”
Customers don’t pay for Spaceborne Computer services.
“We’re not allowed to charge for anything,” Fernandez said. “The other side of the coin is we can’t put anything into production on your behalf. We’re doing lots of proof of concepts.”
Spaceborne-2.5’s extensive Kioxia flash memory addresses one of the underlying concerns of space-based edge processing.
“I’ve got a massive amount of data at the edge. But even if I could process it, where am I going to store it?” Fernandez asked. “Now, I can store lots of DNA sequences. I can store lots of Earth observations. I can store lots of video. I’m super excited about that.”
The computer update has helped HPE and NASA speed up DNA monitoring. In the past, NASA researchers compared each astronaut’s DNA with the human standard genome. If an astronaut had a DNA anomaly, it showed up every day.
Now, HPE has uploaded the human standard genome to Spaceborne-2.5.
“I compare today’s genome of astronauts A to yesterday’s genome of Astronaut A,” Fernandez said. “It’s much better to say to the astronaut, ‘There’s been no change since yesterday,’ than ‘You’ve got the same anomaly as before.’”
AI TRAINING
Spaceborne-2.5 is testing AI for space applications. As experts noted in a recent SpaceNews webinar, AI models, which require extensive computational power and data storage capacity, are trained on massive datasets in the terrestrial cloud.
Cloud processing generates an inference engine, the core software that enables AI systems to make predictions and reach conclusions. Inference engines can be shared with space-based and terrestrial edge processors in a federated learning system. When federation members acquire new data, they can work together to update the underlying AI model.
“Replacing your inference engine completely could take a lot of bandwidth, but if you do it correctly with Docker layers then it could be fast and easy,” Fernandez said.
In a recent experiment, HPE linked an AI model running on Spaceborne-2.5 with the terrestrial AWS Cloud.
“This little old inference engine [on ISS] takes a guess at parameters [for the model] and sends them down to the mothership,” Fernandez said.
The cloud-based AWS mothership tests the parameters against massive datasets in the public cloud and says, for example, “It’s a little bit better, but let’s try I this way.” It then sends back another set of parameters to test. The two inference engines “chit-chat for hours, until they finally get into agreement,” Fernandez said.
Initially, the lengthy chit-chat sessions ended whenever communications between space and the ground were disrupted. Fernandez has written programming instructions that enable discussions to resume when satellite communications are restored.
DETECTING WATER
As spacecraft travel further from Earth, cooperation between edge computers and the terrestrial cloud or powerful space-based motherships will become increasingly important.
For instance, organizations looking for water on the moon are likely to employ multispectral sensors. Rather than sending the multispectral observations to Earth for analysis, companies or government agencies may send the data to nearby edge devices running inference engines that were trained on Earth to detect signs of water.
“In seconds I can tell you that this square meter is dry,” Fernandez said, “and your rover or instrument can go to a new location.”
As inference engines improve over time, lunar datasets could be reexamined.
“If I have hundreds of terabytes of storage and I’ve saved your raw data from the last 10 or 20 or 100 locations that you’ve examined, send me a new inference engine,” Fernandez said. “I’ll reprocess those and say, ‘Hey, we need to go back to site 17 and look again.’ All the raw data stays at the edge and the insight, the thumbs up or thumbs down, is what we send back to Earth.”
EXPANDING FEDERATION
HPE is preparing to install additional federation members on the space station.
“They’ll all be chit chatting back to the mothership,” Fernandez said. “The mothership will say, ‘I learned from Joe we should do this, but Bill said, try this.’ And the mothership, with great compute power and storage of the public cloud, can come up with the best set of parameters and then distribute that to all Federation members.”
“That’s in the works,” Fernandez said.
KEEPING WARM
Meanwhile, HPE is considering ways to commercialize space-based computing. Demonstrating SAP Hana database software in space is high on the agenda.
“Once we can get a commercially supported database in place, that will enable more commercial use of data centers in space,” Fernandez said.
HPE also intends to prove the resilience of space-based applications with HPE Serviceguard, software designed to keep applications running even if a hard drive fails, for example.
Later in the decade, Spaceborne Computer-3 is headed to the moon.
“I feel like I’ve been babied and coddled with all the Spaceborne Computers to date because I’m in the ISS,” Fernandez said. “It’s a whole other world when I get to the moon.”
Heat is not expected to be a problem for Spaceborne Computer-3. The model HPE plans to send has a certified operating limit of 55 degrees Celsius and temperatures at its planned lunar site reach 50 degrees.
The lunar night will be the challenge, though.
“Instead of having to cool my computer, I’ve got to keep it warm,” Fernandez said. “Our engineers are like, ‘Wait for decades, we’ve been worried about how to cool these things and now you’re asking us how to keep them warm on the moon with no air.’ When we get that solved, that’ll open up a whole lot of opportunities.”
This article first appeared in the June 2024 issue of SpaceNews Magazine.
