Space missions are hard. Consider the last two SpaceX Starship launches in November and April; both exploded. In January, a misaligned fuel valve caused the second stage of a Virgin Orbit (VORBQ) rocket to be starved for fuel. The launch failed, and shortly thereafter Virgin Orbit went bankrupt. Successful space missions are numerous—2,474 satellites were successfully launched in 2022—but tricky. The key seems to be getting the right recipe for all the complexities and then repeating it exactly. That’s where artificial intelligence (A.I.) comes into play.
Space missions are also expensive, which makes the hefty cost of A.I. tolerable. The satellite launch vehicle market was $14.4 billion in 2022 and it’s expected to grow at an annual rate of more than 12.5 percent in the next decade, according to Global Market Insights. That works out about $5.8 million per satellite, and that’s just the launch. The manufacturing costs of satellites dwarf the cost of launching them. A spy satellite costs around $100 million, a weather satellite upwards of $300 million, and government and military satellites can cost anywhere from $500 million to $2 billion. So, at a couple of million dollars, even a significant A.I. project is worth it.
A.I. has been used for space exploration by NASA for years. It is estimated the space agency spends $200 million on A.I. each year. The most famous use case is probably the Mars rover. NASA also has used the technology for spacecraft operations, spacecraft health, on-board image processing and land-based data processing.
Major satellite manufacturers like Boeing, Lockheed Martin, Northup Grumman, Thales Alenia, Airbus and SpaceX all use A.I. or machine learning in their products. There’s also a fledgling group of small vendors that specialize in A.I. application in satellites. Examples include Orbital Insight, Planet Labs, Hypergiant, Relativity Space and Sidus Space, just to name a few.
Top reasons why A.I is a good fit for satellites:
- Most AI/ML projects spend at least half of project resource in data cleanup. This has been consistently true for the last 15 years and is not expected to change anytime soon. Satellite ML (machine learning), however, is often machine-generated data, and possibly even new data, which can be better controlled and thus significantly reduce the problem.
- The general theory of A.I. commercialization is that big data owners will eventually win. Be it Google, Microsoft, Meta, Amazon or the U.S. government, these are the entities that own the vast majority of the data and are expected to exploit that advantage. But if satellites can generate new data, and the satellites’ owners have the rights to that data, it will be a nice place to be.
- The amount of data is quite large. Weather satellites collect more than 12 terabytes a day. Processing raw data into actionable answers is essential and ML can play a big role.
- With CPUs and GPUs getting faster, smaller and more energy efficient, AI/ML can be integrated in satellites in more situations, making the systems more responsive and intelligent. While the 200-millisecond round trip delay for a geosynchronous satellite is acceptable for some situations, real-time control often requires on-board processing.
- As discussed at the beginning, the cost of a satellite is high, and it is worth funding AI/ML projects for a variety of use cases.
Promising use cases for AI/ML in satellites:
- On-board processing: Embedding AI/ML in satellites to process data in real time, rather than waiting for it to be transmitted back to the ground, is making more and more sense. Experts in the field want to do even more on-board but need several magnitudes more compute power in a radiation‑hardened satellite CPU.
- Ground-based image analysis and classification: This is more of a classic AI/ML use case. But with tons of new images coming in hourly, the process to get answers from raw data needs all the help it can get.
- Satellite operations and constellation management: With more and more satellites orbiting Earth, keeping them out of each other’s way while optimizing large constellations is becoming intractable. The most important issue is avoiding collisions. Optimizing data collection and transmission and debris management are also becoming increasingly challenging.
- Satellite health monitoring and predictive maintenance: As more and more data gets collected about a given satellite, it can be better maintained, and failures could even be prevented with AI/ML. These techniques are similar to those used for aircraft maintenance for years.
- Enhanced cybersecurity: While security is often not discussed openly, it is becoming more of an issue for satellites. Using AI/ML to detect and prevent attacks is starting to be a standard approach.
John Roy is a managing director with Water Tower Research, a market and investment research firm. He was a former equity analyst at UBS Securities and Merrill Lynch.