When satellites learn to think AI will be ubiquitous in space

Related Vendors

AMD Advanced Micro Devices GmbH

Wuxi InfiMotion Technology Co., Ltd.

Artificial intelligence will be ubiquitous in space and significantly expand the capabilities of modern spacecraft. Advanced sensors and the analysis of massive amounts of data increasingly require faster and more autonomous decision-making processes. AI in orbit can reduce dependence on limited downlink bandwidths by making decisions directly on site.

Ken O'Neill is Space Systems Architect at AMD.

Modern spacecraft are typically equipped with more powerful systems. Examples of this are remote sensing satellites, which take photos or videos with ever higher resolution and faster frame rate or capture a larger number of multispectral and hyperspectral imaging channels. While the development of sensor applications has kept pace with the growing data demand, this is generally not the case with the bandwidth for data downlink. The larger the data sets, the longer the exchange with the ground control stations inevitably takes.

This results in a major challenge: If the decisions resulting from analysing the data need to be made faster and sometimes even in real time, time becomes a critical factor.

To address these issues with insufficient data downlink bandwidths, a large part of the data processing is shifted to systems directly on-site, i.e., in space. However, this requires much more demanding solutions in terms of energy consumption and payload weight, an aspect that must be carefully considered when developing satellites/spacecraft.

Furthermore, both emerging and established players in the space industry are working to reduce the costs of providing hardware for space. This is also evidence of the trend towards a democratisation of space travel and research. Low-cost, low-power solutions (such as SmallSats and CubeSats) can provide an affordable platform for scientific investigations and the demonstration of technology using constellations, swarms and disaggregated systems. This marks another trend: the requirements for semiconductor technologies optimised for space applications are being redefined.

AI capabilities on board

The integration of AI processing units into space-worthy hardware using chips that are robust enough for space use can provide a solution. The key here is enabling compute-efficient and power-saving local inference to identify relevant sensor data and thus relieve the downlink. The ability to make autonomous decisions in orbit supports missions; in some cases, it is even mission-critical. For example, AI is already used in earth observation satellites to detect the presence of clouds in the images captured. If surface details are obscured by clouds, the image can become unusable. In this case, it can be discarded and does not consume either storage space or downlink bandwidth.

When it comes to security applications, where objects on the Earth's surface need to be identified in real time, AI algorithms can also provide important support. For example, they can quickly distinguish commercial ships from warships in object recognition, in order to shorten reaction times and avoid lengthy analysis cycles by humans.

For spacecraft that are supposed to land on planets or asteroids, remote direct control of the landing process from Earth is simply not possible due to communication delays. Thanks to the onboard AI, the vehicle can autonomously identify suitable landing sites in real time.

There is also increasing interest in using AI to monitor the overall condition of systems onboard satellites/spacecraft. In this case, AI is used to detect anomalies in measured parameters (e.g., currents, voltages, temperature levels, mechanical stresses, vibrations, fluid flows and pressures, and magnetic and electric field strengths).

This provides faster error detection without human intervention, which otherwise could take days or weeks. As a complex modern satellite can have several thousand telemetry channels, AI allows for the real-time analysis of all channels, while only a subset of the telemetry channels needs to be available on the ground for human analysis.

Living in space

With the increasing prevalence of AI in space, the industry needs cost-effective solutions for hosting inference workloads. There are various ways to implement AI inference procedures in embedded systems. A common approach is the use of dedicated DSP resources, which are often integrated into microelectronic hardware such as FPGAs, GPUs, TPUs or specialized ASICs. Components like the adaptive AMD Versal AI Core SoCs with their integrated AI Engines (AIEs) are designed to implement the multiplication and accumulation operations required by neural networks much more efficiently.

However, the challenges of preparing systems for use in space will never become obsolete. Spacecraft systems are expensive and cannot be repaired once launched. Therefore, ensuring quality and reliability is crucial.

Since radiation in space is significantly higher than here on Earth, microelectronics components can be exposed to sudden destructive radiation effects (one-time latch-ups, etc.) as well as a gradual deterioration in performance due to the effects of the total ionizing dose.

Example AMD

The Class B qualification and manufacturing test flow is based on the MIL-PRF-38535 Class B specification for the qualification and testing of monolithic ICs. This qualification has been adapted to the advanced organic packaging required for space-grade products like the adaptive AMD XQR Versal AI Core SoCs, and complements the large amount of quality and reliability information already collected about these components under extreme temperature conditions. The components are also tested for radiation effects by exposing them to protons, heavy ions, and gamma radiation in various tests. This ensures the functionality of the spacecraft systems that use these components, and allows the organizations that use these devices to reprogram the hardware after deployment and perform necessary over-the-air (OTA) updates.

Lastly, longevity is crucial. Satellite manufacturers sometimes need support for their products years after launch. By this time, many commercial microelectronic components are already unavailable and obsolete. AMD meets these requirements with a team of specialized quality and reliability engineers, with adaptive SoCs that are tested and characterized for radiation effects. And the production and support of space-grade components are maintained for a long period after the launch.

Diverse Systems

In the last ten years, the demand for commercial image data from space has grown significantly, both for passive (visible, infrared and ultraviolet light) and active imaging (mainly Synthetic Aperture Radar). Most new satellites providing commercial image data fly in Low Earth Orbit (LEO) to minimize the distance between the satellite and the object, thereby ensuring high image quality. By deploying many satellites, satellite operators can increase the frequency with which a satellite passes over an area of interest. More frequent passes can be beneficial for recipients of commercial images, such as organizations monitoring activities in ports, mines, or agricultural areas. The satellites used for such applications often belong to the MiniSat (200 kg to 600 kg) or SmallSat (600 kg to 1200 kg) categories.

In such satellites, AI inference can be very useful for optimizing the use of storage resources and downlink bandwidth, and making autonomous decisions. However, since these compact satellites require small form-factor electronic systems, the AI inference must be performed in devices with minimum space requirements. Leading companies supporting AI in semiconductors optimized for space applications are now able to offer large, fully equipped devices alongside those with fewer resources but much smaller package dimensions. In this way, designers can make the optimal device selection for the specific constraints of their system.

Acceleration of Missions

While the capabilities of satellites or their sensors have dramatically increased, the downlink bandwidth has not grown as quickly. Artificial intelligence is a viable way to reduce the need for limited bandwidth while enabling much faster decision-making – sometimes in real-time – using the captured data. It can be efficiently implemented in SoCs that provide special adaptive AIEs. (mbf)