TLDR: This research paper explores the use of memristors, specifically Phase-Change Memory (PCM) and Resistive Random-Access Memory (RRAM), to accelerate Artificial Neural Networks (ANNs) for on-board guidance and control applications in space. Facing the energy and radiation constraints of small satellites, memristor-based in-memory computing offers a promising solution due to its low power consumption and radiation resilience. Through simulations of a Guidance and Control Neural Network (G&CNET), the study evaluates memristor performance under various non-idealities like noise, conductance drift, and device degradation. Key findings indicate that increasing ‘slices’ (devices per weight) improves accuracy, retraining can restore performance after degradation, and while memristors can learn expert actions, challenges with noise impact accuracy. The paper lays a foundation for future research into energy-efficient and radiation-hardened AI accelerators for space.
The realm of space exploration is increasingly looking towards Artificial Intelligence (AI) to enhance the autonomy and capabilities of uncrewed spacecraft. However, deploying AI in space, especially on small satellites and CubeSats, faces significant hurdles: limited energy budgets and the harsh radiation environment that can damage conventional electronic chips. This challenge necessitates the development of new AI accelerators that are both energy-efficient and radiation-tolerant, while still meeting the demanding computational needs of space applications.
A recent research paper, Guidance and Control Neural Network Acceleration using Memristors, explores a promising solution: using memristors for on-board, in-memory computing AI acceleration. Memristors are a type of electrical component whose resistance depends on the history of the current that has flowed through them, essentially acting as a ‘memory resistor’. This unique property allows them to integrate data storage and computation in the same physical location, drastically reducing the energy consumption associated with moving data between separate memory and processing units—a common bottleneck in traditional computer architectures.
The paper specifically investigates two types of memristors: Phase-Change Memory (PCM) and Resistive Random-Access Memory (RRAM). These devices are not only non-volatile (meaning they retain their state without power) and exhibit minimal energy leakage, but they also show resilience to radiation, making them ideal candidates for space applications. Their ability to perform multiply-accumulate operations, the core of deep neural networks, with high efficiency further underscores their potential.
To evaluate the performance of memristor-based accelerators, the researchers simulated a Guidance and Control Neural Network (G&CNET). This type of neural network is crucial for tasks like real-time optimal control and trajectory generation for spacecraft. The G&CNET studied was a fully-connected feedforward network designed to take system state variables as input and produce optimal control outputs, such as thrust actions for a low-thrust spacecraft.
The simulation setup, utilizing the IBM Analog Hardware Acceleration Kit, allowed for the exploration of various device parameters and non-idealities inherent to memristors, such as noise, conductance drift (where the programmed resistance value changes over time), and device degradation (faulty devices). Experiments were conducted to understand how factors like the number of ‘slices’ (multiple memristor devices used to represent a single weight in the network) impact accuracy, how device degradation affects performance, and whether retraining can restore functionality after degradation. The effect of conductance drift over time was also closely monitored.
The findings revealed several key insights. Increasing the number of slices per weight significantly improved the network’s accuracy, effectively averaging out noise across multiple devices, though with diminishing returns beyond a certain point. Conductance drift did degrade network performance over time, with PCM devices showing a more pronounced degradation than RRAM devices. Crucially, the study demonstrated that retraining the neural network after device degradation could largely restore its performance to nominal levels, even with a significant percentage of faulty devices.
While memristor-based accelerators show immense promise for space AI, challenges remain. The simulated analog networks, despite their potential for energy and latency reductions, exhibited noisier control outputs compared to their digital counterparts. This suggests a need for further research into architectural noise mitigation, output filtering, or adapting neural network requirements for these devices.
Also Read:
- Unlocking Hyperspectral Data: How AI Explainability Guides Band Selection for Better Analysis
- Enhancing Business Process Predictions with Integrated Temporal Logic
This research provides a foundational step for future work, highlighting the potential of memristor-based AI accelerators for space applications. Future avenues include detailed energy consumption comparisons with conventional solutions, longer training durations for improved accuracy, and exploring noise reduction techniques like low-pass filters or averaging results from multiple networks. The technology could also benefit other space applications requiring on-board learning, such as asteroid geodesy or adaptive compression for Synthetic Aperture Radar (SAR) imagery, where low-power inference and continuous in-field learning are critical.
