The revolutionary potential of deploying artificial intelligence (AI) and deep learning on spacecraft is of increasing interest to the space sector. With proven usefulness in applications such as Earth observation, data processing and control tasks, artificial and spiking neural networks have become topics of intense research in the space community. AI generally requires dedicated hardware that is currently unsuitable to spacecraft, as on-board computing is limited by power budget, energy constraints and radiation concerns. Hence, AI is mostly used in ground systems and data must first be sent to Earth to be processed. This imposes design constraints on the spacecraft and a reliance on extensive ground station support. Processing the data at the source with AI can provide a solution to these issues. This creates a need for AI accelerators which can reliably and efficiently perform computations for AI. This PhD project has the goal of developing the first ever memristive hardware platform for AI acceleration in space applications. The core idea is to enable In-Memory Computing by using memristive devices to serve as the main computational component of the accelerator. Computing AI using conventional architectures incur significant costs both in terms of energy and latency due to the impact of data movement. In-Memory Computing solves this by computing inside the memory. Memristors are programmable resistors with memory, and are naturally suited to in-memory computing as they are able to serve as both the element of computation and of memory. Furthermore, they boast radiation hardness and higher energy efficiency. The research will identify a suitable application for on-board AI and then create suitable architecture using simulations. The eventual aim is the construction and characterization of a physical prototype. The addition of on-board AI using such a platform could transform the scope and possibilities of a variety of missions such as Sentinel and ExoMars.
