Enhancing Reliability Estimation for Spaceborne Computers: Methodology from the RAMSES CubeSat Mission

This study introduces a methodology for assessing the reliability of FPGA-based AI accelerators in CubeSats operating in deep space. The approach integrates mission-specific radiation modeling with hardware fault emulation. Using GEANT4 simulations on a detailed 3D model of the spacecraft, realistic Single Event Effect (SEE) rates are derived based on geometry and material shielding. These rates inform targeted fault injection campaigns on the FPGA to measure performance degradation and neural network resilience. Applied to the RAMSES CubeSat, the method shows that accounting for true geometry reduces predicted SEU rates by over 50× and enables accurate Mean Time To Failure estimation. The results demonstrate that coupling precise radiation modeling with hardware-level testing provides a practical and reliable framework for designing robust CubeSat computing systems.