Simulations and Experiments on the Radiation Resilience of Artificial Intelligence

The increasing reliance on Artificial Intelligence (AI) in space exploration necessitates a thorough understanding of its resilience under cosmic conditions. Cosmic radiation can degrade memory, corrupt stored software, and induce bit-flips that impair AI functionality. To address this, we developed a simulation framework that systematically injects bit-flips into different AI architectures to investigate vulnerability to radiation damage across model sizes and storage precisions (32- and 64-bit). The results were validated through gamma irradiation experiments using commercial off-the-shelf (COTS) devices employing SDRAM and EEPROM technology.
Our findings show that certain AI models exhibit remarkable robustness with extensive bit-flips required to significantly degrade the performance. Simple architectures with a large number of weights are particularly suited for radiation-rich environments. Experimental results further reveal a bias toward bit-flips from 1 to 0, but generally pinpoint a notable radiation hardness of COTS devices. 
This work advances the practical application of AI in space, medicine, and energy generation by establishing concrete model selection guidelines and an adaptable framework for predicting software response to radiation.