Probabilistic reliability modeling and soft-error tolerance for neuromorphic spiking neural networks in high-flux radiation environments

Neuromorphic processors enable ultra-low-power edge inference for safety-critical aerospace Artificial Intelligence applications, but ionizing radiation can induce soft errors that degrade spiking neural networks. This paper develops a probabilistic reliability model for spiking neural network inference under single-event upsets in synaptic memory and neuronal state, using a Poisson fault-injection framework that maps particle flux to bit-flip rates. We propose the Hierarchical Temporal Defense, a bio-inspired architecture combining probabilistic input encoding, homeostatic threshold adaptation, and volatility-gated plasticity to suppress error propagation in temporal codes under high linear energy transfer conditions. We also provide a certification-oriented methodology that derives reliability curves and mean time to failure from measurable flux conditions and an explicit functional failure threshold. Hardware-in-the-loop experiments on an Akida neuromorphic processor, using the publicly available Cislunar Anomaly and Risk Dataset, show that baseline spiking neural networks degrade to an anomaly-detection score based on the harmonic mean of precision and recall below 0.48 under severe solar energetic particle event rates. In contrast, the Hierarchical Temporal Defense maintains this score above 0.72 and increases mean time to failure by a factor of 4.5. These results demonstrate a viable pathway for deploying highly efficient, fault-tolerant Artificial Intelligence in extreme radiation environments.