The integration of artificial intelligence into space systems faces fundamental challenges in radiation resilience and energy-efficient computation. Here, we present a neuromorphic computing approach that proposes a strategy to transform these challenges into computational advantages via spintronic technology. We developed spin–orbit torque magnetic tunnel junction crossbar array harnesses radiation-induced fluctuations to enhance Hopfield neural network optimization, converting an environmental constraint into a functional benefit. When exposed to heavy ions (e.g., 209Bi23+), the system demonstrates remarkable radiation hardness with only 1.04% tunneling magnetoresistance degradation while maintaining operational stability. Implemented in a 4 Kb array, this neuro-inspired architecture solves the eight-city traveling salesman problem with 95.2% accuracy at 45.06 nJ energy consumption—outperforming conventional radiation-hardened approaches. Such a complementary experimental and simulation approach elaborates that the measured irradiation conductance fluctuations can be mapped to synaptic weights in a Hopfield network model, significantly enhancing its optimization capability. This work corroborates an emerging paradigm for adaptive, energy-efficient nanoscale artificial intelligence hardware that is designed to thrive in extreme environments, with implications for radiation-resilient neuromorphic architectures and edge computing.
Zhang et al. (2026) studied this question.