Reconstruction of sea surface temperature is critical for marine monitoring, yet conventional edge devices based on complementary metal-oxide-semiconductor (CMOS) technology suffer from memory-wall bottlenecks and radiation vulnerability in harsh marine environments. Here, we propose a neuromorphic computing framework based on radiation-tolerant synthetic antiferromagnetic (SAF) synaptic devices through physical-algorithmic co-design to achieve robust sea surface temperature reconstruction. The fabricated Ta/Ir/Fe0.65Tb0.35/Ru/Co/Pt/Ta-based SAF devices enable field-free magnetization switching via spin-orbit torque, exhibiting multilevel conductance states that naturally emulate synaptic and neuronal functions. Notably, these devices retain over 92% of their performance after 1 Mrad (Si) γ-irradiation, demonstrating inherent radiation tolerance arising from strong antiferromagnetic exchange coupling. By mapping the nonlinear conductance response of SAF onto the cross-attention mechanism of a Perceiver IO architecture, we achieve accurate reconstruction of sea surface temperature fields from sparse sensor inputs. On the National Oceanic and Atmospheric Administration dataset, our system attains a root-mean-square error below 2°C-competitive with deep learning baselines-while projections indicate a potential reduction in energy consumption by an order of magnitude. This work not only advances the application of neuromorphic computing in marine science, but also provides a promising pathway toward "environmentally adaptive intelligent computing".
🌊
