Electrical‐Programming‐Free Neural Crossbars Enabled by Photochemical Weight Patterning

Abstract

ABSTRACT Hardware neural networks based on crossbar architectures require reliable implementation of multilevel synaptic weights, yet conventional electrical programming often suffers from programming interference, device variability, and write–verify overhead. Here, we present electrical‐programming‐free neural crossbars enabled by photochemical weight patterning using photoresponsive PEDOT:PSS devices. Ultraviolet/ozone (UVO) and ultraviolet‐only (UV) irradiation provide fully optical bidirectional conductance modulation, allowing programming and erasing of synaptic weights without electrical assistance. The flexible crossbar array exhibits eight well‐resolved conductance states with low device‐to‐device variation (4.7%), stable retention, repeatable programming/erasing endurance, and mechanical robustness over 2000 bending cycles. Through mask‐assisted parallel photochemical weight patterning, trained weights are directly transferred into a 25 × 4 flexible hardware neural network for fingerprint identification, achieving a hardware accuracy of 86.67%, comparable to the software model accuracy of 89.0%. The platform further supports optical reset and reprogramming for reconfigurable user registration while enabling complete removal of the PEDOT:PSS active layer through water dissolution for substrate recovery. These results establish photochemical weight patterning as a scalable strategy for electrical‐programming‐free, flexible, and reconfigurable neuromorphic hardware.