Machine learning–driven structural optimization of zigzag silk fibroin ionic conductors for spatially resolved wearable motion detection

Abstract Flexible ionic conductors that maintain stable electrical performance under large deformation are critical for wearable sensing. However, their geometric design space couples topology, manufacturability, and deformability in a high–dimensional parameter space that trial–and–error approaches cannot navigate efficiently. Here, we present a data–driven structural design workflow for zigzag–patterned silk fibroin ionic conductors. Zigzag unit cells parameterized by four control points yield 3000 mirror–tiled variants whose biaxial unfoldability, quantified as the mean stretch ratio $${\lambda }_{{mean}}$$ λ m e a n , is labeled by quasi–static simulations. An XGBoost surrogate trained on 15 structural descriptors predicts $${\lambda }_{{mean}}$$ λ m e a n with $${R}^{2}$$ R 2 = 0.97. A continuous–action A2C reinforcement learning agent then uses the surrogate to search for high–unfoldability configurations under manufacturability constraints. Three optimized configurations at $${\lambda }_{{pred}}$$ λ p r e d = 1.27, 1.77, and 2.04 are fabricated as direct–write–printed silk fibroin ionic conductors with ten–channel resistance readout. Under cyclic joint flexion, the devices produce distinct spatial response patterns: the λ = 1.27 device shows a symmetric dual–hotspot distribution, the λ = 1.77 device concentrates response in one dominant channel (peak $$|\varDelta R/{R}_{0}|$$ ∣ Δ R / R 0 ∣ = 0.82), and the λ = 2.04 device achieves multi–hotspot coverage with the highest mean amplitude (0.27). Spatial fingerprints persist across bending speeds (Spearman ρ = 0.75–0.84), establishing a complete pathway from geometric parameterization through reinforcement learning optimization to experimental validation.