Dual-learning framework: Integrating deep and reinforcement learning for optimizing bio-inspired Helmholtz-porous acoustic metamaterials for railway noise mitigation

Urban railway noise pollution poses significant public health challenges, necessitating advanced mitigation solutions. This study presents a novel dual-learning framework to optimize bio-inspired Helmholtz-porous acoustic metamaterials (HPAM) for railway noise absorption between 500–2000 Hz. The framework uses a highly accurate surrogate model, ESKAN (Echo-state module with Kolmogorov-Arnold Networks), which achieved an R2 of 0.987. To find the best material designs, a Q-learning enhanced Covariance Matrix Adaptation Evolution Strategy (CMA-ES) optimization algorithm efficiently explores complex 14-dimensional parameter spaces. This method successfully identifies optimal configurations for various material thicknesses (20–120 mm) and types, including porous materials like melamine sponge and rockwool, and shell materials ranging from soft polymers to rigid metals. Experimental tests on 3D-printed ABS prototypes confirmed the framework’s accuracy. The fabricated samples demonstrated excellent broadband absorption (SAC > 0.8) across the 480–2000 Hz range. This research introduces a generalizable method for designing next-generation metamaterials, overcoming previous optimization limitations while respecting manufacturing constraints.