Resonance-Anchored Residual Surrogate for Spectral Prediction and Inverse Design of Multiple Terahertz Metamaterial...

Efficient forward prediction and inverse design can bypass the complexity and computational cost of traditional full-wave simulations, accelerating the development of terahertz metamaterial absorbers. Most current machine-learning approaches for spectral prediction or inverse design are limited to absorbers with a single topology, requiring retraining whenever the topology changes. Simultaneous forward prediction and inverse design also require separate models that are trained independently. This study proposes a resonance-anchored residual surrogate framework for resonance-frequency prediction and goal-oriented inverse design across multiple absorber topologies. Unlike conventional spectral prediction, the proposed spectral representation consists of a resonance frequency and a continuous local absorption spectrum centered on the resonance peak. The residual network is trained using a physics-guided dynamic anchor loss to improve the accuracy of resonance-frequency prediction, local spectral reconstruction, peak alignment, and absorptivity estimation. On the test set, the model achieves a normalized resonance-frequency MSE of 6.64 × 10^-5, a normalized spectral-waveform MSE of 3.58 × 10^-3, and a physical-domain absorptivity RMSE of 0.0054. Ablation experiments verify the effectiveness of the residual connections and the dynamic anchor loss. The trained surrogate is then frozen and integrated into a projected gradient descent loop for inverse design. Full-wave simulations demonstrate that the optimized structures successfully reproduce the target resonance responses within the sampled design space.