On-demand inverse design of terahertz polarization-converting metasurfaces via physics-constrained extreme learning machines

Repeated inverse design of terahertz metasurfaces requires accurate, reusable predictions of complex scattering responses under physical constraints. We develop an angle-conditioned extreme learning machine (ELM) coupled with coordinate search for fixed-incidence geometry optimization. The ELM learns four complex scattering channels over 401 frequency samples from 7,776 CST-generated unit-cell configurations. Cross-channel equality is imposed by output parameterization, while radial spectral-norm scaling controls the adopted passivity bound. In the unseen incidence-angle test, mean MSE decreases from 0.0975 to 0.0169 relative to the data-driven ELM. Ablation identifies spectral scaling as the source of the accuracy gain; equality wiring supplies the structural constraint. With 2% and 5% labeled target-angle data, mean MSE is 1.2783× 10⁻³ and 5.6210× 10⁻⁴ , respectively. Independent full-wave validation on 50 off-grid targets gives mean complex MSEs of 9.6272× 10⁻⁴ for the proposed pipeline and 9.5264× 10⁻⁴ for the data-driven ELM plus genetic algorithm. Four new PCR-specific designs reuse the same trained model; three satisfy both polarization-conversion and reflected-power specifications at every prescribed band sample. The framework combines inexpensive repeated prediction and geometry search with reusable complex-response objectives and full-wave candidate verification.