Hybrid prediction framework and structural optimization for gap-compensated nested hollow-core anti-resonant fibers

Hollow-core anti-resonant fibers (HC-ARFs) offer low-loss broadband guidance, but the geometric complexity introduced by gap-compensated cladding tubes poses a challenge to accurate confinement loss (CL) prediction. We develop a cascaded hybrid framework that couples an artificial neural network (ANN) with extreme gradient boosting (XGBoost) for rapid broadband CL prediction. The ANN captures the global nonlinear dependence of CL on fiber geometry and wavelength, whereas XGBoost refines the initial predictions by learning log-space residuals from augmented features combining the original inputs and ANN hidden-layer activations. Compared with the standalone ANN, the resulting hybrid model reduces the mean absolute percentage error by 5.5%. When employed as an efficient surrogate for structural optimization, the model identifies a low-loss design within the investigated geometric parameter ranges, with CL below 0.2 dB/km over 1.1-1.45 μm. Independent finite-element calculations for the selected low-loss structure closely reproduce the predicted spectrum, with an average deviation of 0.0064 dB/km, indicating the broader potential of hybrid machine-learning surrogate models to accelerate design-space exploration and optimization for structurally complex HC-ARFs.