Performance prediction of terahertz hollow-core anti-resonant fibers based on a sequential decision network.

Hollow-core anti-resonant fibers (HC-ARFs) serve as ideal waveguides for terahertz wave transmission. To address the low computational efficiency of the traditional finite element method (FEM) in the multi-parameter optimization of HC-ARFs, we propose what we believe to be a novel sequential decision network (SDNet) for predicting the performance of nested terahertz HC-ARFs within the frequency range of 0.5 to 1.5 THz. The model integrates the decision-making principles of the random forest algorithm with the advantages of deep learning neural networks. By utilizing the fiber structural parameters and operating frequency as inputs, SDNet efficiently predicts both confinement loss (CL) and dispersion, classifying each into six distinct levels. Experimental results demonstrate that SDNet achieves prediction accuracies of 94.3% for CL and 97.7% for dispersion, significantly outperforming traditional machine learning algorithms such as Random Forest, K-NN, and Decision Tree. Additionally, analysis of confusion matrices demonstrates SDNet's superior ability in distinguishing between easily confusable categories. This work overcomes the previous limitations of predicting CL only at the 1.55 μm wavelength by extending the analysis to a broader terahertz frequency range and incorporating dispersion classification predictions, thus providing an efficient and accurate approach for the multi-parameter collaborative optimization of HC-ARFs.