Use of machine learning to efficiently predict the confinement loss in anti-resonant hollow-core fiber

The fundamental mode confinement loss (CL) of anti-resonant hollow-core fiber (ARF) is efficiently predicted by a classification task of machine learning. The structure-parameter vector is utilized to define the sample space of ARFs. The CL of labeled samples at 1550 nm is numerically calculated via the finite element method (FEM). The magnitude of CL is obtained by a classification task via a decision tree and k-nearest neighbors algorithms with the training and test sets generated by 290700 and 32300 labeled samples. The test accuracy, confusion matrices, and the receiver operating characteristic curves have shown that our proposed method is effective for predicting the magnitude of CL with a short computation runtime compared to FEM simulation. The feasibility of predicting other performance parameters by the extension of our method, as well as its ability to generalize outside the tested sample space, is also discussed. It is likely that the proposed sample definition and the use of a classification approach can be adopted for design application beyond efficient prediction of ARF CL and inspire artificial intelligence and data-driven-based research of photonic structures.