Optimizing machine learning approaches for multimode interference fiber sensors: classification versus regression in wide-range refractive index detection

Abstract The optimization of machine learning (ML) approaches for multimode interference (MMI) fiber sensors plays a critical role in enhancing wide-range refractive index (RI) detection for applications in biomedical diagnostics, industrial monitoring, and environmental assessments. ML-based models improve sensing accuracy and resolve RI ambiguities by effectively interpreting complex spectral responses. However, choosing between classification and regression models presents a challenge, particularly when balancing discrete RI categorization with the need for continuous, high-precision measurements. This study systematically evaluates Decision Tree (DT), Support Vector Machine (SVM), and Neural Network (NN) models under both classification and regression frameworks to determine their effectiveness in ML-enhanced MMI fiber sensing. The findings reveal that classification models efficiently segment RI ranges, while regression models achieve superior predictive accuracy and continuity, with improvements exceeding 90% (NN: 99.27%, DT: 99.05%, SVM: 95.47%). The results underscore the advantages of regression-based ML approaches for uninterrupted and precise RI measurements, providing valuable insights for optimizing ML methodologies in next-generation fiber optic sensing applications.