Refractive index measurement based on multimode fiber specklegram analysis using a deep neural network

We present a refractive index (RI) measurement method based on specklegram analysis using a multi-taper multimode fiber (MMF) and a deep neural network. A bi-conical MMF with multiple tapers was fabricated using arc discharge to enhance sensitivity to variations in the surrounding RI. Specklegrams were recorded during the controlled dripping of sodium chloride solutions with RIs ranging from 1.336 82 to 1.344 65. After preprocessing and data augmentation, a ResNet-34 model was trained to establish a quantitative mapping between specklegrams and RIs. The trained model achieved a root-mean-square error (RMSE) of 8.9 × 10−5 in discrete RI prediction. In the continuous RI tracking, the temporal evolution of the predicted RIs closely followed the ideal mixing curve, yielding a RMSE of 7.3 × 10−5. These results demonstrate that deep-learning-based specklegram analysis offers a feasible, accurate, and cost-effective solution for RI sensing.