Generalizable full-spectrum learning for recalibration-free refractometry with reflection-mode tilted fiber Bragg gratings

A tilted fiber Bragg grating (TFBG) is a powerful refractometric sensor, but its complex multimode spectrum demands rigorous demodulation and specialist expertise. Our convolutional learning pipeline converts these spectra directly into surrounding refractive index (SRI) estimates, making sophisticated TFBG sensors accessible to a much broader range of users. We evaluated an established preprocessing and convolutional learning pipeline on independently fabricated reflection-mode TFBG sensors at tilt angles of \(7^\circ \) and \(8^\circ \) , without calibrating or adapting the model to either independent test sensor. Four model-development sensors at each angle were used to establish the pipeline, and one independently fabricated test sensor at each angle was reserved for the final independent test. The pipeline used deterministic mode equalization, with its amplitude scale fixed from the training sensors, and five multiscale models based on convolutional neural networks (CNNs) in each training fold, using five random seeds. The results demonstrate spectral prediction across independent sensors and between the two measured tilt angles using a pipeline established entirely from the training sensors. Four additional sensors at \(6^\circ \) and \(9^\circ \) , angles absent from training, were also predicted without recalibration. These results provide a reproducible foundation for broader validation across additional fabrication batches of reflection-mode TFBG sensors.