Machine-learning-assisted phase transition detection of organic compounds based on optical fiber waist-enlarged fusion taper structures

This paper proposes an optical fiber evanescent wave sensor for phase transition detection of organic compounds, which was validated using n -octadecane. The sensor is constructed by arc-discharge splicing single-mode fiber (SMF) into a waist-enlarged fusion taper (WEFT) structure using a fiber fusion splicer. When two WEFTs are connected in series, they form a Mach–Zehnder interferometer (MZI). Since n -octadecane has different refractive indices in its solid and liquid states during the phase change, the change in refractive index causes variation in the interference dips in the spectrum, enabling the distinction between the solid and liquid states. However, traditional wavelength and intensity tracking methods require precise numerical analysis, limiting their practical applications. Therefore, we propose using machine learning to assist the WEFT structure in phase change detection. During the heating and cooling processes, the K-means algorithm is first applied to classify the solid and liquid states, corresponding to the two phases of the transition. Subsequently, a Gaussian mixture model (GMM) is used for optimization, allowing for accurate differentiation between the liquid and solid states of n -octadecane. The results show that during the heating and cooling processes, after training on the spectral data, the average silhouette coefficients were 0.8619 and 0.8813, respectively, and the log-likelihood values were −21.8062 and −1.175. The sensor we propose has a simple structure and is easy to manufacture. Combined with machine learning algorithms, it holds great potential for application in the field of phase change energy storage.