Hard optical fiber imaging components as two-dimensional array imaging materials comprised of tens of millions to billions of micro- and nano-scale optical glass fibers have extensive applications in numerous optical systems, including low-light-level image intensifiers and space target detection systems. Resolution and contrast are crucial indicators affecting the performance of these components. Elemental diffusion at the core-clad interface results in a gradient distribution of material refractive index rather than the ideal abrupt distribution, thereby degrading the resolution and contrast of the components. It is necessary for the demands for refined manufacturing processes of these components to gain a comprehensive understanding of the diffusion behavior and distribution status of elements at the core-clad interface. At present, energy dispersive spectrometry (EDS) is often employed as a characterization tool for elemental distribution status, yet it suffers from a low image resolution (20 nm) and lacks corresponding quantitative calculation methods. In this study, high-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) was utilized to acquire high-precision images of elemental spatial distribution. Furthermore, a multi-model fusion machine learning approach was proposed to address the challenges of characterization and quantitative analysis in the study of elemental migration at the core-clad interface of hard optical fiber imaging components.

In this study, monofilaments undergoing various thermal processes were fabricated by a drawing technique. High-Angle Annular Dark-Field (HAADF) images were captured and utilized to train Tree-SVM/LDA models. These models enabled pixel-level core-clad classification and x–y two-dimensional spatial classification, thereby yielding quantitative results on elemental interface diffusion. In addition, the reliability of the models was also analyzed.

The Tree-SVM/LDA model is grounded on the following assumptions, i.e., element Si as a key component in the formation of both core and cladding glass network structures remains stable during the diffusion process. The deviation distances of other elements are benchmarked against that of Si, the average deviation (D) is defined as an arithmetic mean of the migration distances of all elements within the tested range, serving to characterize the overall diffusion behavior of elements rather than individual extreme values, elements undergo unidirectional migration, consistent with the direction of the concentration gradient, manifesting as the perpendicular direction to the interface in monofilament samples, the transition zone at the core-cladding interface is formed through the superposition of diffusion behaviors of different elements, and the algorithmic procedure of model encompasses the following steps (i.e., constructing training and prediction datasets; training the decision tree model; developing a labeled prediction dataset; and obtaining classification hyperplane parameters).

The results obtained by SVM and LDA models both indicate the thermodynamic principles at higher resolutions, with the deviation D_K∈[10, 40]nm, D_Ba∈[0, 11]nm, and D_Ti∈[0, 13]nm. At lower resolutions, where the width of a single pixel in the raw data is 3 nm, the computational error of the models is amplified threefold, leading to the results that deviate from actual principles. At 810 ℃, the SVM model exhibits a poor data separability due to the low content of Ti in the cladding region, resulting in abnormal calculation results. This phenomenon underscores the superiority of the LDA model over SVM in terms of training set diversity, demonstrating a greater generalization capability. The calculation results of the SVM model and LDA model show a good consistency, while also presenting regular differences, i.e., the boundaries obtained by the LDA model tend to skew towards the cladding region relative to those of the SVM model. This regularity may be associated with differences in model characteristics and the element distribution patterns in the HAADF images of the core and cladding regions, having a further in-depth research. In general, the SVM model achieves higher accuracy than the LDA model. However, the LDA model possesses a better generalization capability.

The migration capabilities of the elements K, Ti, and Ba intensified as the wire drawing temperature increased. Specifically, the average migration distance of element K could range from 10 nm to 40 nm, while that of elements Ti and Ba could span from 0 to 13 nm. Affecting the reliability of models were since the spatial resolution of the raw data could be in atomic or sub-atomic scale, models could be sensitive to data anomalies, thus requiring repeated training, elements with great diffusion capabilities could decrease the accuracy of the model, while simultaneously enhancing its generalization capability. The SVM model demonstrated superior accuracy, compared to the LDA model, having a weaker generalization capability. The SVM model could be recommended at a high quality of the raw data, whereas the results of the LDA model could be more realistic at a low quality of the raw data.