Data-Driven Spectral Prediction of Black Dyeing in Recycled Polymer Microfibers via Multi-Output Regression

This study presents a spectrum-based data-driven modeling framework for predicting the wavelength-dependent K/S spectrum from dye recipes in the dyeing process of black dope-dyed recycled microfiber fabrics. Instead of relying only on scalar color coordinates, the proposed framework directly predicts the K/S spectrum over the 400-700 nm range. This enables a more direct interpretation of color reproducibility while preserving the physical basis of color formation. Using samples dyed with combinations of yellow, red, and blue dyes, the problem was formulated as a multi-output regression task in which the K/S value at each wavelength was treated as an output. Because black-shade data are characterized by high inter-wavelength correlation and limited spectral variation, eight regression models were systematically compared under these constrained conditions, and partial least-squares regression (PLSR) was selected as the final model. The selected model showed high agreement between the predicted and measured K/S spectra. When the predicted spectra were converted into CIELAB coordinates, the mean color difference between predicted and measured colors was 0.79 in terms of $\Delta E_{ab}^{*}$ , indicating a high level of color reproducibility. These results show that the proposed framework can effectively model black-shade dyeing data with a limited number of samples while maintaining both spectral agreement and perceptual accuracy. The framework provides practical support for data-driven dye recipe design and quality validation and may contribute to improved color control, process efficiency, and reproducibility in polymer-based microfiber dyeing processes.