Data-Driven Color Recipe Recommendation System for Reactive Dyeing of Cotton and Blended Knit Fabrics

Abstract

Reactive dye recipe development for cotton and blended knit fabrics depends on historical shade-matching, laboratory trials, and CMC-based color evaluation in industry. This thesis presents a data-driven recipe recommendation framework built from industrial shade-matching records. The dataset includes target shade information, fabric and fiber composition, dye formulations, and instrumental color measurements. The framework used historical similarity retrieval, dye-name prediction using supervised learning, dye-percentage regression and $\Delta E$-based quality prediction. The historical baseline performed better for candidate retrieval. It achieved 0.7821 component dye accuracy, 0.6154 exact sequence accuracy, and 0.7692 top-3 dye-set accuracy. For top-1 prediction, the Gradient Boosting recipe outperforms baseline, with 0.8269 component accuracy, 0.6538 exact sequence accuracy, and 0.6538 top-1 dye-set accuracy. Its dye-percentage model achieved 0.2750 total recipe percentage MAE. The Extra Trees quality model predicted maximum $\Delta E$ with 0.1492 MAE, and 72.44% of predictions were within $\pm$0.20 $\Delta E$. The results show that a hybrid framework beats single model: historical retrieval improves candidate coverage, supervised learning improves top-1 recipe prediction, and $\Delta E$ estimation supports risk-based ranking.

Keywords: Reactive dye recipe; Shade matching; Machine learning; $\Delta E$ prediction; Cotton-blended knit fabric