Fast prediction of optimal reaction conditions and dyeing effects of natural dyes on silk fabrics by lightweight integrated learning (XGBoost) models

Abstract

There is a lot of repetitive work involved in exploring the dyeing performance of natural dyes. To improve the experimental efficiency, save material, reduce time costs and shorten the research cycle, this study collects and analyses the literature data of 350 natural dye experiments to construct the Natural Dyes Dataset, and achieves rapid prediction of the optimal reaction conditions and dyeing effects of natural dyes using a lightweight integrated learning model. The size of the trained XGBoost model is only 562 KB; only the name of the dye and its approximate chemical composition need to be input to predict the results of the reaction environment pH, colour fastness to washing (CFW) and colour fastness to rubbing (CFR) of the natural dye on silk fabrics with the highest K/S in a very short time of 52 ms. The prediction accuracies for pH, CFW and CFR in the validation set are as high as 94.12%, 93.75% and 100%, respectively, and 77.78%, 91.67% and 83.33% for the real test set, with both validity and transferability. The integrated learning approach provides valuable guidance for exploring the dyeing performance of natural dyes with very small deployment costs and a very short inference time, expanding the possibilities of cross‐application of the disciplines of machine learning and textile dyeing.