Bidirectional BPNN-based color prediction framework for wide-gamut color-spun yarns: Enhancing digital control in sustainable textile manufacturing

This study addresses the challenge of bidirectional color prediction in wide-gamut color-spun yarns through the development and validation of robust Backpropagation Neural Network (BPNN) models. Unlike traditional physical color models that struggle with the complex optical interactions in fibrous structures, our approach leverages neural networks' capacity to model non-linear relationships directly from experimental data. We designed a comprehensive grid-based color mixing framework using four strategically selected primary fibers (Grey, Cyan, Magenta, and Yellow), from which 66 distinct yarn samples were produced and characterized. Two specialized BPNN architectures were developed: one predicting blend ratios from spectral reflectance data, and another forecasting resultant CIELAB values from known blend ratios. The models were trained, validated, and tested using industry-standard metrics, achieving high accuracy with mean squared error values consistently below 0.03 for ratio prediction and color differences (ΔE) generally around 2.0 for color prediction. Practical validation involved producing yarns based on BPNN-predicted recipes, which demonstrated close visual and spectrophotometric agreement with target colors. Beyond addressing the limitations of conventional color theories, this bidirectional prediction framework offers significant practical implications for textile manufacturing, enabling both precise recipe formulation from target colors and accurate color forecasting from known recipes. This research advances textile coloration toward more sustainable, digitally-integrated production processes with color reproduction capabilities approaching modern digital color standards.