A chemically guided Vision Transformer for interpretable near-infrared spectral analysis of multicomponent textile blends

Abstract

Accurate quantification of multicomponent textile blends by near-infrared (NIR) spectroscopy is hindered by overlapping absorption bands, baseline variation, and nonlinear interactions between fiber components. To address these challenges, we propose a chemically guided Vision Transformer (CG-ViT) that incorporates domain-specific chemical information—such as characteristic absorption peaks and functional group signatures—into the attention mechanism to improve predictive performance and ensure chemically grounded interpretability. The architecture combines a convolutional patch embedding to preserve local spectral continuity, a chemical-prior mask to emphasize wavelength regions of known relevance, and a gating module to adaptively control the influence of these priors for each sample. The approach was evaluated on 533 cotton–polyester–spandex textile samples, each measured at five independent positions, yielding 2665 spectra in total. CG-ViT achieved an overall RMSE of 0.7675 and R² of 0.999 on the test set, with component-wise RMSEs of 0.9443 for cotton, 0.8618 for polyester, and 0.3644 for spandex, all with R² ⩾ 0.998. These results exceeded the performance of established baselines, including PLSR, CNN, ANN, and standard 1D-ViT models. Ablation experiments verified the contribution of each module, and attention analysis demonstrated clear correspondence between model focus and chemically meaningful spectral features. Robustness assessments under cross-validation, Gaussian noise, and baseline drift conditions showed stable accuracy (R² ⩾ 0.97), indicating strong resilience to typical instrumental variations and supporting the method’s application to reliable, real-time, nondestructive textile composition analysis.