From Wet Chemistry to Portable NIR: Robust High‐Throughput Compositional Models for Biomass Pretreatment

ABSTRACT This study demonstrates a robust chemometric framework for high‐throughput compositional analysis of lignocellulosic biomass using a portable near‐infrared (NIR) spectrometer. Sugarcane bagasse solids produced by pressurized solvent fractionation (PSF) under batch and continuous conditions were analyzed to predict cellulose, hemicellulose, and lignin contents. To enhance calibration representativeness, a matrix‐preserving densification strategy was introduced, in which blends of precharacterized samples—rather than synthetic mixtures—were generated, combined with controlled variations in moisture (4–25 wt% d.b.). Spectra (908–1676 nm) were preprocessed using Savitzky–Golay derivatives, standard normal variate, mean centering, and external parameter orthogonalization (EPO) to mitigate scattering and moisture effects. Partial least squares (PLS) regression models were built and validated through SPXY and random subset approaches. Optimized single‐constituent models exhibited high predictive accuracy for lignin (RMSEP 3.2–3.5 wt% d.b., R p 2 ≈ 0.96), cellulose (3.46–4.1 wt% d.b., R p 2 ≈ 0.94), and hemicellulose (1.1–1.5 wt% d.b., R p 2 ≈ 0.98), with moderate model complexity (four to six latent variables). EPO significantly reduced moisture‐related variance, improving calibration robustness and lowering latent‐variable requirements for lignin and hemicellulose. Conversely, multivariate PLS models with multiple dependent variables required higher dimensionality and yielded inferior performance, confirming the superiority of constituent‐specific calibrations.