Machine learning-driven prediction and optimization of microcrystalline cellulose and cellulose nanocrystal production from Miscanthus biomass

The non-destructive and cost-effective prediction of microcrystalline cellulose (MCC) and cellulose nanocrystal (CNC) production performance is critical for optimizing Miscanthus -based biorefinery processes. This study employed a judicious sampling strategy initially by integrating PCA-based K-means clustering with centroid-proximity screening, systematically identifying 49 representative clones of M. sacchariflorus and 38 representative clones of M. lutarioriparius that are genetically diverse from a nationwide collection. The above clones were combined with 12 cultured polyploid M. lutarioriparius to form the sample set of the model. Analysis of variance (ANOVA) and summary statistics confirmed that polyploid M. lutarioriparius was the optimal biomass feedstock for MCC and CNC production, exhibiting superior yields and enhanced physical properties. Correlation and path analyses indicated that cellulose content, crystallinity index, and degree of polymerization positively promoted MCC/CNC yields and their physical properties, whereas other components negatively affected them. Leveraging the optimized sample sets, three machine learning algorithms were employed to develop predictive models for estimating the yields of MCC/CNC and their associated physical properties based on biomass quality traits. Among these, the extremely randomized tree regression (ETR) model exhibited remarkable robustness (R 2 train > 0.88) and superior predictive accuracy (R 2 test > 0.87). Based on MCC and CNC yield data, four superior germplasm resources were selected, with G07 exhibiting the highest performance (MCC yield: 39.55 %; CNC yield: 31.64 %). This study not only provides high-throughput tools for forecasting MCC and CNC production but also highlights chromosome doubling as an effective strategy for enhancing biomass quality, thereby advancing sustainable and competitive biorefinery systems. • Systematic sample set optimized via integrated PCA-K-means with centroid-proximity. • Polyploid M. lutarioriparius exhibits superior MCC and CNC production performance. • CC, CrI, DP boost MCC/CNC yields and their physical properties. • ETR showed high robustness (R 2 train >0.88) and accuracy (R 2 test >0.87). • SHAP elucidates the biomass-to-product mechanistic linkages of ETR.