Engineering an interpretable chemometric pipeline for sugarcane wine fermentation using synthetic spectra and explainable ensembles

Background: Real-time spectroscopic monitoring of sugarcane wine is becoming increasingly important for quality and safety control. However, it remains less studied than established fermentations such as beer or grape wine, especially from a chemometric perspective. Objective: To develop a mechanistically grounded synthetic spectral generation framework that enables interpretable machine learning for fermentation monitoring without extensive experimental calibration datasets. Methods: We integrated kinetic fermentation modelling based on extended Monod-inhibition equations with realistic spectral simulation incorporating multi-scale noise artefacts. Nine machine learning architectures (PLS, Random Forest, Gradient Boosting, DNN, CNN, LSTM, ResNet, Transformer, and Stacked Ensemble) were evaluated using SHAP-based explainability analysis and Bayesian bootstrap uncertainty quantification. Results: Tree-based models achieved exceptional performance on purely synthetic validation data (R² up to 0.997, RMSE ≈ 1.1 g/L), but all architectures collapsed when evaluated under simulated real-world conditions that introduced unmodelled matrix variability and instrument artefacts (R² from −0.01 to −1.88). The simplest PLS model degraded the least but still failed to reach acceptable predictive accuracy, indicating a fundamental gap between the synthetic training distribution and realistic deployment scenarios. Significance: These results show that even careful mechanistic, noise-aware synthetic spectra cannot guarantee successful domain transfer by themselves. Synthetic data remain valuable for model prototyping, architecture screening, and interpretability analysis, but must be complemented by targeted experimental calibration and synthetic-to-real adaptation strategies. Exploiting this limitation is critical for chemometric practice, as it reframes synthetic pipelines from “replacements” to “augmentation tools” for real fermentation monitoring.