A data-driven framework for modeling, interpretation, and multi-objective surrogate optimization of dairy spray drying using heterogeneous literature data
Food and Bioproducts Processing · 2026-12 · 160 · 631-648
DOI: 10.1016/j.fbp.2026.10.022 ↗Kurz gefasst
Fachartikel in der Zeitschrift Food and Bioproducts Processing, 2026-12. Verfasst von Mehri Karbasi und 1 weiteren Autor:innen. Die Arbeit nennt 79 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Institute of Food and Agriculture.
Abstract (Originalsprache)
Simultaneous control of dairy powder quality during spray drying is challenging because feed composition, processing conditions, and powder properties interact nonlinearly. This study developed a unified machine learning framework integrating prediction, interpretation, and multi-objective optimization using heterogeneous literature-derived datasets comprising 110–178 observations per target from 35 studies. Four regression models were evaluated for yield, final moisture content, median particle size, and residual thermal ratio using nested leave-one-out cross-validation, with imputation and hyperparameter tuning performed within each fold. XGBoost achieved the highest R² for yield (0.679), moisture (0.781), and thermal ratio (0.921), whereas random forest performed best for particle size (0.799); differences from competing models were within uncertainty except for thermal ratio. Leave-one-study-out validation revealed limited cross-study transferability: R² decreased to 0.523 for moisture, 0.307 for particle size, and 0.223 for thermal ratio, while yield predictions did not generalize across studies. Out-of-sample model interpretation identified physically plausible process–quality relationships and discrepancies from training-set attributions. Multi-objective optimization generated 200 Pareto-optimal candidate conditions; incorporating applicability-domain and particle-size constraints increased predicted specification compliance from 6.5% to 100%. However, broad jackknife+ prediction intervals, whose coverage falls below nominal for unseen studies, and the absence of independent experimental validation preclude conclusions about operational reliability. The principal contribution is a methodological framework for synthesizing dispersed literature data to explore process trade-offs and guide experimental design. The framework identifies candidate operating regions that require independent experimental validation before operational deployment.
Förderung
- National Institute of Food and Agriculture
Zitieren
APA
Karbasi, M., & Xu, C. (2026). A data-driven framework for modeling, interpretation, and multi-objective surrogate optimization of dairy spray drying using heterogeneous literature data. Food and Bioproducts Processing, 160, 631-648. https://doi.org/10.1016/j.fbp.2026.10.022
Vancouver
Karbasi M, Xu C. A data-driven framework for modeling, interpretation, and multi-objective surrogate optimization of dairy spray drying using heterogeneous literature data. Food and Bioproducts Processing. 2026;160:631-648. doi:10.1016/j.fbp.2026.10.022
BibTeX
@article{Karbasi2026,
author = {Karbasi, Mehri and Xu, Changmou},
title = {A data-driven framework for modeling, interpretation, and multi-objective surrogate optimization of dairy spray drying using heterogeneous literature data},
journal = {Food and Bioproducts Processing},
year = {2026},
volume = {160},
pages = {631--648},
doi = {10.1016/j.fbp.2026.10.022}
}
EUVIMED