Machine learning in the citrus essential oil industry: A comprehensive review on process optimization, authentication, and flavor and bioactive compound development

Background

Citrus essential oils (rich in flavor and bioactive compounds) are extracted from citrus peels (constituting processing byproducts accounting for nearly 40–50 % of fruit mass), enabling comprehensive citrus utilization while mitigating environmental contamination. The escalating complexity of extraction process parameters, analytical techniques, and compound research and development generates large-scale, complex datasets, necessitating applications of advanced computational analytics (machine learning, ML) for efficient data processing and analysis.

Scope and approach

This review systematically presented the general workflows for construction of ML model (encompassing data collection, pre-processing, model selection, and model evaluation), with a focus on their applications (including process optimization, yield prediction, traceability, adulteration authentication, key flavor compound characterization and identification, and prediction and synthesis of flavor and bioactive compound). Subsequent sections critically addressed predominant challenges and limitations, culminating in proposed future outlooks.

Key findings and conclusions

ML has permeated all stages of the essential oil industry, with its applications progressively transitioning from traditional machine learning algorithms (relying on manual feature extraction) to deep learning algorithms (employing neural networks for automatic feature learning) since about 2020. This technological evolution has significantly enhanced production efficiency, quality control, flavor regulation, and bioactive compound development. To further accelerate ML implementation in this sector, future efforts should prioritize: establishment of comprehensive public databases; development of domain-specific data pre-processing methods; identification of key features for quality control; advancement of hybrid and multi-task modeling; necessity of computational simulation and biological verification; enhancement of interpretability and commercial applicability.