Comparing Scientific Machine Learning and Data-Driven Neural Networks for Postharvest Quality Prediction with a Tomato Case

Fruits and vegetables are highly perishable, resulting in significant food loss and waste in the supply chains. Postharvest quality prediction of fruits and vegetables is challenging and traditionally addressed through two main steams of modeling approaches: mechanistic models and data-driven models. Mechanistic models are explainable and data-efficient but lack flexibility for capturing complex patterns within the data. Data-driven models, especially deep learning, can capture complex data patterns but require extensive data with low interpretability. To address these limitations, Scientific Machine Learning (SciML) has emerged as a hybrid approach that embeds domain knowledge to neural networks to improve the learning efficiency. This research makes a twofold contribution to the existing literature. First, It refines the framework introduced by Faroughi et al. (2024) by broadening its scope and standardizing the definitions of SciML models. This is achieved by adopting the typology of knowledge-guided, knowledge-encoded, and knowledge-informed neural networks as replacements for the classifications of physics-guided, physics-encoded, and physics-informed neural networks. Second, it applies the knowledge-informed model to a practical use case: predicting the overall visual quality of tomatoes post-harvest and comparing the results with those of purely data-driven neural networks. The findings demonstrate SciML’s potential to improve predictive accuracy by effectively addressing data challenges commonly encountered in traditional deep learning methods.