Multi-modal zero-shot prediction of color trajectories in food drying

Color change during food drying is an important indicator of both process dynamics and product quality. Although prior studies have investigated color changes, there is a lack of methods capable of accurately and efficiently predicting color change trajectories, i.e., the continuous evolution of color over the course of drying. Further, existing approaches are typically limited to discrete post-drying measurements and low-dimensional color representations, resulting in limited predictive capability and poor generalization across varying process conditions. This paper develops a multi-modal data-driven method for zero-shot prediction of color trajectories under unseen drying conditions. Specifically, each color trajectory is characterized as a weighted sum of predefined basis functions, thereby reducing the prediction problem to estimating a compact set of coefficients. We develop a multi-modal modeling approach to automatically extract informative features from both process parameters and initial sample images, which are then fused to predict these coefficients. To further improve generalizability, we introduce a similarity-informed training data selection strategy that filters the training set to include only conditions most relevant to the target evaluation scenario. Collectively, these elements distinguish the proposed framework from existing drying-quality prediction approaches by enabling full-trajectory prediction under unseen process conditions. The proposed method is validated on two distinct case studies, namely, cookie drying and apple drying. Under a zero-shot learning setting where training and evaluation conditions are non-overlapping, the model achieves root mean square errors (RMSEs) of 2.12 for cookie drying and 1.29 for apple drying, reducing errors by over 90% compared with a baseline long short-term memory (LSTM) model. Ablation results show that multi-modal fusion and similarity-informed training each independently improve prediction accuracy, with the greatest gains when combined. Overall, these results demonstrate the potential of data-driven trajectory prediction to support real-time quality monitoring and decision-making in food drying, and motivate future validation across broader food products, drying systems, and quality attributes.