Application of Deep Learning Methods and Gas Chromatography-Mass Spectra to Develop Predictive Models for Specialty Coffee Odor

Abstract. Olfaction, or the sense of smell, is an important sensory aspect for human beings to shape the world of flavor. The odor is composed of various volatile organic compounds (VOCs) that generally could be detected by gas chromatography-mass spectrometry (GC-MS). However, the relationship between odors and the associated VOCs is not easy to be interpreted. In recent years, deep learning (DL) approaches have been used to solve complicated prediction and data-mining problems in the field of odor analysis. This study aims to develop predictive models of five odor categories (floral, citrus, berry, fermented, and nutty) in specialty coffee by machine learning and deep learning methods. In this study, 362 specialty coffee samples were collected along with the GC-MS spectra and the cuppers‘ report to develop odor predictive models. Machine learning methods (e.g., support vector machine and random forest) and deep learning methods (e.g., convolutional neural networks, CNN) were applied for model development, and the performances were compared. The CNN-based model presented the best performance with an average F1 value of 0.666 and an average accuracy of 0.790. Model visualization was further implemented to present the features of five odor categories learned by the CNN model. Preliminary results showed that deep learning methods are promising in predicting five targeted odor categories using GC-MS spectra; further research will focus on conducting in-depth data analysis to interpret model features.