Machine Learning for Flavor Development

Abstract

Currently, most flavor development is performed by chemists who experimentally iterate many times to find flavors that best fit specified requirements. In this study, three potential algorithms for automating and accelerating this process are examined and implemented, and the results are analyzed. The apriori algorithm, which generates probabilistic association rules, fails because it only predicts the coexistence of high-frequency features. The generative adversarial network (GAN) fails because the generator only exploits a small part of the solution space. However, the variational autoencoder (VAE) successfully recovers purposely omitted features 95.6% of the time, making it the most promising algorithm.