In the world of fragrance, where AI-enabled features like Osmo’s Odor Map or Givaudan’s Myrissi tool are ushering in a new era of speed and efficiency in perfume product development, deep neural networks (DNNs), which are a specific type of AI technology, hold plenty of potential.

DNNs have more than three layers including output and input layers, where each layer trains on the pervious layer’s outputs, thus meaning the deeper the layer one goes, the more advanced and complex features can be recognized. 

The “Automatic scent creation by cheminformatics method” research paper by Manuel Aleixandre and Dani Prasetyawan, led by Professor Takamichi Nakamoto from the Laboratory for Future Interdisciplinary Research of Science and Technology (FIRST) in Japan, published in Scientific Reports, explores the potential of the medium for scent production.

The researchers trained the DNN to automatically create an odor profile with the addition of one scent descriptor. In order to predict odor descriptors and generate its own compositions, the model analyzed the mass spectrometry data of 180 essential oils, with 94 essential oils that had odor descriptor data for generating fragrance profiles. There were 39 descriptors in total, like “sweet,” “resinous,” and “balsamic.” With the development of an odor reproduction technique, the model was able to generate scents with varying ratios of odor component mixes. Random essential oil mixtures were added to improve accuracy and generalization. Sensory evaluations from human participants were used to make sure the perceptions aligned.

Once all this data had been collected and processed, the model was able to generate its own compositions, which were then assessed by human evaluators. The DNN scored the highest accuracy in predicting the odor descriptor “floral” and lower for the descriptor “woody.” The total balanced accuracy was 0.736 (out of 1).

DNNs offer superior reading of a large number of datasets, pattern identification, and prediction precision, reducing time and cost in the process for the manufacturers. “As DNN models continue to evolve, they could enable the creation of personalized fragrances tailored to individual preferences. Additionally, this approach could be extended to other sensory domains, such as taste, where similar methods could be used to craft personalized flavor profiles,” Nakamoto states.