A framework for predicting odor threshold values of perfumes by scientific machine learning and transfer learning

Area of Science:

• Computational chemistry
• Machine learning
• Sensory science

Background:

• Odor thresholds are vital for the perfume industry but difficult to measure.
• Empirical models can estimate odor thresholds from molecular structures.
• Scientific machine learning offers novel predictive strategies.

Purpose of the Study:

• To develop and evaluate a machine learning framework for predicting chemical odor thresholds.
• To leverage transfer learning for improved odor threshold prediction accuracy.
• To compare the proposed model against benchmark methods and existing correlations.

Main Methods:

• A transfer learning strategy combining graph convolutional networks (GCNs) and feedforward neural networks (FNNs).
• GCNs predicted semantic odor descriptors, with outputs serving as input for the FNN.
• The FNN estimated odor thresholds based on molecular structures.

Main Results:

• The transfer learning-based model significantly outperformed a benchmark model lacking transfer learning.
• The proposed method showed better predictive performance than a literature correlation and a dummy regressor.
• This indicates the efficacy of transfer learning in odor threshold prediction.

Conclusions:

• Transfer learning offers a powerful and accurate approach for predicting odor thresholds.
• The developed framework provides a valuable tool for the perfume and fragrance industries.
• This method facilitates the estimation of odor properties from molecular information.