Some fragrance ingredients are clearly perceptible even in small amounts, while others smell gentle. Can this “odor strength” be predicted from molecular structure before anyone actually smells the substance?
In May 2026, Peter Fichtelmann and Julia Westermayr at Leipzig University published a study addressing this question. They organized existing odor descriptions for 2,393 molecules and trained AI to learn the relationship between molecular structure and strength. In a test dividing strength into four categories, 72% of predictions matched the original classification. However, performance varied across categories, and validation on a separate dataset depended on concentration conditions.
Dividing odor strength into four categories
How can odor strength be expressed numerically in the first place? Even for the same substance, perception changes with dilution and the person smelling it. Large datasets measuring all fragrance ingredients under identical conditions are also insufficiently available.
The researchers therefore collected descriptions from the fragrance information website The Good Scents Company and the chemical database PubChem. After addressing errors in structural information, duplicates, and other issues, they organized the data into four categories: odorless, low, medium, and high. Rather than newly smelling and comparing every substance, the study prepared accumulated records for machine learning.
These four categories have an order. Predicting “medium” for a substance originally recorded as “high,” for example, is a smaller error than predicting “odorless.” The study incorporated this distance between categories into its evaluation. The main metric used to select models weighted the errors in each category equally so that performance was not determined only by the categories with the most data.
Although this makes strength numerically tractable, it does not imply ratios such as “high smells three times as strong as low.” It is a way of treating perceived strength as a rough ordering.
Representing molecules with numbers
The input given to AI concerns the atoms making up a molecule and how they are connected. This information is converted into numerical measures of molecular weight, structural features, and properties related to affinity for water or oils. Such numerical representations of molecular features are called molecular descriptors.
The study compared several representations and learning methods, including approaches that use molecular substructures as markers. It then examined in detail a method that averages the answers of three predictive models using molecular descriptors.
Thus, “predicting from molecular shape” does not mean looking only at a molecule’s visual outline. It means finding relationships between various kinds of information derived from atomic connectivity and previously recorded strength. The AI did not itself smell the ingredients or directly measure nasal receptor responses.
How accurate were the predictions?
The researchers divided about 80% of the collected data into a training set and 20% into a test set. They also grouped structurally very similar molecules on the same side of the split to avoid inflating apparent performance by including close analogues in both sets.
In that test, predictions combining three models matched the original classification in 72% of cases. Most mistakes were between neighboring categories, such as low versus medium or medium versus high.
However, the best-predicted category was medium, which had the most data. Low, which had fewer data, was also the most difficult to predict. The 72% figure does not mean that all strength categories were predicted equally well.
The researchers also compared predictions with data from another study in which people actually smelled and rated substances. Even after excluding molecules too structurally similar to those used for training, predicted categories corresponded to human ratings at lower concentrations. Performance declined, however, at the relatively concentrated condition of a tenfold dilution.
The model does not include concentration as an input. This reveals a limitation: it cannot directly calculate how strength changes as the same molecule is progressively diluted.
Molecular features used by the AI
What clues did the model use? Analysis of contributions to prediction highlighted molecular weight, shape, and features related to polarity. Polarity is an uneven distribution of electrical charge within a molecule and is related to properties such as affinity for water.
For an odor to be perceived, molecules must enter the air, reach the nose, and pass through a mucus layer to reach olfactory receptors. Molecular size and properties affect this journey. The features emphasized by the model were consistent with knowledge of how odor molecules reach their targets.
However, this is an analysis of the clues used for prediction. It does not establish a simple rule that a particular shape must smell strong. Multiple features contribute jointly, and no clear tendency was found for particular properties to be especially important at particular odor-strength levels.
Conclusion
AI trained on past odor records was able to estimate the broad strength of fragrance ingredients from molecular structure. In lower-concentration smelling tests from another study, molecules predicted to be stronger also tended to receive higher intensity ratings. The work shows that structure-based predictions can offer clues to human perception.
However, because concentration is not an input, the model cannot tell us how much to dilute a substance to achieve a target strength, nor does it predict the intensity of a blended perfume. Its potential is as an aid alongside smelling tests: using predictions to narrow candidates when searching for new ingredients, then having people smell them at the concentrations actually intended for use.
Original paper
Peter Fichtelmann and Julia Westermayr
Machine learning for smell: ordinal odor strength prediction of molecular perfumery components
RSC Advances, 2026, 16, 25696–25704. First published May 14, 2026.
DOI: https://doi.org/10.1039/d6ra01805j
Full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC13182554/
Original paper license: CC BY 3.0
https://creativecommons.org/licenses/by/3.0/
