“Sciences provide an understanding of a universal experience, Arts are a universal understanding of a personal experience” –Mae Jemison

In our previous post we explored various tools (perfume/fragrance wheels) for helping perfume professionals and consumers identify categories of odor. Although these are great tools, we must acknowledge that they are the output of individuals with unique interpretations on fragrance classification. Another way to categorize fragrance, perhaps in a more global or universal context, lies under the lenses of science, chemistry, statistics and machine learning.

In this post we explore how researchers categorize the hugely complex world of odor molecules.

A Scientific Approach to Fragrance

There is over a century of published scientific research on odor involving a multitude of fields including chemistry, physics, neuroscience, psychology, biology, and machine learning, with applications from understanding human perception of smell to the creation of biomimetic artificial noses. We will focus exclusively on odor/fragrance classification.

Some of the earliest research with reference to odor classification dates to the mid 1800’s. This tended to focus on human perception of smell. More recent research explores a wide array of topics, but one of the most relevant to classification is a field of research that explores Quantitative Structure-Odor Relationship (QSOR) models, which attempts to identify the relationships between molecular structure and odor perception. 

In order for researchers to identify a relationship between the chemical structure of an odor and what a person perceives the smell as, they need data. Specifically, they need people (called subjects or a panel) to smell a large number of different odors and describe how they interpret the smell. 

A Modern Scent Perception Dataset

In 2015 such a dataset was provided by Andreas Keller & Leslie B. Vosshall in their paper “Olfactory Perception of Chemically Diverse Molecules”. In this paper they describe how they tested the perception of 481 different molecules in 61 untrained subjects (meaning they were not fragrance experts). They asked each subject questions about how familiar, intense, and pleasant each smelled in addition to asking how they would rate 20 different descriptors. 

Scent descriptors ordered from most to least pleasant:

• Sweet
• Flower
• Edible
• Fruit
• Bakery
• Warm
• Spices
• Grass
• Cold
• Wood
• Garlic
• Fish
• Burnt
• Acid
• Chemical
• Ammonia/Urinous
• Sweaty
• Sour
• Musky
• Decayed

DREAM & Google’s Brain Team Tackle Scent

This dataset was subsequently used in a project called the DREAM Olfactory Prediction Challenge, where challengers competed to create models that could predict smell descriptors based on molecular structure. In 2018, Google Research’s Brain Team in collaboration with Arizona State University, the University of Toronto, and others created a state-of-the-art model that outperformed the winner of the DREAM challenge. In their paper “Machine Learning for Scent: Learning Generalizable Perceptual Representations of Small Molecules”, they created a new dataset of over 5,000 molecules labeled by fragrance professionals with 138 odor descriptors.

One of the products of this research shows how often certain scent decriptors are used per molecule. The top five most commonly used descriptors (from most to least common) were: fruity, green, sweet, floral, and woody.

Another relevant finding was revealed in how certain words used to describe fragrance are clustered. They identified the following clusters of odor descriptors:

• Dairy - Yogurt, Milk, Cheese
• Fruity - Apple, Pear, Pineapple
• Bakery - Toasted, Nutty, Cocoa, Popcorn, Coffee
• Clean - Pine, Lemon, Mint
• Alcohol - Rum, Cognac, Malt
• Spice - Cinnamon
• Savory - Onion, Beef 

The odor co-occurance matrix with major scent families annotated:

The analysis shows common-sense and natural groupings of fragrance families. In addition to this they created an accurate mapping of odor space, showing how different scent families are inter-related. Below is a 2-dimensional representation of how their model interprets the odor space. A illustrates how odors that are not usually associated do not overlap. B shows groupings of several different odor categories.

Conclusion

The scent research that we’ve reviewed here is concerned primarily with the connection between molecular structure and perception of smell. It shows how challenging it is to identify relationships between what something smells like and what its chemical structure looks like. It also shows that the utility in this understanding doesn’t translate perfectly to how professional perfumers map different scent families. 

Like many things in life, we can look to the science for insight, but it often doesn’t contain a complete picture. The process of creating fragrance is something that always has and always will require both science and art. As researchers continue to untangle the complexity of scent we will continue to expand our true understanding of what fragrance is and how we experience it.

With every published paper on olfactory research, comes with it the potential for future innovation in perfumery and fragrance engineering. At WAVE we are always looking to the latest in research to invest in the future of fragrance. If this is something that interests you, please let us know! We're hiring!