Tuesday, February 4, 2025

Fragrance Generators



Text generators and image generators and even video generators have been hitting the streets and hitting our screens for a while now, but let's not forget the molecule generators, maybe we could call them chemical generators and make it real confusing (they're generating possible chemicals, like the formula for a chemical not yet known by science, not actual chemicals; that's for a different robot). 

Perfume engineering uses trial-and-error to find new fragrant chemicals. That's how we do everything before we know how to do it. It's very inefficient. So now we're trying to use machine learning to take at least some of the guessing out of all this. And it works, sort of. 

Machine learning gives us new molecules to work with, but it doesn't tell us how to combine those molecules with others. It can't predict the perceived intensity of the just-discovered but not-yet-created molecule.

Molecules yes, perception no.

Perfumes have no copyright protections; they are protected by the inability for people to guess the composition. You can get the molecules right, but not the amounts relative to each other; changing the ratio of even two molecules from 10:1 to 10:3 is enough to mess up the overall effect.

So this effort is to predict the overall effect of a bunch of molecules mixed toegther, not just one but a bunch together. They train a neural net using molecules and words. I can't really tell which word-dataset they're using, because it seems to be proprietary, and based on Teixeira et al's 2014 Perfumery Radar 2.0 (https://sci-hub.se/10.1021/ie403968w).


Using AI to replicate odors and validating them via experimental quantification of perfume perception
Mar 2024, phys.org

via Norwegian University of Science and Technology: Bruno C. L. Rodrigues et al, Molecule Generation and Optimization for Efficient Fragrance Creation, arXiv (2024). DOI: 10.48550/arxiv.2402.12134



I'm skeptical because the list is so short, but it might be as simple as this: citrus, fruity, green, floral, herbaceous, musk, oriental, and woody; and based on the rationale that this small group represents 75% of the odor space (Teixeira 2010). 

There are less common descriptors such as leather, gourmand, aldehydic, balsamic, and herbal which are used only twice. One dimensional descriptors are tobacco, modern chypre, floral oriental, soft oriental, mossy woods, dry woods, and mint, among others. Some molecules didn't come with their own notes, so Good Scents was used as a reference.

Here's a list of the lexica for odor descriptions mentioned in the Perfumery Radar text - Calkin and Jellinek, Jaubert, Roudnitska, Edwards' Fragrance Wheel, Zarzo and Stanton, Boelens-Haring and Thiboud. Apart from these, each fragrance company or perfumer has their own that they've developed over the years. The Perfumery Radar 2.0 itself uses a base layer with eight olfactory families, and two additional layers: an outer layer with seven descriptors and an inner layer with 17 descriptors. 

And here's the info from their table on the most used descriptors by fragrance companies: floral, woody, citrus, fruity, green, oriental, chypre, aromatic, fouger̀e, musk, spicy, ambery, marine; used in different ways by Givaudan, Osmoz, International Flavors & Fragrances, Symrise, Frutarom, MANE, Societ́é Franca̧ise des Parfumeurs (SFP), The Fragrance Foundation, Avon, Fragrantica, LaLoff.

They talk about the difference between using common words for olfactory perception and the more limited set of words used by expert perfumers, which is an important part of constructing these lexica, and also that each fragrance house maps the olfactory space in its own way. 

via Chemical Engineering Department of the Norwegian University of Science and
Technology, Laboratories of Separation and Reaction Engineering and of Catalysis
and Materials and of Chemical Engineering at University of Porto, and SIA Murins Startups in Latvia:  BC Rodriguex et al. Molecule Generation and Optimization for Efficient Fragrance Creation.  

No comments:

Post a Comment