Researchers use machine learning to repeat and replicate fragrances

AI is changing how we take into consideration preserving scents, potentially rescuing rare fragrances which can be on the point of disappearing. 

Idelfonso Nogueira and his team on the Norwegian University of Science and Technology demonstrated that AI could craft formulas to recreate perfumes.

The study involves profiling fragrances by their “scent family” – descriptive terms reminiscent of “spicy” or “musk” – and their “odor value,” which gauges a smell’s intensity. 

For example, they found one fragrance with a high odor value for “coumarinic” (harking back to vanilla) and one other that was notably “alcoholic.” 

The researchers intend to make use of this process to preserve obscure and rare odors – like those obtained from changing natural environments or plants on the point of extinction – by replicating them from a single sample.

It may additionally create an efficient, replicable process for creating perfumes, a task that typically requires considerable trial-and-error. 

Methodology breakdown

The first step involves analyzing goal fragrances to grasp their scent profiles, including their scent families (like “spicy” or “musk”) and their intensity. 

Then, using a Gated Graph Neural Network (GGNN) trained on an unlimited database of molecules, the team generates recent molecules that might potentially replicate the goal fragrance. 

The study methodology. Source: Via ArXiv

This process involves two phases: training the GGNN with known molecules to learn the connection between molecular structures and their scents after which generating recent molecules that match the specified scent profile through transfer learning.

Nogueira explains that the perception of smell changes attributable to the physical and chemical interactions molecules undergo with air or skin. To replicate the unique scents accurately, they chose AI-generated molecules that evaporated similarly to those in the unique fragrances, addressing the challenge of capturing the ephemeral nature of “top notes” and the longevity of “base notes.”

After generating a variety of molecules, the team selects people who best match the goal fragrance’s scent profile based on their vapor pressure and fragrance notes. 

The final stage involves optimizing the perfume formulation to match the unique scent. This process considers the intensity of various scent families within the fragrance and adjusts the fragrance’s molecular composition accordingly.

One of the fragrances produced from this method closely replicated the unique, with minor deviations in its “coumarinic” and “sharp” notes. The other was almost a precise match. 

The authors suggest that expanding the database to incorporate more complex molecules could enhance the accuracy of those AI-generated fragrances, offering a less expensive and more sustainable solution for the perfume industry, which currently faces high costs and long development times.

Nogueira is poised to check the technology further by experiencing a few of the AI-generated fragrances firsthand in a colleague’s lab in Ljubljana, Slovenia. “I’m very excited to smell them,” he states.

AI has quite a few novel uses, reminiscent of when researchers used neural networks to geo-locate wine’s unique flavor profiles to light up the chemical basis of terroir.