Introducing Scentalytics AI Platform
Author
Milan PavlovicScentalytics AI Platform is a set of custom trained AI models which determine various characteristics of perfumes based on their notes. To demonstrate the capabilities of the platform, we developed an interactive web application, called AI Perfume Creator, which guides you through the virtual perfume creation process.
Introduction
Scentalytics AI Platform consists of 8 related neural networks, each designed to predict a certain aspect for a given perfume composition. Those aspects include perfume accords and type, then gender, seasonal and daytime suitability, as well as potential consumer sentiments and ratings for the desired combination of perfume notes.
When compared to the current state-of-the art AI models, known as large language models (LLMs), our neural networks are much smaller, easier to manage and more energy efficient. Since they are trained only on the specific data about perfume compositions, we believe they offer comparable performance with the more general large language models, but at much lower computational cost.
The whole platform currently consists of the following main models:
- Accords - calculates the proportions of the presence of different accords in the perfume composition
- Types - calculates the probability of composition belonging to different fragrance families
- Genders - predicts the most probable gender suitability for the given combination of notes
- Weather - determines if a composition is more suitable for warm, cold or moderate weather
- Daytime - determines if a perfume creation is more suitable for a day or night wear
- Sentiments - calculates scores for the five different consumer sentiment categories
- Rating - predicts the potential user rating value
AI Perfume Creator
This interactive web app demonstrates capabilities of our five main AI models. As you add or remove different perfume notes, the AI models in the background are constantly calculating properties of the current perfume composition, and the charts are being updated accordingly. Such an approach guides the user to create a desired and personalized perfume creation with the help of AI, but still leaves the main creative process to human.
Once you’re done with your creation, you can check the real existing perfumes which are similar to your virtual composition. To find them, we don’t use pure keyword search, but rather the semantics of notes, accords and fragrance family of your creation. This is the reason why the results won’t always contain perfumes with exactly the same notes as yours, but they should be similar to your composition in terms of combined semantics of the main aspects of your fragrant creation.
This perfume recommendation engine is a part of our advanced scent profiling service, and is powered by one of our neural networks. To find similar perfume creations, it searches the database of little less than 40 thousand known perfumes. This database will be updated over time to contain even more existing products.
The last, but not least feature is an Example formula. This highly experimental feature relies on the one of the newest and most powerful large language models out there. It enables perfumers and fragrance enthusiasts to see the suggested example formulations of chemical compounds for the given composition of perfume notes.
Current limitations
Our models are trained on publicly available online data about perfume notes, compositions and other properties. Currently, they can’t get more refined input as detailed perfume formulations and they may not always produce the expected results. However, if given the right data, we can develop tailored AI solutions for specific needs.
Future plans
In the near future, we plan to develop a model which can suggest a perfume composition based on input characteristics which are currently shown as outputs on charts in our AI Perfume Creator app. Such a model would represent more AI-centric approach to the virtual perfume creation when compared to the current human-centric style.
We also plan to improve our types model, by adding more fragrance families which a perfume can belong to. We are intensively working on the next upgrade of our platform, to include models which can associate a perfume composition with personalities, emotions, colors and even sounds.
Conclusion
We hope that our work will help and inspire perfume creators in their search for new olfactory experiences. If you have any questions, suggestions or collaboration requests, feel free to drop us a message to hello@scentalytics.com. We will be more than happy to help you.