How AI helps perfumers explore new combinations

A perfume can begin with a suggestion from a machine and end with something a person chooses to wear. In June 2019, O Boticário launched Egeo On Me and Egeo On You in Brazil after working with Symrise and IBM’s Philyra fragrance system.[1]

The collaboration has since extended to other fragrances. Symrise lists the 2021 Kalos Tech launch for Belcorp’s Ésika brand and Tros AI Awakening in Thailand in 2022. Its current platform description reports an archive of 3.5 million formulas and information about more than 2,000 raw materials. These are the supplier’s figures.[1]

The AI explores a recipe space

IBM describes Philyra as a system that learns from formulas, ingredients and previous commercial results. A brief specifies the intended audience and desired character, and the software proposes candidate combinations.[2]

The team’s 2017 workshop paper describes generating candidates and scoring them against several objectives. Its novelty analysis compared combinations of scent notes in existing perfumes: a different recipe alone does not necessarily create a recognisably different smell.[3]

For O Boticário, a master perfumer adjusted the formulas to emphasise particular notes and improve their persistence on skin.[2]

A good scent still needs the right application

Symrise’s account of developing the system supplies a revealing early mistake: one candidate smelled good but had been formulated for a candle. That was the wrong answer to the brief. The intended product had to become part of the model’s instructions and evaluation. The same company account explains that ingredients can behave differently in soap, detergent and fine fragrance.[4]

This is why an extensive recipe archive has value beyond the number of entries it contains. A useful suggestion must connect an appealing smell with the material and product in which it will be used. Symrise describes its development process as repeated algorithm refinement, rather than a successful formula appearing from the first experiment.[4]

Berlin let the public help shape a fragrance

A separate 2019 project added feedback from people visiting the Bikini Berlin shopping centre. According to Symrise, 4,300 people participated between 20 May and 29 June; 1,100 supplied impressions of the scents. Their responses informed the blending process for Berlin 3.0, which the company presented at IBM’s Frankfurt Think Summit that October.[5]

That was a participatory design exercise. The reported figures describe involvement, rather than a controlled comparison proving that AI made a better perfume. Its interest is the working loop: people encountered an actual scent, their reactions became further information, and the formulation could develop around those responses.[5]

The brief can include the ingredients’ future

Philyra 2.0, announced in June 2022, added criteria for renewable and readily biodegradable ingredients. Symrise said perfumers could use it both for new creations and for revisiting established fragrances. The announcement describes capabilities and goals; it does not quantify an independently measured reduction in the environmental impact of a finished bottle.[6]

For a perfumer, that creates an additional creative problem worth solving: retain an attractive character while changing which materials provide it. A promising alternative has to satisfy several demands together. The contribution of AI is to make more of those alternatives available for a person to examine, mix and judge. The final object is a physical sample, ready to be assessed as its character develops on skin.[6]

Report an error or suggest a correction