At first this sounds simple, but in practice it is highly complex. “We have around two million formulas for fragrant mixtures in our system, and a large number of them have been sold to customers in the past,” says Christian Schepers, describing the scope of the task. The manager in Global Business Support in the Fragrance division at Symrise is in charge of the project. “This data is a true treasure if you know how to use it.”
The fragrance recipes play highly varied roles here. Some of them are “icons in the world of perfume,” as Schepers calls them, which have been sold unchanged again and again since they were first created. Others, however, are more short-lived. Just a couple hundreds of new fragrances are launched each year – from limited-edition Christmas perfumes to seasonal mixtures for detergents or shampoos – for which thousands of test samples must be prepared in order to find the right product. “Our initial approach was that digitalization can help us cover these ever-shorter product cycles and make development more efficient.”
Even computers have to learn
The process is based on algorithms that are initially fed by programmers, but which will increasingly be fed and improved by artificial intelligence. IBM was sent all of the recipes in encrypted form, meaning without the real names of the products or their raw materials. “The mixtures are our capital and we don’t want to just give it away,” explains Christian Schepers the reasoning behind this step. The algorithms were then programmed so that they could – very simply put – analyze the structure of a fragrance, the application areas, factor in various markets and sales figures, and ultimately design new fragrances based on this foundation. “For example, we put in a fragrance for a detergent in Latin America for the spring. At the push of a button, the system found between 50,000 to 70,000 formulas and filtered out the ones that have the highest potential for success within a second,” says Schepers. But most of the attempts didn’t run quite so smoothly. “The first tests were miles away from working. After that, we refined the algorithms more and more.” One fragrance, Schepers recalls, smelled good but was created for use in a scented candle. The type of application also has to be factored in, just like dozens of other parameters.
The first milestones
The first products to hit the market were two perfumes for the Brazilian manufacturer O Boticário, the world’s number three in the industry. “It was very important for us to attract such a big partner in order to bring more attention to this innovative approach,” says Schepers. With the help of Philyra, the project began in January 2018. The market-ready product was presented at the World Perfumery Congress in Nice in June 2018 and went on sale in Brazil in June 2019. The automated creation is a product of the future, not least in considering the topic of regulatory requirements. If an ingredient is no longer allowed in a part of the world for regulatory reasons, it will be swapped out in the recipe. Most of the time there aren’t 1:1 alternatives, however. “We often have to incorporate multiple ingredients in the formulas to reach the desired result. This is very time-consuming,” explains Schepers. “Artificial intelligence will become part of the perfumers’ toolbox to solve this,” says the manager. “With this tool, they can their use creativity, resourcefulness and inspiration even better than before.”