Symrise and IBM: New Fragrances Created by AI
Symrise AG, a leading German producer of fragrances and flavors, uses an artificial intelligence (AI) software developed with IBM to create new fragrances more efficiently.
Headquartered in Germany, Symrise AG is a leading producer of fragrances and flavors with € 3.2 billion in revenues in 2018 [1]. The company designs and produces aromatic ingredients for perfumes, personal care and household products, as well as food and beverages. Very often, these flavors determine if the consumer subconsciously likes or dislikes a product. Symrise’s additives are responsible for such a wide range of product characteristics—freshness in toothpaste, the special scent of perfumes, etc.—that the average consumer interacts with the company’s products up to 20–30 times a day [2].
The process of inventing new scents is the core of the company’s business model. To stay competitive, Symrise needs to be at the forefront of product innovation while frequently introducing new products to the market. This innovative capability is what differentiates the company from its competitors, especially in its fragrance business. However, crafting new scents combines art and science and is, as such, a process difficult to fully control and predict. Therefore, Symrise is looking to push the boundaries in its invention process. The company recently tested artificial intelligence (AI) software called Philyra developed in cooperation with IBM [3].
Philyra uses advanced machine learning algorithms to create new and innovative fragrances [4].
Even after years of experience and training, skilled perfumers are largely dependent on trial and error to find new promising combinations of ingredients. This could soon change. The software Philyra uses machine learning algorithms to scan through thousands of raw materials to identify novel combinations of fragrance raw materials and to autonomously design new formulas. The algorithm is trained with 1,300 fragrance raw materials (synthetics and natural extracts), historic formulas (Symrise contributed a database with 1.7 million tested combinations), and information as to which formulas were previously successful [5]. By adding information about specific target customer groups, preferences, and sales data, the software can even suggest a formulation that had not been tried before and is, for example, particularly likely to be successful among Brazilian millennials—which was the first product actually introduced for the cosmetics company O Boticário, the second-largest beauty chain in Brazil [3]. One of the two versions smells like “fenugreek seeds, green cardamom pods, carrot seed, all wrapped with a milky, buttery, rich base note” and the other “is a fruity, floral scent—seemingly aimed at girls—and has scents of Osmanthus tea with lychee and patchouli” [3]. Both have received extremely positive reviews.
It turns out that the algorithm is particularly helpful when it comes to combinations that seem unlikely to harmonize [5]. In contrast to human beings, the Philyra software is not biased due to the complete absence of any predisposition to the world of smells and fragrances. Only after Philyra comes up with a novelty, an experienced perfumer fine-tunes the product and emphasizes certain features to give the fragrance a unique character. This approach offers the advantage of letting the highly trained (and well-paid) perfumer concentrate on the art piece of the innovation process without spending time on the combinatory exercise.
However, this collaboration between man and machine also creates challenges. Every perfumer has a certain style, prefers certain raw materials with which they are experienced, and often wants to create a perfume that will be recognized as their composition—like a cook who has a personal note. The Philyra program requires the perfumer, who tends to be a strong personality, to radically change their methods, which could lead to some frustration and objection. Symrise has tried to mitigate these challenges by publicly emphasizing that the software will never replace perfumers but is supposed to support them and allow them to be even more creative [6]. This assurance also seems important in making sure that the association of fragrances with craftmanship, arts, and creativity does not vanish in the eyes of the consumer, as this could over time push down consumers’ willingness to pay.
It seems likely that solutions such as Philyra will become standard tools in the development process of new fragrances. As the competitive advantage of Symrise will depend more and more on data analysis and machine learning, the company will have to increasingly hire expertise and insource data capabilities. With its vast database of fragrances, the company is uniquely positioned to develop proprietary software that will help it to secure a competitive advantage. Collaboration between data scientists and perfumers seems only to be a continuation of the art and science that have produced fragrances from time immemorial. While the collaboration with IBM was used to gain publicity, future generations of AI software will be employed behind the curtain to protect the magic that surrounds our favorite perfumes.
[2] https://www.symrise.com/our-company/#introduction
[4] https://www.ibm.com/blogs/research/2018/10/ai-fragrances/
Well considering I wrote about Watson I feel the duty to comment on this post! I really like that you focused on the tension this raises between human and machine – from my own research, I felt that IBM goes to great lengths to incorporate human intuition into its AI. I wonder to what extent the perfumer’s intuition is being baked in as a “hybrid” approach or whether the AI runs more standalone according to certain metrics set by perfumer. The former approach seems to more in the spirit of collaboration, but the second is more what you would get “out-of-the-box” meaning faster to get up to speed and simpler from a process perspective to implement. Would be interesting to see how that machine – hybrid approach is being implemented and whether it has implications on other industries.
I think this is super interesting. Having previously worked directly with master perfumers, I can empathize with the push back that incorporating AI into the fragrance development process undermines the art form of perfumery and could diminish the credibility of the perfumers themselves. However, behind every perfumer, there is an entire lab of scientists working on the formulas – so already we have a blend of art + science. It is an extremely iterative process and a single perfumer will never have enough time to explore and discover every combination, so I see Philyra’s role as both assistant and new source for inspiration. I would also like to entertain the idea that the database can be used to identify combinations with raw materials that are more sustainable to impact the value chain. Fragrance is extremely subjective and I would love to see the technology pushed even further to “learn” how certain formulas react with various skin types. While I don’t have a solution to accomplish this yet, one of the toughest parts of selecting between modifications of a formula is evaluating how different individual skin chemistry responds to it. Thank you so much for sharing!
Interesting! I did not hear about Philyra before today, but I’ve spent the last hour trying to understand how AI can be used to create materials with new properties (fragrance, in this case).
It looks like Symrise and IBM use ‘distance’ between raw materials to create the most novel fragrances. Does novelty always translate to appeal in fragrances? I wonder how they can further couple this process with human perfumers to create a feedback loop that allows them to manufacture new combinations of fragrances without having to resort to distance as a measure.
This is a radical idea , but can Philyra integrate ‘fragrance sensors’ (like electronic noses – https://en.wikipedia.org/wiki/Electronic_nose) into the learning process so that this system can constantly generate new perfumes with only minimal human intervention from the perfumers?
Fragrance creation and AI, this is such an unexpected combination! I wonder at what frequence does the Philyra software come up with new proposal? Would it be around 10 or 100 a day? I would assume that starting from a certain amount, the frustration of the perfumer would be lowered since making a choice between a large number of combinations would give him/her a feeling of agency and the possibility to still express a personal note while benefiting from the AI’s intervention. Anyway, this is a great post, thanks for sharing!