Episode 4: When Engineers Met a Master Perfumer
MIT Media Lab’s How to Generate (Almost) Anything series paired domain experts with generative models to test how creative disciplines could evolve alongside machines. In Episode 4, perfumer Christophe Laudamiel guided student engineers through the fundamentals of fragrance structure. Neural networks trained on open-source recipe data proposed ingredient lists, while participants evaluated the accords in a studio setting.
The experiment underscored how machine suggestions can provoke novel ideas without replacing trained noses. Laudamiel emphasised classical frameworks—including the top, heart, and base note pyramid—before interpreting the AI’s proposals. The outcome was a trio of wearable prototypes that reflected both algorithmic exploration and human curation.
Industrial Platforms: Philyra and the Data Advantage
Shortly after MIT’s showcase, Symrise and IBM Research introduced Philyra, a machine-learning model trained on more than 1.7 million anonymised formulas, ingredient costs, and market briefs. The system ranks combinations according to target demographics and regions, letting perfumers adjust accords with insight into price constraints and anticipated performance. Human evaluators still conduct organoleptic tests, adjust dosages, and ensure perfume note pyramid fundamentals are preserved while remaining compliant with IFRA guidelines.
When Brazilian retailer O Boticário launched two Philyra-assisted fragrances in 2019, Symrise highlighted that AI reduced formulation time from months to weeks while enabling competitive pricing. Crucially, the perfumers involved still authored the creative brief, selected the final palette, and shaped the storytelling.
Tools Perfumers Use Today
Industry houses now treat AI as a creative assistant embedded in their laboratories. Givaudan’s Carto pairs an ingredient-pouring robot with predictive analytics: perfumers sketch accords on a touchscreen, and Carto suggests dosages based on historical success metrics. Firmenich’s Scentmate™ extends that concept to entrepreneurs and indie brands, auto-generating briefs, mood boards, and ready-to-evaluate bases that comply with IFRA standards.
These platforms streamline repetitive tasks—balancing volatility curves, forecasting stability, or filtering for vegan ingredients—leaving perfumers freer to craft the overarching theme, story, and auteur touch that algorithms cannot replicate.
Personalisation and Retail Experiments
Beyond laboratory use, consumer-facing brands experiment with quizzes, skin chemistry sensors, and AI clustering models to recommend bespoke blends. Retail concepts in Shanghai, New York, and Dubai invite shoppers to adjust note intensity in real time while the algorithm keeps the formula IFRA-compliant. Some direct-to-consumer startups ship discovery kits, learn preferences through feedback loops, and generate follow-up scents post-purchase.
While these offerings rely on data science, they still require human oversight to evaluate ingredient sourcing, ensure preservation systems work, and translate analytics into poetic marketing narratives.
Ethics, Bias, and the Irreplaceable Human Nose
AI models inherit bias from the datasets they learn, which skews toward Western fine fragrance launches of the last four decades. Without human correction, algorithms risk reinforcing narrow aesthetics or overlooking culturally significant ingredients. Perfumers also remain the arbiters of safety, sustainability, and storytelling—areas where quantitative models still struggle.
Responsible teams pair data scientists with evaluators, review training sets for breadth, and set guardrails around IFRA limits, biodegradability, and ethical sourcing. The most successful collaborations treat AI as an exploratory partner while celebrating the intuition built through years of sensory practice.