The Scent Graph
The Scent Graph is a directed multi-relational graph of olfactory entities. Nodes represent atomic notes, accords, fragrances, fragrance families, and user-side taste vectors. Edges represent compositional, perceptual, and behavioural relationships: which notes belong to which fragrance, which notes co-occur in which accords, which fragrances are similar by perceptual profile, which notes a given user has liked or disliked.
Every interaction inside the Fragnatique app updates the graph. A liked fragrance propagates affinity to its constituent notes and accords. A scanned bottle anchors a new context to a known fragrance node. A Scent Lab session adds a candidate formulation that the AI evaluates against the graph before producing a recipe.
The 680 atomic notes
We use 680 atomic notes as the smallest indivisible unit of olfactory representation. An atomic note is a named raw material or naturalistic accord that perfumers and buyers recognise as a single olfactory object: bergamot, iris, oud, sandalwood, galbanum, ambroxan, ISO E Super, white musk, vanillin, and so on.
Each atomic note carries a structured profile: olfactory family or families, position in the notes pyramid (top, heart, base, or distributed), characteristic descriptors (powdery, woody, animalic), volatility class, common usage range in formulation, and an indicative IFRA category where applicable.
The 3,000 fragrances
The 3,000 fragrances cover both designer and niche houses, with strong coverage of houses that historically test the depth of any AI perfume advisor: Creed, Tom Ford, Maison Francis Kurkdjian, Byredo, Diptyque, Penhaligon's, Mancera, Initio, Le Labo, Parfums de Marly, Xerjoff, Amouage, Andrea Maack, Hormone Paris, and many others. Every fragrance is decomposed into its notes pyramid, accord radar, vibe gauge, and metadata such as season, time-of-day, longevity, and sillage.
The 3,000-fragrance dataset is cached on-device so the offline scanner and the match score work without a network connection.
Eight olfactory families
Fragnatique uses an eight-family taxonomy as the top-level olfactory organisation: citrus, floral, woody, oriental and amber, fougère, chypre, gourmand, and aquatic. The taxonomy aligns with the dominant industry references (the Fragrance Wheel by Michael Edwards, the Société Française des Parfumeurs classification) while remaining stable across modern reformulation cycles.
Family membership is non-exclusive. A modern release can sit primarily in one family and inherit characteristics from another, which the graph represents with weighted edges to multiple family nodes.
The six-axis accord radar
Each fragrance is profiled on a six-axis accord radar: woody, fresh, sweet, earthy, fruity, and green. The radar is a perceptual fingerprint, not a chemical inventory. It compresses dozens of micro-descriptors into a shape that humans can compare quickly and the algorithm can index efficiently.
The accord radar is the input to the perceptual similarity edges in the graph. Two fragrances with the same family but different radars are presented as different choices; two fragrances with similar radars across different families are surfaced as cross-family matches in the discovery flow.
How the match score works
The match score compares a user's personal taste vector against the structured profile of every fragrance in the graph. The taste vector is built from three sources: the 12-question profile quiz, explicit likes and dislikes, and (optionally) photo style analysis from the AI vision layer.
The score is a percentage, but it is always paired with a written explanation that names the specific notes and accords that drove the decision. Reasoning matters more than the number: a 92 percent match without an explanation is less useful than an 86 percent match that says why.
Scent Lab AI: GNN + QSPR
The Scent Lab uses two model families together. A Graph Neural Network (GNN) reads the structural relationships in the Scent Graph and predicts harmony, novelty, and perceived character of a candidate formulation. A QSPR (Quantitative Structure-Property Relationship) model predicts physical properties like volatility, stability, and a 90-day aging trajectory from the molecular composition.
The two outputs are reconciled into a single recommendation surface: a numeric harmony and success score, a sensory and volatility profile, an indicative cost breakdown, and an editor pass that turns the underlying numbers into a plain-language recipe.
The IFRA-aware compliance layer
Every Scent Lab output is checked against the latest International Fragrance Association (IFRA) standards (currently Amendment 51, 2023). The compliance layer flags raw materials that exceed allowed concentration limits for the intended product category, surfaces the 26 EU-listed allergens that require declaration, and produces the safety section of the manufacturing PDF.
IFRA-aware does not replace a regulatory dossier or an independent safety assessment. It is a first-pass check that prevents an indie perfumer from sending an obviously non-compliant formula to a contract manufacturer. Final responsibility sits with the brand or formulator.
Validation and updates
The Scent Graph is validated continuously against three signals. First, internal consistency: notes pyramid, accord radar, and family assignment must agree across structurally similar fragrances. Second, perceptual ground truth: the algorithm outputs are checked against blind sniff panels and against published reviews from recognised noses. Third, user feedback inside the app: when a high-scoring match consistently fails to convert across many users, the graph is re-weighted.
Major updates cycle quarterly. Note-level corrections and new fragrance additions ship more frequently. The on-device cache is refreshed on each app update.
Sources and provenance
The Scent Graph is constructed from a combination of public manufacturer disclosures, the published industry references for olfactory family taxonomy, structured aggregation across fragrance databases, hand-curation by the Fragnatique team, and direct relationships with niche houses and retailers (including the retail partnership with Fragrance Gallery Tallinn).
Note-level descriptors and family assignments cite Michael Edwards' Fragrances of the World classification, the Société Française des Parfumeurs reference, and the Wikipedia Fragrance Wheel entry as cross-checks. IFRA standards are referenced from the official ifrafragrance.org documentation.
The Scent Graph itself is a proprietary asset of Appthos Studio OÜ. Press, partner, and academic inquiries about access or licensing can write to the support page or directly to contact@appthos.com.