Industries / Food, farming and forestry / Flavors and fragrances
Creating a flavor or scent
Creating a flavor or scent is choosing which of thousands of materials go into a formula, and in what amounts, so that the mixture smells or tastes as intended in the product that will carry it.
6
Milestones achieved, from 1882 to 2019.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1882
Perfume built on a synthetic
Paul Parquet created Fougère Royale for Houbigant around an accord, a small blend that reads as one smell, of coumarin, oakmoss, geranium and bergamot. The coumarin was made in the lab. A 2001 review by Givaudan chemists says that with its launch "the new era commenced" for perfumes using synthetic ingredients.
Credit
Paul Parquet; Houbigant
1889
Perfume built in notes
Guerlain's Jicky used lab-made vanillin as well as coumarin. Simon Cotton, a chemist at the University of Birmingham, writes that it is often thought of as the first modern perfume built in three "notes": a top note of the lightest molecules that dominates for the first few minutes, a middle note that lasts for hours, and a base of heavy molecules that remains longest and slows the others' evaporation. Jicky is still sold.
Credit
Guerlain
1960
Meat flavor from heated sugars
Charles Gerard May and Ian Douglas Morton of Lever Brothers patented a way to make a meat flavor by heating cysteine, an amino acid, with sugars; adding ribose made it taste more like beef. This uses the Maillard reaction, the reaction between sugars and amino acids that browns cooked food. A 2000 review in ScienceAsia calls Morton's 1960 British patent the first to claim that heating cysteine with ribose, glucose or xylose gives a meat-like flavor, and many later reaction-flavor patents followed it.
Credit
C. G. May and I. D. Morton, Lever Brothers
1975
Scent sampled from living flowers
Roman Kaiser, a chemist at Givaudan, worked mainly from 1975 on analyzing and rebuilding natural scents. His headspace method encloses a living flower in a glass vessel for up to two hours and catches its scent on an adsorbent trap for analysis, without picking it. Swissinfo reported in 2006 that he and colleagues had analyzed more than 1,900 scents from about 9,000 plant species; Givaudan's field program, ScentTrek, won a FiFi technology award in 1996.
Lessona picked flower and a living one differ in chemistry, so what is sampled decides what is rebuilt.
swissinfo.chimg.perfumerflavorist.comgivaudan.comapi.openalex.org
Credit
Roman Kaiser, Givaudan; Braja Mookherjee and colleagues at IFF, who published on "live vs. dead" flowers in 1990
1988
Trial blends mixed by machine
At the end of 1988 Henkel's Fragrance Center in Düsseldorf installed a computer-run laboratory mixing station. Valves dosed liquid materials from tanks under nitrogen onto scales, weighing from 10 mg to 1,000 g, and a robot moved the containers on and off the scales, so that 100 to 200 or more trial blends could be made in 24 hours. Crystalline materials still had to be added by hand at manual stations, each checked by bar code. In July 1999 LUZI, a Swiss fragrance maker, began running a KUKA robot cell that loads and unloads its dosing system three shifts a day.
Lessonmachines took over the liquids first, and solids stayed with people.
Credit
Alexander Boeck, Henkel; LUZI AG, IMT Anlagen and KUKA
2019
Software proposes perfume formulas
Philyra, a system built by IBM Research for Symrise, learned from Symrise's archive of fragrance formulas, about 1.7 million by one account and around two million by Symrise's, and proposed new formulas aimed at a chosen group of buyers. Two perfumes for the Brazilian brand O Boticário, Egeo ON Me and Egeo ON You, went on sale in Brazil in June 2019 after Symrise perfumer David Apel adjusted the program's formulas.
Lessonneither company has said how much of each sold formula came from the program and how much from the perfumer.
Credit
IBM Research; Symrise; David Apel; O Boticário
Watch for
AI cuts development time
IFF estimates that AI cuts 25 to 30% of the time needed to develop and make a perfume, AFP reported. One of its tools, used by IFF perfumer Tanguy Guesnet, looks at the molecules in a chosen ingredient and suggests other ingredients that share them.
Achieved whenIFF or an outside party publishes development times for comparable briefs made with and without the tools.
Claimed by
IFF, November 2026
Generated food aromas, blind-tested
Researchers at the National University of Singapore and the Leibniz Institute for Food Systems Biology at the Technical University of Munich built KFO-Atlas, 896 key odorants drawn from 2,282 food aroma profiles, and KFO-Gen, a model that writes aroma formulas for a chosen food category. In their own blind test, 30 experienced panelists put 73% of 100 generated formulas in the intended category, and one meat-like aroma built only from plant-derived odorants was judged meat-like by all assessors.
Achieved whenanother laboratory mixes formulas from the released model and an independent panel confirms the categories.
Claimed by
J. Zhang, D. Zhang and colleagues, Advanced Science, September 2026
Open
01
Perfume formulas from old patents
An open set of complete fragrance formulas, each ingredient with its amount, taken from the worked examples in expired fragrance patents and tagged with the scent family each patent gives, such as citrus, floral or woody, plus a model that predicts the family from the ingredient list alone.
The briefCredit · unclaimed02
An open reaction-flavor map
An open data set of small heated batches of food-grade amino acids, such as cysteine and glycine, with sugars, such as ribose and glucose, at set pH, temperature and time, with the aroma compounds measured in each, and a model that predicts the main aroma compounds for conditions it has not seen.
The briefCredit · unclaimed
Open 01 · the brief
Perfume formulas from old patents
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open set of complete fragrance formulas, each ingredient with its amount, taken from the worked examples in expired fragrance patents and tagged with the scent family each patent gives, such as citrus, floral or woody, plus a model that predicts the family from the ingredient list alone.
Why it is still open
Philyra, the formula-writing software in the 2019 entry above, learned from Symrise's private archive of about two million formulas. Laura Sisson wrote in ACS Omega in 2025 that no public data set holds many blends of aroma chemicals with their amounts and expert odor labels, "although they likely exist in some proprietary form within major flavor and fragrance houses". Searches in November 2026 found no open set of formulas taken from patents.
Built with what exists
Full-text patents from Google Patents and the USPTO's bulk data, searched in the patent classes for perfumes (C11B 9/00) and perfume preparations (A61Q 13/00); a language model such as Claude to pull the formula tables out of patent text, with a person checking each one (the Note "Seven more flagged" covers a Claude-based reviewer of documents); PubChem's free API to match ingredient names to structures; RDKit; a gradient-boosting classifier as a first model.
First step
Pull 50 expired patents from class A61Q 13/00 and copy their example formulas into a table by hand, to fix the format before any automatic extraction.
Done looks like
At least 2,000 formulas from expired patents, each linked to its patent; a second reader re-checks 100 chosen at random and the error rate is published; a model that names the stated family for 200 held-out formulas, scored against always guessing the most common family.
Open 02 · the brief
An open reaction-flavor map
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open data set of small heated batches of food-grade amino acids, such as cysteine and glycine, with sugars, such as ribose and glucose, at set pH, temperature and time, with the aroma compounds measured in each, and a model that predicts the main aroma compounds for conditions it has not seen.
Why it is still open
Reviews name the amino acid, the sugar, temperature, time and pH as what decides a Maillard flavor. Published work optimizes one product at a time and reports only the chosen recipe: a 2026 study used a neural network to tune one seafood reaction liquid and gave its optimum (121 °C for 64 minutes at pH 6.5). Searches of papers and the web in November 2026 found no open data set spanning these factors.
Built with what exists
Food-grade amino acids and sugars; pressure-rated headspace vials in a heating block at up to 121 °C, in a fume hood in a university food-chemistry laboratory; GC-MS; a design-of-experiments plan and a gradient-boosting model such as XGBoost. Nothing is tasted.
First step
Run an eight-batch pilot (two amino acids, two sugars, two temperatures) in duplicate and check that the GC-MS repeats agree.
Done looks like
At least 200 runs with their peak tables released, and a model that predicts the presence of 20 chosen marker compounds on 40 held-out runs, with its accuracy for each compound published.