Industries / Mining and materials / Chemicals and petrochemicals / Specialty chemicals
Formulation
Formulation is choosing which ingredients go into a product such as a paint, a glue or a flavor, and in what amounts, so that it does its job.
5
Milestones achieved, from 1941 to 2023.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1941
Paint thinned with water
Sherwin-Williams chemists made Kem-Tone, a fast-drying, washable water-based paint for household use. In 1948 Glidden's research center in Lakewood, Ohio developed an interior latex paint, a paint whose binder is a fine dispersion of plastic particles in water.
Credit
Sherwin-Williams; Glidden
1956
Glue set by surface moisture
Harry Coover of Eastman Kodak patented cyanoacrylate adhesives. The patent, filed in 1954 and granted in 1956, describes bonding parts quickly without heat or pressure by moistening a surface and spreading a thin film of the liquid between the parts. The National Inventors Hall of Fame says the discovery led to industrial, consumer and medical uses, best known as superglue.
Credit
Harry W. Coover, Eastman Kodak
1967
Three numbers for solubility
In his 1967 doctoral thesis at the Technical University of Denmark, on the importance of solubility in surface coating formulation, Charles Hansen showed that how well a solvent dissolves a polymer or a pigment can be described with three numbers: one each for dispersion (van der Waals) forces, polar forces and hydrogen bonding. Formulators use these Hansen solubility parameters to choose solvents and to judge which ingredients will mix, since materials whose three numbers lie close together tend to mix.
Credit
Charles Hansen
2002
Butter flavor harms workers
In May 2000 eight former workers at a microwave-popcorn plant were reported to have severe bronchiolitis obliterans, a scarring that blocks the lungs' smallest airways. A 2002 study of the plant's current workers found that airway obstruction rose with their exposure to diacetyl, the main ketone in artificial butter flavoring. In 2012 scientists at the US National Institute for Occupational Safety and Health (NIOSH) found that 2,3-pentanedione, another butter-flavor chemical, damaged rats' airways much as diacetyl did.
LessonA substitute ingredient needs its own evidence that it is safer than the one it replaces, because a chemical that tastes alike can act alike in the body.
Credit
Kreiss and colleagues, New England Journal of Medicine; Hubbs and colleagues, NIOSH
2023
A model names smells
A team led by Brian K. Lee and Alexander B. Wiltschko trained a graph neural network, a model that reads a molecule as a network of atoms and bonds, to predict how a molecule smells. On 400 molecules it had not seen, its predicted odor profile came closer to a trained panel's average than the median panelist did, as reported in Science. This lets new aroma ingredients for perfumes and flavors be screened on a computer before anyone makes them. An outside developer published open code, OpenPOM, that reproduces the model.
Credit
Lee, Wiltschko and colleagues; OpenPOM by GitHub user ARY2260
Watch for
AI halves coating development
Advanced Polymer Coatings (APC), a maker of polymer coatings and linings, says Albert Invent's AI tool for planning experiments and tuning formulations cut its development time by 50%. The report names no baseline project or period and cites no outside check.
Achieved whenAPC or an outside party publishes development times for comparable projects run with and without the tool.
Claimed by
Advanced Polymer Coatings and Albert Invent, July 2026
AI-found fragrance ingredients
Osmo says its AI models screen billions of molecules for new fragrance ingredients, and in May 2026 it presented Tornadew, a watery melon and cucumber scent it says had never been made before. Osmo also says it filed more fragrance patents in 2025 than the rest of the industry combined. Avery Gilbert, who writes the FirstNerve newsletter on smell, counted five published patent applications and no granted patents for Osmo in 2025.
Achieved whena perfume or flavor house outside Osmo sells a product containing one of these ingredients, or an independent sensory panel confirms that the odor is new.
Claimed by
Osmo, January and May 2026
Tint batches corrected in advance
Tints are the concentrated colors mixed into paint. Axalta says TintMaster AI, an internal system that predicts a tint batch's color from past production records and adjusts the recipe before the batch is made, raises the share of batches that are right the first time by up to 29% "in certain manufacturing scenarios". The system won a 2026 Edison Award, judged by a panel of business and academic leaders; no test has been published.
Achieved whenAxalta or an outside party publishes before-and-after counts of color corrections per batch over a full year of production.
Claimed by
Axalta, April 2026
Open
01
Recipe tables from patents
A free, checked dataset of the worked examples in adhesive and coating patents. Each row gives one example's ingredients, their amounts, how it was cured, and the test results the patent reports, such as lap-shear strength (the force needed to slide two glued strips apart) or pencil hardness. Add a simple model that predicts a test result from the recipe, scored on patents it has not seen.
The briefCredit · unclaimed02
Formulation optimizers, raced fairly
A public test in which formulation software proposes recipes one round at a time against a simulator built from measured data, scored by how many simulated experiments each needs to reach a target. Open-source optimizers run first, and vendors are invited to run theirs on the same simulators.
The briefCredit · unclaimed
Open 01 · the brief
Recipe tables from patents
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A free, checked dataset of the worked examples in adhesive and coating patents. Each row gives one example's ingredients, their amounts, how it was cured, and the test results the patent reports, such as lap-shear strength (the force needed to slide two glued strips apart) or pencil hardness. Add a simple model that predicts a test result from the recipe, scored on patents it has not seen.
Why it is still open
CheMixHub, the largest public benchmark for predicting the properties of chemical mixtures (June 2025), covers solvents, ionic liquids, fuels, drug solubility, polymer electrolytes and smells, and has no task for paints or glues. CrowdChem, a Tokyo company, trained a model on about 700,000 process records from nearly 9,000 patents and papers, including adhesives and paints, curated by human experts, and its paper states no public release of that data. A 2026 study that used machine learning to suggest two new anticorrosion coatings worked from 492 salt-spray results on recipes with 148 ingredients, a dataset its authors call relatively sparse.
Built with what exists
Google Patents full text, searched by patent class (C09J for adhesives, C09D for coating compositions such as paints); a small AI model built for extraction and long documents, such as Meta's Muse Ember (see the Note "The jobs agents repeat"); RDKit, an open-source chemistry toolkit, to link ingredient names to structures where a patent gives a chemical name; CheMixHub's code for data splits and baseline models.
First step
Pull 50 epoxy adhesive patents from Google Patents and enter their example tables by hand, to settle the columns and how to record trade names (a supplier's brand name for a resin).
Done looks like
At least 3,000 worked examples from at least 500 patents on two product types, two-part epoxy adhesives and water-based acrylic paints. Two people re-enter 200 randomly chosen examples by hand, and the dataset agrees with them on at least 95% of ingredient amounts. A baseline model's error on held-out patents is published with the data and code under an open license.
Open 02 · the brief
Formulation optimizers, raced fairly
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test in which formulation software proposes recipes one round at a time against a simulator built from measured data, scored by how many simulated experiments each needs to reach a target. Open-source optimizers run first, and vendors are invited to run theirs on the same simulators.
Why it is still open
Olympus, the open benchmark suite for experiment-planning software from the Aspuru-Guzik group, holds 43 datasets, from chemical reactions to solar cells and color mixing, and none for adhesives or coatings. A 2021 comparison of optimizers on five measured materials datasets, by Liang and colleagues, used silver nanoparticles, perovskites, a polymer film and two 3D-printed structures, again with no coating or adhesive. Vendor claims such as APC's 50% (see Watch for) give no baseline that others can rerun.
Built with what exists
Olympus, which builds simulators from measured data; BoTorch, an open-source library for Bayesian optimization (a method that chooses the next experiment from the results so far); the measured epoxy adhesive data published in 2019 and hosted by Japan's National Institute for Materials Science, which started from 32 adhesive samples and added three rounds of five; recipe tables from the item above as they appear.
First step
Put the NIMS epoxy adhesive results into a table and fit a first simulator with Olympus.
Done looks like
Simulators for at least three measured coating or adhesive datasets; at least five optimizers, including random picking as a floor; 50 repeat runs each; the median number of experiments needed to reach 95% of the best measured result, published with all code.