Manufacturing
PFAS-free reformulation with AI: what your formulation history is worth
By SourceX Editorial · Updated
Short answer
AI PFAS-free reformulation works best when a coatings or adhesives maker has linked formulation history: formula versions, raw-material properties, test results and failed trials. Failed trials are often the most useful records, because they show a model what does not work. That history is worth most inside your own lab, and sometimes beyond it under strict terms.
Key takeaways
- AI-guided reformulation needs structured history that links each formula version to its test results.
- Failed and abandoned trials show a model the boundaries of what works, yet many labs never record them consistently.
- Raw-material property data is often the weakest link, especially for additives supplied under confidentiality.
- Formulas are usually trade secrets, so any outside use needs a narrow scope and reasonable secrecy measures.
- Organizing ELN, LIMS and spreadsheet history speeds reformulation whether or not any record is ever licensed.
What does AI-guided reformulation need?#
AI-guided reformulation needs a record of what a lab has already tried and how each attempt performed. Formulation informatics tools suggest candidate formulas that hit a target, such as water and oil repellency or flow and leveling without a fluorinated additive, by learning relationships between ingredients, process conditions and test results.
Removing PFAS often means replacing fluorosurfactants or fluoropolymer additives that delivered several properties at once, which is why reformulation can take many rounds of trials. A model trained or tuned on a maker's own history can narrow the search, but only if that history is structured enough for software to read.
The same records serve lower-VOC reformulation and other substitution projects, such as replacing a discontinued resin. A lab that organizes its history for PFAS work is building a reusable asset, not a one-off project file.
Which formulation records matter most?#
The formulation records that matter most link a specific formula version to its measured performance. The table ranks common lab records by usefulness for reformulation work and notes where they usually fall short.
| Record | What it teaches a model | Usefulness | Common gap |
|---|---|---|---|
| Formula versions with exact quantities | Which ingredient changes moved which properties | High | Versions overwritten in spreadsheets |
| Test results linked to formula versions | The performance outcomes the model learns to predict | High | Results filed by project, not by formula |
| Failed and abandoned trials | The boundaries of what does not work | High | Not recorded, or recorded without results |
| Raw-material properties and lots | How ingredient characteristics affect outcomes | Medium to high | Supplier data incomplete or confidential |
| Process conditions (mixing, cure, application) | Effects the formula alone does not explain | Medium | Kept in notebooks or operator memory |
| Customer feedback and field complaints | Real-world performance beyond lab tests | Medium | Kept in email, not linked to formulas |
Why failed trials are worth more than you think#
Failed trials are worth more than most labs assume because a model learns as much from what did not work as from what did. A history containing only successful formulas shows the model a narrow, biased slice of the design space and nudges it toward options the lab already ruled out.
Labs lose failed trials when results stay in personal notebooks, when spreadsheets overwrite earlier versions, or when a project closes without a summary. Asking chemists to log every trial with its result and a one-line reason for abandoning it costs little and improves both AI tools and the next chemist's starting point.
Raw-material data is the weak link#
Raw-material data is often the weakest part of formulation history because much of it comes from suppliers. Technical data sheets, safety data sheets and certificates of analysis describe ingredients at different levels of detail, and suppliers often withhold exact composition for proprietary additives.
For PFAS-free work, identifying which historical formulas contained fluorinated additives is itself a records task. Additive trade names, supplier disclosures and internal raw-material codes have to be mapped to each other, and supplier confidentiality terms limit what can be shared outside the company, including with a software vendor.
Getting formulation history into usable shape#
Getting formulation history into usable shape is mostly a records project, and it can begin before any AI tool is chosen. Prioritize the product families that need PFAS-free versions first, rather than trying to digitize the whole archive.
- Inventory where history lives: ELN, LIMS, spreadsheets, paper notebooks and project folders.
- Give each formula version a stable ID and link its test results to that ID.
- Recover failed trials from notebooks for priority product families.
- Map raw-material codes to suppliers, trade names and known fluorinated content.
- Standardize test methods and units, and note when a method changed.
- Flag customer-specific formulas and supplier-confidential data.
Where formulation data should live during an AI project#
Formulation data should live wherever the maker keeps the most control that its team can actually operate. The hosting choice decides who can see formulas, whether a vendor can learn from them, and how hard it is to leave the tool later.
Whichever option is chosen, record who can access the data, which product families were loaded and what the vendor contract says about training and deletion. That record is part of the reasonable measures that keep formulas protected.
| Option | Control over formulas | What to check |
|---|---|---|
| Tool installed on the maker's own servers | Highest; data stays inside | Internal IT capacity and model update path |
| Vendor cloud, dedicated environment, training off | High if the contract is clear | Training, retention and deletion terms; export rights |
| Vendor cloud with shared models across customers | Lower; learning may benefit other customers | Whether your data improves models others use |
| Research partnership with a university or lab | Varies by agreement | Publication rights, IP ownership and confidentiality |
Illustrative: a coatings maker inventories its lab history#
Illustrative: a fictional industrial coatings maker needs PFAS-free versions of several products that relied on fluorosurfactants for wetting and leveling. Recent work lives in an ELN, older formulas sit in spreadsheets, test results are in a LIMS keyed by sample number, and the earliest history is in paper notebooks.
The VP of R&D has the team link LIMS samples to formula versions for the affected product families and recover failed trials from notebooks where results were written down. Raw-material codes are mapped to suppliers, and every fluorinated additive is tagged.
The linked history goes into a formulation informatics tool hosted under the maker's control, with vendor training on its data switched off. Chemists use the tool to rank candidate formulas before running trials. The company keeps its formulas private and treats the cleaned history as a strategic asset.
What formulation history is worth outside the lab#
Formulation history is worth most inside the lab, and any outside value has to be weighed against trade secret risk. Formulas are typically protected as trade secrets, and under 18 U.S.C. 1839(3) that protection depends on the owner taking reasonable measures to keep the information secret.
Some makers may still license narrowly scoped records under strict terms, such as structured trial outcomes with exact formulas abstracted, test method data, or history from discontinued product lines. In SourceX Enterprise Data Value Framework terms, lab history tends to rate well on uniqueness, domain expertise and human-generated signal, and an exclusive license raises price; preparation cost, which is heavy when formulas must be abstracted, reduces net value.
SourceX starts with a metadata-only fit check, reviews supplier and customer terms in the Rights step of the SourceX five-step transaction, and the maker approves every step. A license grants defined use for a defined purpose; ownership of the lab history never leaves the maker.
Frequently asked questions
Do we need an ELN before using AI for reformulation?
Not necessarily, but you need structured, linked records, and an ELN is the usual way to keep them going forward. Historical spreadsheets and notebooks can be converted for priority product families. Starting an ELN or a consistent spreadsheet template now improves the data every new project adds.
Can supplier data sheets be used to train or tune models?
Use them internally with care, and check supplier terms first. Technical data sheets are often public, but confidential composition disclosures may carry use restrictions. Keep supplier-confidential data inside the company and out of any vendor-hosted tool whose terms do not meet those obligations.
Will an AI vendor own what it learns from our formulas?
That depends on the contract. Check who owns models tuned on your data, whether the vendor may use your data to improve products for other customers, and what happens to your data at termination. Ask for training on your data to be off by default unless you agree otherwise in writing.
Is formulation history from discontinued products worth keeping?
Often yes. Old formulas and their test results show how ingredients behaved across many conditions, and that history can inform reformulation of current products. Before discarding legacy notebooks or retiring old systems, record what they contain and archive the parts that link formulas to results.
Does reformulation history help with regulatory and customer questions?
It can support regulatory and customer documentation, such as showing when fluorinated additives were removed from a product and what replaced them. Keep a clear trail from raw-material changes to product records. Reporting duties vary by jurisdiction and change over time, so confirm current requirements with regulatory counsel.
Sources
- Under 18 U.S.C. 1839(3), information qualifies as a trade secret only if the owner has taken reasonable measures to keep it secret and it derives independent economic value from not being generally known to, and not readily ascertainable through proper means by, another person who can obtain economic value from its disclosure or use. Source
Related resources
See if your company qualifies
A short company assessment. No data uploads are needed.