Formulation R&D with AI on your own lab data for paints, adhesives and plastics

What AI does in formulation development for paints, coatings and adhesives

AI in formulation R&D learns from the recipes your lab has already made and tested, predicts the properties of a recipe nobody has mixed yet and proposes the next batches most likely to reach the target or to teach the most. With the tens to few hundred documented recipes a typical industrial lab has per product family, the suitable methods are statistical models that run on any computer. A GPU is needed for reading old paper records, for a language model over lab notebooks and for neural networks on molecular structures.

The models learn which raw materials in which amounts, under which process conditions, gave which viscosity, gloss, adhesion, pot life or curing time.

This guide is written for heads of R&D and lab managers in coatings, adhesives, sealants, lubricants, plastics compounding and building chemicals, and has a technical section for the IT team that will run the system. It differs from our guide to a GPU server for chemistry and materials research, which covers molecular dynamics, DFT and double precision in research labs: here the subject is product formulation from your own test data.

Recipes are among the most sensitive data a company has. The ki-syndikat.de catalogue of AI use cases puts it in three words on its plastics page: "Rezepturen sind Betriebsgeheimnisse", recipes are trade secrets. That is the main reason to keep this work on hardware you control; our article on trade secrets, NDAs and AI tools covers the business side.

Machine learning on small formulation data: what tens or hundreds of recipes support

Formulation data is small by machine-learning standards. Take as an example a product family with 150 documented recipes, each with up to 20 raw materials, a few process settings and a handful of measured properties. For data of that size, methods that give a prediction together with its uncertainty are the usual choice.

Gaussian process models. A Gaussian process predicts a property and also says how sure it is. The scikit-learn documentation (version 1.9.1) describes the prediction as probabilistic, so that one can compute confidence intervals and decide where to measure next. The same page names the limit: these models "lose efficiency in high dimensional spaces – namely when the number of features exceeds a few dozens". A recipe described by 20 raw-material amounts and three process parameters stays below that limit; a recipe with 200 possible ingredients needs grouping first, for example by raw-material class.

Bayesian optimisation. Bayesian optimisation uses such a model to choose the next experiments: it proposes recipes where the predicted property is good or where the model is most uncertain, the lab mixes and tests them, and the model is updated. BoTorch, an MIT-licensed library built on PyTorch, implements the methods, and its README recommends the Ax platform to users who are not researching Bayesian optimisation themselves.

A published example shows the scale. In a 2022 paper in a Royal Society of Chemistry journal (DOI 10.1039/D2DD00029F), university researchers working with an adhesive manufacturer built a self-driving laboratory that optimised the resin-to-hardener ratio of a two-part epoxy for bond strength. It used Ax with a Gaussian process and batch acquisition; each round tested five formulations in triplicate, and the optimal region was clearly defined after four rounds. That is 20 formulations by our arithmetic. The paper names no GPU; the optimisation itself is light work, and the expensive part is the lab.

Gradient boosting and linear models. For predicting properties from many historical recipes, gradient-boosted trees and plain regression are often the first models to try. They also run on the processor, and they show which raw materials drive a property, which a chemist can check against experience.

Getting recipes out of Excel, scans and cards

Before any model, the recipes need to be in one table. The ki-syndikat page on recipe documentation in plastics lists where they are found today: Excel files, PDF scans, paper notes, handwritten production cards, PDF lab reports, scanned batch documents and supplier datasheets. It quotes a 2023 survey of 266 German industrial companies in which 76 per cent reported massive knowledge loss through staff changes, which is the business reason to start here even before any prediction.

Reading the paper records is GPU work. A vision-language model, an AI model that reads a page as an image, transcribes each card or report into fixed fields: raw material and supplier, amount, unit, process step, temperature, time, test method and result. The ki-syndikat page gives 95 per cent or more OCR accuracy for clean print and 60 to 75 per cent for hard-to-read handwriting without custom training, citing an OCR vendor's benchmark, and it describes the handwritten production cards of the 1990s and 2000s as often the oldest and most critical part of the records. It also states the consequence: every recipe extracted by OCR needs review by a specialist before it counts as a productive recipe. Plan the chemist's review time into the project; it is usually the largest effort.

Two details save work later. Use one name per raw material and map trade names and supplier codes to it in a table. And record units and test methods with every value, because a viscosity measured with a different spindle or temperature is a different number.

We supply lab workstations and servers with RTX PRO 5000 and RTX PRO 6000 cards for this work, and our Private AI/ML service sets up the transcription and the lab assistant; send us a description of your recipe data and the properties you test, and our team replies with a proposal.

Which steps need a GPU

scikit-learn 1.9.1 Gaussian process documentation, BoTorch README on GitHub, Chemprop 2.3.1 documentation and README, all read 25 October 2026; "runs on" is our assessment for data sets of hundreds of recipes.

BoTorch's README lists "native support for highly parallelized modern hardware (e.g. GPUs)" among its advantages, so the same code runs faster on a card when the models grow. With a few hundred recipes the gain is small; the card earns its place with the language and document models.

Molecular models. Chemprop, an MIT-licensed project, provides message passing neural networks for molecular property prediction, and its documentation (version 2.3.1) has sections on multicomponent models. Such models need a molecular structure for each component. Many raw materials in formulations are blends, resins or polymers sold under a trade name, for which no single structure exists, so these models help most where the ingredients are defined substances such as solvents, monomers or additives.

A lab assistant over notebooks and datasheets

The most immediate benefit for many labs is a language model that answers questions over the lab's own documents: earlier reports, notebook entries, raw-material datasheets and safety data sheets. A chemist asks which binders were tried for a given substrate, what happened to adhesion at low temperature, or which supplier changed a grade, and gets an answer with the source documents. Our guide to RAG on company data explains how such a search works; our guides on lab reports in product development and safety data sheets with local AI cover neighbouring tasks.

Qwen3.8-27B is a suitable open model for this: its card describes it as a native vision-language model, it is published under Apache 2.0, and its FP8 checkpoint is 30.9 GB, 28.75 GiB. By our users-per-card rule, 90 per cent of the memory the driver reports less 3 GiB per card, one RTX PRO 5000 72 GB holds 61.5 GiB, as our comparison of two RTX PRO 5000 72 GB or one RTX PRO 6000 Max-Q derives. That leaves 32.8 GiB of cache. Our Qwen hardware guide puts this model's cache, stored in FP8 as in all our sizing, at about 1.14 GiB per conversation of 32,768 tokens, linear-attention state included, so one card serves 28 such conversations at the same moment by our estimate. A lab team of 20 to 40 people rarely has more than a few questions running at once. The same model can transcribe recipe cards, since it reads images.

Workstation or server for the lab

NVIDIA RTX PRO 5000 product page (read 25 October 2026); usable memory and power by our estimates as in our RTX PRO 5000 and Max-Q comparison and the entry configuration on our AI servers page (cards, processor or processors, plus 15 per cent for the rest of the system).

Two RTX PRO 5000 72 GB in a workstation. The NVIDIA RTX PRO 5000 Blackwell is a 300 W dual-slot card with an active cooler; NVIDIA's page lists 72 GB of GDDR7 with ECC, 1,344 GB/s for the 48 GB model and MIG with up to two isolated instances. Two cards draw 600 W, and the workstation about 1.1 kW at the wall by our estimate, on an ordinary 230 V circuit. One card runs the lab assistant, the other runs transcription or training, or splits into two isolated instances for two jobs; NVIDIA's datasheet gives two instances of 36 GB for the 72 GB card, which NVIDIA's MIG guide does not list yet. Our guide to buying an RTX PRO 5000 72 GB workstation lists the parts.

Two RTX PRO 6000 Max-Q in a workstation. The NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition has 96 GB at 1,792 GB/s in the same 300 W per card. Two of them draw the same power as two RTX PRO 5000 72 GB and hold 83.0 GiB per card instead of 61.5, room for about 47 conversations of 32K tokens with the same model, and each card splits into two 48 GB or four 24 GB MIG instances. If the budget allows, this is the stronger lab workstation; the RTX PRO 5000 pair suits labs where the second card mainly runs smaller jobs.

A server for the company. Where other departments also plan AI work, the entry server on our AI servers page with two RTX PRO 6000 Server Edition cards, two AMD EPYC processors and 384 GB of memory serves the lab together with them, at about 1.8 kW by our estimate, in a server room. Access rights then decide who sees recipes: development sees formulations, production sees the parameters it needs.

Notes for the IT team

1. Software stack. BoTorch needs Python 3.11 or later and PyTorch 2.0.1 or later per its README; the documentation of Chemprop and scikit-learn showed versions 2.3.1 and 1.9.1 on 25 October 2026. The RTX PRO Blackwell cards are compute capability 12.0, so the PyTorch build must include support for it, as our guide to RTX PRO Blackwell software support explains.
2. Serving. Run the assistant model with vLLM behind an LLM gateway with keys and a query log. The Qwen3.8 card says thinking mode is on by default and can be switched off per request; keep it on for analysis questions and off for transcription.
3. Data model. Store recipes in a database with one row per component and batch, units and test methods in their own fields, and a link to the source document of every value.
4. Access. Recipes are trade secrets: separate roles for development, production and quality, keep the server off the internet, and log access to the recipe database.
5. Backups. The recipe database becomes the lab's memory; include it and the scanned originals in the backup plan.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build lab workstations with two RTX PRO 5000 72 GB or two RTX PRO 6000 Max-Q and servers with RTX PRO 6000 Server Edition cards to order, assembled and tested before delivery, delivered within the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up the transcription, the lab assistant and the search over your documents; tell us which product families and documents you would start with, and we propose a first step.

FAQ