AI-guided formulation & recipe development (coatings, adhesives, cosmetics, food & flavours)

A formulation is a mixture of known ingredients whose properties depend on proportions, processing and interactions that are hard to model from first principles. Formulators today rely on experience and trial-and-error. Small-data machine learning and Bayesian design of experiments (DoE) learn from the lab's own history, predict performance, stability, sensory profile and cost, and propose the most informative next experiments, so fewer lab iterations are needed to hit a specification or to replace a restricted ingredient.

Problem & value

• Today: formulation chemists in specialty chemicals, paint, adhesives, cosmetics, food and flavour houses iterate by experience and classical DoE. Knowledge sits in individual heads, spreadsheets and unstructured ELN entries; staff turnover loses it. Each cycle (mix, cure or age, test) takes days to weeks.
• AI value: reuse historical experiments as training data, search the mixture space with surrogate models and uncertainty-aware acquisition functions, and capture knowledge in a structured platform. Typical targets: faster time-to-spec, fewer experiments, cheaper substitutes, and faster reformulation when regulation removes an ingredient.
• Evidence level: vendors report large experiment reductions (for example Citrine says customers cut the number of experiments by 50–90% depending on the project, in an interview [2]). Treat these as vendor claims that depend on data quality and problem type; run your own back-test before committing.
• Regulatory pull: REACH restrictions, the EU Cosmetics Regulation ingredient bans and sugar/salt reduction goals force repeated reformulation with unchanged performance [4][5].

Variants

• Coatings, inks, adhesives, polymers: multi-objective optimisation (viscosity, cure, adhesion, gloss, VOC, cost); drop-in replacement of restricted substances (for example PFAS, certain biocides, solvents).
• Cosmetics and personal care: texture, stability and sensory targets; claim and ingredient-list compliance under Regulation (EC) 1223/2009 and the product safety report [4].
• Food and beverage: sugar, salt or fat reduction, plant-based analogues (NotCo's Giuseppe is a well-known example), shelf-life; Nutri-Score as an optimisation target or constraint; Novel Food screening (Regulation (EU) 2015/2283) when a candidate ingredient lacks pre-1997 EU consumption history [8].
• Flavours and fragrances: creation from a brief, reformulation to match a benchmark, sensory-panel-in-the-loop (for example Givaudan Carto in the fragrance field).
• Pharma excipients/drug products: adjacent but under GxP; platforms exist, validation burden is higher.
• Private use: recipe tweaking for home cooks is covered in PERS-004, not here.

How it works

1. ELN/LIMS + raw-material DB + sensory/test data
2. data curation (units, ingredient ontology, outlier and failed-run handling)
3. featurisation (composition fractions, process settings, ingredient descriptors / fingerprints)
4. surrogate models (GP / ensembles / small neural nets, optional property foundation models)
5. constrained, multi-objective Bayesian optimisation (qEHVI-class acquisition under mixture constraints)
6. proposed experiment batch (formulator reviews, edits, approves)
7. lab runs
8. results back into ELN
9. retrain (active-learning loop)
10. regulatory/cost screening before scale-up

• Models: Gaussian processes and tree ensembles dominate for tens to a few thousand experiments; they give calibrated uncertainty, which drives exploration. Graph or descriptor features of ingredient molecules help transfer across recipes.
• LLM layer: extracts recipes and results from legacy reports and ELN free text, translates a customer brief into target specifications, and answers questions over formulation knowledge ("what did we try with this thickener?").
• Human-in-the-loop: chemists set objectives, constraints and acceptable ranges, approve each batch, and own safety and regulatory sign-off. The model suggests; it does not release a product.

Data preparation

The decisive data are past experiments: recipes, process settings and measured properties. In an SMB they sit in Excel sheets, lab notebooks (often paper), ELN/LIMS exports, ERP raw-material masters and PDF test reports, and much knowledge lives in the heads of senior chemists.

Typical gaps: inconsistent ingredient names and units, missing process parameters, failed runs that were never recorded (yet they are highly informative), mixed measurement methods, scanned reports and spreadsheets with merged cells.

Preparation steps: - Map data sources and gaps against the target properties (PREP-001, PREP-002); find shadow Excel files (PREP-003). - Digitise notebooks and old reports with OCR (PREP-006) and tidy spreadsheets into one table per experiment (PREP-011). - Build an ingredient master with synonyms, CAS numbers and units (PREP-012, PREP-015). - Record failed runs and define a standard capture form for new experiments (PREP-008). - Define which properties count as success and keep a hold-out test set (PREP-018).

Tools: an ELN or a simple database, OpenRefine, pandas, RDKit for descriptors. Plan for 50-60 % of the total effort in data work; a few hundred clean experiments beat thousands of messy ones. Use GUIDE-003 for the readiness check.

Implementation paths

Which path for whom

• SMB: Path A, because platforms such as Uncountable or Citrine bundle data capture and optimisation for teams without data scientists; insist on EU hosting and data export, and move to Path C if recipes are extremely sensitive.
• Enterprise: Path C, because formulations are core IP, volumes across sites justify an internal platform, and the Bayesian-optimisation stack is light enough for HW1.
• Developer: Path C, because BoTorch/Ax, RDKit and an open extraction model give full control over mixture constraints.

Path A — Buy (off-the-shelf)

• Formulation/materials-informatics platforms: Uncountable (R&D data platform plus ML for formulation-heavy industries), Citrine Informatics (sequential learning), Albert Invent (chemistry-native AI; raised a USD 55M Series B led by TCV in Oct 2026 per press coverage [see Sources]; customer and partner claims are vendor-reported), Polymerize, MaterialsZone, Kebotix-style autonomous offerings.
• Food/flavour: NotCo's Giuseppe (licensing model), Givaudan's in-house tools (not for sale as software), plus sensory-analytics vendors.
• EU-based or EU-hostable: check data residency, as most leaders are US-headquartered; ask for EU hosting, IP ownership of trained models and exit/export of data. Smaller EU vendors and consultancies exist, but verify case studies.
• Choose A when: you have 1+ product families with at least a few hundred documented experiments and no in-house data-science team.

Path B — Build on cloud APIs

• LLMs (Claude, Mistral, Gemini, GPT in EU regions) for ELN text extraction and brief parsing with structured outputs; tabular ML in your cloud tenant (Azure ML, Vertex AI, SageMaker, or Databricks).
• Optimisation with open Bayesian libraries on managed compute. Integrate ELN/LIMS via API (for example Benchling, LabWare, Dassault BIOVIA, SAP).
• Choose B when: you have data engineers, need custom constraints, and want to avoid platform lock-in.

Path C — Self-host (open source)

• Bayesian optimisation: BoTorch/Ax (PyTorch), scikit-optimize, Optuna, GPyTorch, plus pyDOE/DoEgen-style classical DoE [3]. Mixture designs (simplex-lattice, extreme-vertices) belong in the first rounds.
• Cheminformatics: RDKit for ingredient descriptors; open materials/chemistry models (for example MACE, ChemBERTa-class encoders) when ingredient structure matters.
• LLM extraction: an open-weight model of 7-30B (for example Qwen, Mistral, Llama families) served via vLLM or Ollama.
• Hardware: HW1 suffices for tabular/GP work; HW2 only if you add local LLM extraction.

Path D — Custom training

Worth it for large, long-lived product portfolios (hundreds of thousands of records across sites): train in-house property predictors or fine-tune a molecular/materials foundation model on proprietary data. Needs curated, consistently measured data and a data-science team; failed experiments must be recorded, because negative results are very informative.

State of the art (2026)

• Bayesian optimisation with mixture constraints and batch acquisition is the practical workhorse; commercial platforms wrap it for non-programmers and add LLM assistants (for example Albert Invent's assistant features; verify current product names) to query knowledge and draft experiments [1][2].
• Spec-to-formulation assistants propose starting recipes from historical analogues; quality depends entirely on data coverage.
• Self-driving labs (SCI-007) close the loop physically for powder/liquid dosing; most EU Mittelstand firms still keep chemists in the loop with manual or semi-automated mixing.
• Limits and failure modes: small, biased datasets (only successful recipes recorded); batch-to-batch raw-material variation; extrapolation outside the explored region; unmodelled ageing and scale-up effects (lab mixer vs production); sensory targets are noisy and panel-dependent; secondary sources conflict on vendor details, so verify claims in pilots. Models rarely replace stability and safety testing.

Shared resources & combinations

• Same stack (tabular-ml + optimization-solver): OPT-031 (blend/feed/concrete optimisation), OPT-009 (process parameter tuning on the machine), SCI-003 (new materials; SCI-014 uses known ingredients).
• Upstream/downstream: SCI-018 (analytical data interpretation feeds property data), SCI-007 (lab automation closes the loop), SCI-015 (in-silico toxicology and REACH/cosmetic safety screening of candidates), DOC-023 (SDS extraction and restricted-substance register), DOC-027 (label and claims proofing).
• Shared infrastructure: ELN/LIMS data pipeline, ingredient ontology and a vector or graph index of lab reports are reusable by the other R&D cases; GPU needs are modest.

Effort, cost & KPIs

• Path A: weeks to pilot, 3-6 months to embed; subscription pricing, plus data-migration effort that usually dominates.
• Path B/C: 3-9 months with one data scientist and a formulation lead; main cost is data cleaning and ontology work, not compute.
• KPIs: experiments (or lab days) to reach specification, time-to-market per formulation, hit rate of proposed batches, share of reformulations passing first scale-up, raw-material cost per kg, number of records usable for training.

People & change

Management decides on budget and on how much IP may leave the company; involve the department lead (R&D head) and the formulators who will use the proposals. IT handles ELN/LIMS integration and hosting, data owners clean the legacy records, and legal or compliance reviews supplier NDAs and, for cosmetics and food, the regulatory consequences of new ingredients (GUIDE-010).

Works council: depends. Pure experiment data does not monitor people, but ELN and platform logs record who ran which experiment, which can make the system suited to monitor performance under §87(1) no. 6 BetrVG. Inform the council early under §90(1) no. 3 and agree that logs are not used for performance evaluation (GUIDE-004). Public bodies involve the staff council.

DPO: personal data is limited to user accounts and author names; an AVV with the platform vendor is still required (GUIDE-005).

AI literacy: support formulators with training on uncertainty estimates and extrapolation limits (GUIDE-006). Start with one product family and 2-3 chemists as pilot group, compare proposals against their own picks, and set clear go/no-go KPIs (GUIDE-009). Typical concerns are "the model takes over my craft" and trust in black boxes; address them by letting chemists set constraints, approve every batch and own safety sign-off (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: minimal risk; no Annex III category for internal R&D tooling. AI literacy duty (Art. 4) applies. No Art. 5 prohibited practice is involved. Sensory panels with staff or consumers are ordinary research, not emotion recognition. [6]
• GDPR: mostly no personal data; panelist data, consumer-test responses and researcher identities in ELN logs are personal data, so apply consent, minimisation and pseudonymisation.
• Sector rules: REACH/CLP (restrictions, authorisation, SVHC, SDS) [4], Cosmetics Regulation 1223/2009 (Annex II banned substances, safety assessment) [5], Novel Food Regulation 2015/2283, EU food-additive and labelling rules, flavouring rules (Regulation 1334/2008), and biocidal products rules. The AI proposes; qualified safety assessors sign off.
• IP and trade secrets: recipes are core IP, so prefer contracts with no training on your data, EU hosting, and export rights; protect under the EU Trade Secrets Directive.
• Mitigations: keep model suggestions in an audit log, require a compliance screen before pilot batches, validate with confirmation runs, and never skip stability, safety and toxicology testing.

Getting started

1. Pick one product family with a clear spec, 200+ documented experiments and a painful bottleneck (for example a restricted-ingredient swap).
2. Audit data: export ELN/LIMS records, harmonise units and ingredient names, and include failed runs.
3. Back-test: hold out the latest experiments and check whether a baseline model (GP or random forest) would have predicted them; run 1-2 vendor trials on the same data.
4. Run 3-4 closed-loop rounds of 5-10 proposed experiments each, with chemists approving every batch.
5. Add compliance and cost screening, then extend to further families.

Sources

1. Albert Invent raises USD 55M Series B (Oct 2026) — https://en.wowtale.net/2026/10/15/235504/
2. Citrine Informatics — AI cuts R&D experiments (vendor interview) — https://greyb.com/blog/citrine-informatics-scouted-interview
3. BoTorch — Bayesian optimization in PyTorch — https://github.com/pytorch/botorch
4. ECHA — REACH regulation overview — https://echa.europa.eu/regulations/reach/understanding-reach
5. Regulation (EC) No 1223/2009 on cosmetic products — https://eur-lex.europa.eu/eli/reg/2009/1223/oj
6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. Huntsman with Citrine Informatics (AZoM) — https://www.azom.com/article.aspx?ArticleID=25176
8. Regulation (EU) 2015/2283 on novel foods — https://eur-lex.europa.eu/eli/reg/2015/2283/oj
9. The SaaS News — Albert raises USD 55M Series B (Oct 2026) — https://www.thesaasnews.com/news/albert-raises-55m-series-b/