EngineeringNewVerified Real deployment at a named or precisely described company, with a measured result from a primary source.

AI proposes material formulations so R&D runs fewer lab experiments (Resonac, formerly Showa Denko)

Added 2026-10-19 · Last updated 2026-10-19

Role

Engineering / R&D (materials and formulation development) · Typical users: R&D chemist, materials scientist, formulation engineer, lab manager, in-house data scientist

The problem

Developing a functional material such as an adhesive, film, coating or compound means trying many combinations of raw materials and ratios until several properties meet target at the same time, often properties that pull against each other (strong adhesion but no residue when peeled off, transparency plus adhesion to copper). Skilled developers rely on knowledge and experience and repeat formulate-and-test cycles. Experiment records are scattered, and the qualitative information (spectra, structure images, notes) gets lost when data is reformatted for analysis.

How AI helps

Materials informatics: AI models trained on the company's own past experiments predict the properties of a formulation and propose candidate formulations that should meet the required properties (Showa Denko Materials cites Bayesian optimization for multi-objective targets). Developers then test only the most promising candidates. At Resonac, results typed into electronic lab notebooks flow automatically into the database and the models (MLOps), so developers always work with up-to-date predictions, and a data scientist who doubts a prediction can click back to the raw experiment data.

Step by step

1. Pick one development project with a clear multi-property target and a stack of past trials; record how many experiments and weeks similar projects took.
2. Gather past trials into one table: raw materials and ratios, process conditions, measured properties, plus the failed trials (they teach the model where not to go).
3. Standardize how new experiments are recorded (an electronic lab notebook or a shared template), so every new result can feed the model.
4. Train a model on past trials and have it propose a small batch of candidate formulations, with predicted properties and uncertainty; the chemist removes anything unsafe, unavailable or outside known chemistry.
5. Run the batch, add the results, retrain, and repeat until the target is met; compare experiments and time with the baseline.
6. Train a few developers as in-house data users (Showa Denko Materials certifies about 20 in-house data scientists a year) and extend to other projects.

Tools

• Resonac's in-house MI app and MI platform (built with QUATRE-i SCIENCE and Hitachi), with the BIOVIA Notebook electronic lab notebook from Dassault Systèmes
• Other tools to evaluate: materials informatics platforms (for example Citrine Informatics or Uncountable), and design-of-experiments software (JMP, Minitab) as a first step
• Any LLM with a data-analysis or code tool (ChatGPT, Claude, Copilot, Gemini) with the prompt below, on an export of your past formulation trials, to pick the next experiments

As reported by the source; most figures are vendor or self-reported and not independently verified.

Human review: cautions

• The chemist checks every proposed formulation for safety, compatibility and feasibility before it is mixed; models do not know your hazards or supply limits.
• Models extrapolate poorly outside the range of past trials; treat out-of-range proposals as exploration, not predictions.
• Results depend on clean, consistently recorded experiments; Resonac built an electronic lab notebook and data pipeline first, and had to win over researchers who found the notebooks time-consuming.
• The 30% figure comes from AI performance tests; the one-third figure is one adhesive project. Both are company-reported.
• Formulations are core intellectual property: do not paste them into a public chatbot; use an approved enterprise tool.

You are an experienced formulation scientist helping me plan the next round of lab experiments.

I will paste: (1) the target properties with limits [property, target, must-have or nice-to-have]; (2) a table of past trials [trial ID, raw materials and amounts or ratios, process conditions, measured properties, notes]; (3) constraints [raw materials allowed, ratio limits, safety or regulatory limits, cost limits].

Do this, using code on the data:
1. Check the table: missing values, units, duplicate trials; list issues instead of fixing them silently.
2. For each target property, show which ingredients and conditions seem to drive it, and where properties trade off against each other.
3. Propose 5 to 8 next trials that balance exploring untested regions with improving on the best trials, each with the reasoning and the expected direction of each property (not invented precise values).
4. Flag any proposed trial that falls outside the tested range, and anything I should check for safety or compatibility before mixing.
5. Say how I will know after this round whether to continue or change the approach.

Do not invent raw materials or data that are not in my tables.
Use only the data I paste in; if something is missing, write 'not found' or list it as a question instead of guessing. Do not send, order, post or change anything in any system: produce drafts for my review and approval.

Related cases

• AI predicts simulation and test results from your past data, so engineers run fewer prototypes
• Generative AI trained on past designs proposes new product designs that meet cost and manufacturability targets (Eaton)
• Find existing parts before designing or buying a new one (AI shape search)
• AI reads customer tenders and specification packages clause by clause and builds the deviation list (ROBEL, Bosch Manufacturing Solutions)

All 8 Engineering use cases →

Sources

• https://www.resonac.com/news/2022/11/02/2167.html
• https://www.resonac.com/news/2021/12/07/2071.html
• https://www.resonac.com/jp/news/2022/11/02/2164.html