Small data, expensive experiments
Small datasets are common in coatings R&D. The experiments behind them are often expensive and time-consuming.
How can we build reliable predictive models from limited data and use them to decide what to test next? Dr. Ivanov will explore this question in his talk.
What the talk covers
The talk covers how probabilistic modeling, uncertainty quantification and Bayesian optimization can support more efficient experimentation. He will also discuss how smart agents can help build and automate the workflows around these methods.
Predictive models for small data
We designed STOCHOS, our predictive engine, for limited, expensive datasets. Its predictions come with an uncertainty band, so you see where the model is reliable and where it is not. A model is trustworthy only inside the range its data covers, and it has to be checked before its results are used. Our products track that range and show it.
For coatings R&D, recipes and process conditions can enter as numeric or categorical inputs, with constraints such as a mixture totaling 100 percent. We use models trained on existing experiments to explore and screen formulations; the final recipe still requires confirmation in the lab.
Working toward fewer experiments in coating formulation
We use Bayesian optimization for experiments to select the next lab tests from model predictions and their uncertainty. After those tests, the results are added to the data and the model is trained again. Failed experiments and results that violate constraints also inform which regions to avoid.
Our approach to AI for coating formulation applies this adaptive loop to targets such as color, gloss, and viscosity, including targets that compete.
Date and place
- Date: Tuesday, 27 April 2027
- Session: Digitalisation II
- Conference: European Coatings Show Conference 2027, 26 to 28 April 2027
- Venue: NürnbergMesse, Nuremberg, Germany
- Time: Not yet announced
If you are attending ECS 2027, we would be happy to connect. Get in touch to arrange a meeting.
Coatings R&D at PI Probaligence
We exhibited at the European Coatings Show 2025, where automated design of experiments and intelligent control of high-throughput experimentation drew the strongest visitor interest. At the GDCh Lacktagung 2025 in Aachen, we gave a talk and had a booth on AI-driven adaptive design of experiments for coatings development.
Learn more about our work in paint and coatings: AI for Paint and Coatings.
