Recently, our machine learning team spent a day at Beiersdorf's AI Month, the internal event the skincare and personal-care leader runs to bring its scientists and data teams up to speed on where AI is heading. Alongside my colleagues Andrej Patoski and Alessandro Zambusi, I gave a talk called "Accelerating R&D with Bayesian Optimization: Inside Uncountable's ML Engine." This post is the written version: a short account of what we covered, and the takeaways for any R&D team weighing how machine learning fits into experiment design.
What We Shared at Beiersdorf's AI Month
The premise of the talk was simple. Designing experiments is one of the most expensive things an R&D team does, and too often the next experiment is chosen by intuition or by changing one variable at a time. There is a smarter way. Probabilistic machine learning, specifically Gaussian processes and Bayesian optimization, lets the data itself point to the experiment most worth running next. The rest of this post walks through how that works, and how it is built into Uncountable.
Why Is Traditional Experiment Design So Inefficient?
Traditional experiment design wastes trials because it explores the formulation space slowly and without memory. One-variable-at-a-time testing changes a single factor while holding the rest fixed, which misses interactions between variables and burns through runs. Even classical design of experiments (DOE), a statistical method for planning multi-variable tests, is usually set up once and rarely revisited as results come in. In formulation-heavy work, where a single experiment can take days and real materials cost money, every redundant run is margin gone and time lost.
What Is Bayesian Optimization, and How Does It Work?
Bayesian optimization is a method for finding the best inputs to a system in as few experiments as possible by learning from every result along the way. It works as a loop. First it builds a model of how your inputs, such as ingredients and process conditions, relate to the outputs you care about, such as strength, viscosity, or stability. Then it uses that model to choose the next experiment, balancing two goals: testing where the model predicts strong performance (exploitation) and testing where the model is most uncertain and could learn the most (exploration). You run that experiment, feed the result back, and the model sharpens. Each cycle moves you closer to target with far fewer trials than guessing or brute force.
Why Gaussian Processes?
Gaussian processes are what make this work with the small, expensive datasets that real labs produce. A Gaussian process is a probabilistic model that predicts not just a value but how confident it is in that value. That confidence estimate is the key ingredient: it is what lets Bayesian optimization judge where exploring is worth the cost. Unlike data-hungry deep learning, Gaussian processes perform well with the modest number of experiments a lab can realistically run, which is exactly the regime R&D operates in.
How Does This Live Inside Uncountable's Platform?
In Uncountable, this engine runs on top of your structured, centralized experiment history rather than a one-off spreadsheet. Because formulations, process conditions, and results are captured in one connected data model, the models can learn from everything your organization has run, not just the project in front of you. A scientist describes the formulation constraints and the performance targets, and the platform suggests the next experiments most likely to hit them, with the reasoning visible rather than hidden in a black box. The sequencing matters here: structure first, AI second. A model is only as good as the connected data underneath it, which is why the data layer comes before the algorithm.
Key Takeaways
For anyone who could not be in the room, here is what we would want an R&D leader to walk away with:
- Let the data choose the next experiment. Gaussian processes and Bayesian optimization turn experiment design from a series of educated guesses into a guided loop that improves with every result.
- Uncertainty is the point, not a footnote. Because a Gaussian process reports how confident it is, the method knows where it is worth exploring, which is what makes progress possible on small, costly datasets.
- Structure comes first, AI second. The engine is only as strong as the connected data underneath it. A single structured record of your experiments is what turns machine learning from a pilot into a capability.
- The goal is to augment the scientist, not replace them. The method encodes a scientist's constraints and goals, then handles the combinatorial search, so experts spend their time on the science rather than on bookkeeping.
See It on Your Own Formulations
Uncountable's ML engine is built into the same platform where your R&D, QC, and product data already live. Book a demo and we will show you how Bayesian optimization applies to your work.
