The problem: too many formulations to test
Even three ingredients and three target properties give you more possible formulations than any lab can run. A fixed grid spreads your runs evenly across the whole space, so most of them land far from the best region.
AI design of experiments can pick better runs, but teams only act on a suggestion when they can see why the model made it.
What this demo shows
- The model predicts each property from your past runs, along with how confident it is.
- Exploration picks spread out across promising areas you haven't tested yet; exploitation picks target the best expected result.
- All suggestions stay inside your ingredient and process limits, such as binder plus solvent at 80% or less.
- Each round learns from the last, so adaptive rounds spread out and then close in on the best region.
- In an acrylic topcoat project, the suggested formulations show predicted gloss, viscosity and hardness, the sampling regime for each, effect sizes, and training accuracy against measured results.
Why it matters
Fewer runs to reach a formulation that meets spec, and a clear reason behind each one. Scientists can see what the model expects, how much each ingredient drives each property, and how close past predictions landed, before they commit lab time.
See it with your own data
This video is part of DOE and Predictive Tools. To see how it would work with your own formulations, request a demo or talk to us about an ML proof of concept. Our team replies within 30 minutes during business hours.
Read the transcript
You can't test every formulation. So why trust the experiment the model picks? It predicts each property from your past runs, along with how confident it is. Some picks spread out across promising areas you haven't tested yet. Others target where it expects the best result. All of them stay inside the ingredient and process limits you set. Measure the results and rerun, and the next round learns from them.
Take binder and solvent. A fixed grid spends runs across the whole space. Adaptive rounds spread out, then close in on the best region.
In Uncountable, check it yourself: each suggestion's predictions, and whether it explores or targets the best result, how much the model expects each ingredient to change each property, and how close predictions land to measured results.
