Design of experiments: why this experiment? Make the next one count
A new target lands: more flexural strength without losing elongation. You can't test every formulation, so the team spreads a grid across the design space or reruns last year's best recipe. And a suggestion that hides its constraints, targets and model sends scientists back to what they already know.
Uncountable's design of experiments and predictive tools train on the experiments you already have, predict each property with its confidence, and suggest the next round. Every candidate respects the formulation and process limits you set, each round balances exploration with regions that look promising, and Effect Sizes show which factors sit behind a suggestion. After you run the round, a predicted-versus-actual plot shows how close the model came.
This solution brief is for R&D leaders and formulation scientists who want fewer, better-chosen experiments that their teams will actually run.
Inside this solution brief:
- One guided flow from design setup and training set to the submitted design
- Constraints, performance targets and a predictive model for each property
- Exploration and exploitation in every round
- Effect Sizes and predicted-versus-actual checks
- Bodie AI Agent for analyzing results and recommending next experiments
- A worked example: one round of suggestions for four mechanical targets
Make the next experiment count. For more depth, read the guide Shortening the design-make-test-analyze loop.
FAQs
How Uncountable's design of experiments tools use your existing experiments and property targets to suggest the next formulations to test, from constraints and predictive models to exploration and exploitation, Effect Sizes, predicted-versus-actual checks and Bodie.
The training set comes from your existing experiments in Uncountable, so the first round of suggestions builds on past work rather than starting from a blank grid.
Exploration suggests experiments in areas where the model needs more information. Exploitation suggests experiments in regions that already look promising. Each round balances the two, and every suggestion is labeled with its sampling regime.
Each suggestion shows the constraints it respects, the target ranges and its predicted value for each output. Effect Sizes rank the feature weights for each output, linking a suggestion to the factors behind it.
Add the measured results after you run the round. On the predicted-versus-actual plot, points near the diagonal show closer agreement. You can tighten constraints or exclude an outlier from the training data and run the DOE again.
