Predictive Models Built on Your Chemistry
Most formulation models are trained on ingredient percentages alone. The model learns which named materials appeared in which recipes, not why those materials worked. That makes it fragile at exactly the moment it is most useful: a new supplier grade arrives, the name is unfamiliar, and nothing in the training data connects it to anything the model already knows. The learning curve starts again, and the experiments start again with it.
Uncountable holds ingredient attributes and experiment results in the same system. Each ingredient is a reusable record carrying the physical properties your chemistry depends on: molecular weight, density, glass transition temperature, particle size, or anything else you measure. Input calculations weight those properties by the amount of each ingredient in a formulation, producing a physical profile for every experiment, such as weighted average molecular weight, weighted average Tg, or average D50. Those calculated features become predictors alongside the ingredient amounts, so a model trained on your history learns the relationship between chemistry and performance rather than between a product code and a result. Because the same attributes describe every raw material in the library, a grade the model has not seen before is scored on its properties. Physical constraints you set are carried into the optimization loop, so suggested experiments stay inside the ranges your formulation has to respect.
See how it works on your own formulation history by booking a personalized demo.
FAQs
A predictor derived from the physical properties of the ingredients in a formulation rather than from the ingredient names. Weighted average molecular weight across a resin blend is a chemistry-informed feature; "60% Acrylic A" is not.
Any numerical attribute stored against an ingredient. Molecular weight, density, glass transition temperature, and particle size are common, but the set is defined by your chemistry, and once an attribute is recorded it is available to every model built on that library.
The new grade is described by the same attributes as everything else in the library, so its formulations carry the same calculated features the model was trained on and can be scored without retraining.
Constraints are set on the calculated physical properties, not just on ingredient amounts, so optimization only proposes experiments whose predicted profile falls inside the ranges you defined.
The model is trained on your own experiments and your own ingredient library, so it learns relationships that hold in your chemistry rather than patterns drawn from someone else's data. Pattern recognition only works on history captured in a comparable format, which is why the attributes and the experiment results sit in the same system instead of being assembled fresh for each modeling exercise.
