Problem
A natural cosmetics recipe goes through planning, suitability review, prototyping, evaluation and revision, and final approval, and the evaluation and revision step repeats many times. Researchers set a formulation from experience, make a prototype, get feedback, and start over if it falls short.
Natural ingredients vary in color, scent and marker compound content from batch to batch, so the same formulation sheet can still give different properties. A single factory is expected to have a few thousand formulation records at most and a few hundred at least, and most regional cosmetics companies are too small to run their own AI team.
TendiLab Solution
We predict target properties from existing formulation data and new experiment results, and recommend new formulation ratios that meet the desired properties. The goal is to turn a hundred attempts into ten combinations recommended by AI.
First we check whether the recipe data we received is enough to predict the target properties. If it is, we predict with the property prediction foundation model. If the data falls short, we gather formulation data in the same format from the literature, add it, and train again.
The screen researchers use shows each prediction together with which ingredients contributed to the property. If exposing ingredient information is a concern, we can start from data with the ingredient names replaced by aliases.
Target KPIs
- Turn one or two local natural ingredients into data the AI can use.
- Derive prototype recipes that contain those ingredients.
- Success means getting the desired result when the product is made as the AI recommends.
Next Steps
We have finished the preliminary discussions and defined the technical requirements together. Next come organizing the natural ingredient data and deriving prototype recipes. We will share results after the project ends.