How AI Flavor Profiling Works
One of the most common questions we get is: “How does an AI actually understand flavor?” It’s a fair question. Taste is subjective, cultural, and deeply personal. So how do we train a model to predict whether a formulation will taste good?
The Training Data
Our models are trained on a dataset of thousands of commercial beverage formulations paired with sensory panel scores. These panels — conducted by trained food scientists — rate beverages on dozens of attributes: sweetness intensity, acidity balance, bitterness, mouthfeel, aftertaste, and overall acceptability.
By mapping ingredient combinations and ratios to these sensory scores, our models learn the relationships between chemical composition and perceived flavor.
Feature Engineering
Raw ingredient lists aren’t enough. We engineer features that capture the chemistry:
- pH prediction based on acid and buffer concentrations
- Brix estimation from sugar and sweetener blends
- Flavor compound interactions — some ingredients amplify each other, others mask
- Viscosity modeling from hydrocolloid and fiber content
- Color prediction from natural and artificial colorant concentrations
The Model Architecture
We use an ensemble approach combining gradient-boosted trees (for tabular ingredient data) with a small transformer network (for capturing complex ingredient interactions). The ensemble outputs predicted scores across our sensory dimensions, plus confidence intervals.
Practical Application
When you adjust a slider in Beverage Creator — say, increasing citric acid by 0.1% — the model instantly recalculates predicted flavor scores across all dimensions. You can see in real-time how that change affects sweetness perception, acidity balance, and overall score.
This tight feedback loop is what lets brands iterate in minutes instead of weeks.
What’s Next
We’re actively working on incorporating consumer preference data (not just trained panels) and expanding our models to predict flavor stability over shelf life. More on that soon.