Machine Learning for Sourdough Prediction

Machine learning is good at one thing — finding patterns in data. Sourdough has a lot of data. Here's what's actually predictable, and what isn't.

What ML can predict well

Bulk fermentation duration — given temperature, starter percentage, hydration, flour type, and starter activity, models can predict bulk completion within 15 minutes.

Cold proof timing — straightforward exponential function of temperature and starter strength.

Oven spring potential — given dough temperature, hydration, and proof percentage, models can predict expected rise.

Internal temperature curves — for any dough mass and oven temperature, models can predict when to remove.

What ML predicts moderately well

Crumb structure — better with cameras and image analysis. Hydration and fermentation time correlate strongly, but bench shaping technique matters more.

Flavor profile — fermentation conditions correlate with sour/sweet/complex, but flour quality and water mineral content matter heavily.

Crust color — depends on bake temperature, time, and steam — predictable, but home oven variance is high.

What ML predicts poorly

Shaping quality — depends on the baker's hands. No data substitute.

Aesthetic scoring outcomes — too dependent on blade angle, pressure, and dough condition at the moment of scoring.

Whether you'll like the bread — taste preference is personal.

One-time anomalies — bad flour batch, broken oven, contaminated starter.

What good models look like

Not magic. They're regressions on a lot of bakes.

A good fermentation model uses:

• Ambient temperature (most important)
• Dough temperature
• Starter percentage
• Hydration
• Flour protein
• Time of last starter feeding

Output: estimated time until bulk completion, with confidence interval.

Where ML helps real bakers

Schedule planning — "Based on your kitchen at 73°F, your dough will be ready to shape at 1:45 PM, and ready to bake at 7:30 AM tomorrow."

Anomaly detection — "Your bulk is taking 30% longer than usual; check starter activity."

Flavor recommendations — "For more sour flavor, increase your cold proof from 12 to 24 hours."

Recipe scaling — adjusts not just ingredients but timing and bake parameters.

Where ML is overhyped

Generative recipes — current LLMs hallucinate baking ratios. Don't trust a recipe without verifying baker's percentages.

Computer vision crumb analysis — works in research labs, less in home kitchens with phone cameras.

Personalized AI bakers — marketing.

Building your own data

Even without ML, logging your bakes systematically gets most of the benefit:

• Date, ambient temperature, starter activity (1–10)
• Hydration, starter %, flour blend
• Bulk start, bulk end (note duration)
• Folds (when, how many)
• Shape, proof type, proof duration
• Bake time, bake temperature, internal temp at done
• Result (1–10 scale, plus notes)

After 50 bakes, patterns emerge. After 100, they're statistically significant.

What I'd want from a sourdough AI

• Reliable fermentation timing predictions
• Schedule optimization based on my calendar
• Flavor outcome predictions based on my variables
• Anomaly detection on my starter
• Recipe scaling that handles fermentation correctly

What I wouldn't want:

• Auto-generated recipes
• "Personality" in the assistant
• Forcing me to use a specific brand of equipment
• Cloud requirements for offline tasks

The limit

The best baker in the world wouldn't be replaced by a model. The model assists. The hands and judgment are still the work.