Brewery Fermentations in the 21st Century
Fermentation us a critical step in the production of beer. During the process the sugar in the wort will be consumed, and new yeast cells will be formed with the production of ethanol, carbon dioxide, and flavour compounds.
A reproducible fermentation process is essential for the development of a consistent flavour profile.
The fermentation process may be impacted by:
- Temperature: low temperature reduces yeast activity leading to a slower process.
- Yeast viability: Low viability will also lead to a longer fermentation process.
- Oxygenation: Yeast requires oxygen during the early stages of fermentation for healthy growth and reproduction. Insufficient oxygenation can lead to slower fermentation rates.
- Wort Gravity: High gravities which contain higher concentrations of fermentable sugars, can result in slower fermentation as yeast may struggle to ferment all the sugars completely.
- Yeast Strain Characteristics: Different yeast strains have varying fermentation characteristics, including fermentation rates. Some strains ferment more slowly than others.
- Contamination: Unwanted microorganisms can inhibit yeast activity and slow down fermentation.
In summary, a fermentation is a relatively complex process with many factors influencing the final flavour profile.
Under normal circumstance it is expected that the process would follow a ‘standard’ temperature and gravity profile, which consists of five distinct stages.
How can data science help us achieve a more consistent product?
Machine learning algorithms are very good at learning from experience, and typically in breweries there is substantial history. Unfortunately this history is sometimes scattered across a combination of both electronic and paper based records. Nevertheless these records allow us to train a model.
This could be either a supervised model based on the KNN algorithm, in which the fermentations have been previously labeled as good or bad or alternatively an unsupervised model. However, in both case it needs to be trained with all the relevant information.
On completion of training the model can be deployed in the production environment. In this example it is fed with live production data which is used to classify the fermentation as being either good or failing. Where a failing fermentation is detected it could either:
- Alert the shift manager and then it is the responsibility of the shift manager what remedial action should be taken.
- Alternatively the output could be used to feed a rule engine.These rules have previously been provided by an SME which describes what remedial action should be carried.
This blog has focused on the fermentation process. However the general approach described is equally applicable to many batch industries.