Back-tracing Detected or Predicted Flaw-compounds to their Production Issue in Beer through Probabilistic Density Clustering in Graphical Databases


Back tracing beer flaws


Notice anything backwards, missing, misleading, or wrong?


Authors


Ryan Ahn
Director of Research & Development, Founding Team

Jason Cohen
Founder & Lead Data Scientist

Abstract

Flaw-compound recognition is a powerful tool for beverage producers. Traditionally, finding a flaw compound in a complex beverage product required the use of analytical lab equipment such as GC-MS or HPLC, or by hiring and extensively training tasters for a sensory panel. The incredible power of the human palate and nose are often overlooked and, as we show in this paper, offers a more cost-effective alternative method of determining the cause of flaw-compounds via probabilistic density clustering in graphical databases. The methods employed stem from chemical-informatics, machine learning, and sensory science.

Background

The human palate can detect tens of thousands of different chemical compounds, some at incredibly low thresholds; this is what truly matters when it comes to making a consumer product. Consumer purchasing decisions are based on expectations and experience of hedonic enjoyment from a flavor profile. Thus, sensory-based quality control is a critical component for any food or beverage producer with a valuable brand. At the end of the day, sensory has final say as to whether or not a particular batch of beer makes it to market. Conventional individual flaw recognition by sensory evaluation involves “spiking” a complex beverage with a known flavor standard, such that it is above the aroma and flavor thresholds of the average person. From this, expert panelists will train their palates to recognize that flaw. This is both expensive and unreliable. If a panelist gets sick, confused, or incepted, it can drastically change what he/she tastes. Through the Gastrograph System, panelists do not have to know what a flaw tastes like. Over the 24 different flavor attributes and mouth feels, we train our models on the unique “fingerprint” each flaw leaves on our system. From this, we can quantify the probability of a specific flaw, taint, or contamination from purely sensory data without the panelists even knowing the specific taste of a particular flaw compound.

A challenge has always been extracting meaning out of sensory data. Naturally, the first questions any brewer would ask are what caused the problem, why, and how to fix it. Through extensive collection of data from purposefully contaminated beers, Analytical Flavor Systems has developed methods for individual flaw recognition. The process that produces each flaw-compound leaves behind a unique “fingerprint” on the flavor profile of a beer. This allows for the identification of where in the production process a specific flaw-compound was produced. This analysis is quite powerful as it allows brewers to leverage that data to solve these problems in the future. The true value of sensory is being able to leverage sensory data to trace particular flaws back to a certain issue during production.

Methods

Flaws, taints, and contaminations can arise in beer due to a variety of issues - from low quality ingredients and bacterial contaminations, to bad packaging and poor cleaning practices. It can often be difficult determining exactly what causes these flaws solely based on the taste or chemical composition. The fundamental assumption behind back-tracing chemical flaws to a particular issues during the beer making process is that certain flaw-compounds, precursors, and cofactors occur together. Analytical Flavor Systems has built proprietary graphical databases that contain the production process for making beer, potential issues that can occur at each step, and flaws and flaw-groups that can occur because of them. These databases contain thousands of nodes per step and trace the entire bio-chemical pathway of every flaw, taint, and contamination that can appear in beer at each stage of the production process. Paired with machine learning and artificial intelligence techniques, including our Flaw Identification Analysis, we can back-trace to the production step where the problem occurred and make actionable recommendations to improve your production process and prevent similar flaws in the future.

For example, if our system flagged a sub-set of the following flaws: cis-3-hexanol, acetic acid, p-menthene-8-thiol-3-one, acetaldehyde, benzaldehyde, and trans-2-nonanol as contaminants in a particular beer with a high degree of certainty, our system can back-trace that combination of flaws to the issue of post-production oxidation. These are common flaws that arise from too much oxygen in the headspace of beer bottles, a bad bottling line, beer that is stored too long in a growler, or a faulty keg.

Furthermore, not all of these flaw compounds would necessarily have to be flagged for our system to back-trace and identify this issue. Our system searches for clusters of high probability around a particular issue-of-origin node within our graphical databases to determine where the compound was most likely to have been produced. The image below is a simplified partial schematic of our graphical database showing how a yellow issue node (production issue where chemical flaws originate) can be traced back by high likelihoods of occurrence of the red chemical flaw nodes.

A simplified representation of AFS's graphical beer flaw database. In this example, all red flaw nodes had some significant likelihood of being present in the beer. These flaws were all 'traced back' to the yellow issue node.

The other fundamental assumption we make is that every production issue or set of production issues has a unique set of chemical flaws. We use probabilistic clustering and our graphical databases to flag and trace-back production issues to let the brewer know what went wrong and how to prevent that from reoccurring. Different flaws are given different weights in the determining of which issue they could arise from depending on many factors such as beer style and flavor interactions when occurring with other flaws and flavor compounds.

Results & Future Work

Through probabilistic density clustering, Analytical Flavor Systems is able to effectively determine the most likely causes of flaws, taints, and contaminations, thereby leveraging sensory data to its most actionable and informative conclusion.

These types of individual flaw recognition models and flaw-tracings via probabilistic density clustering are far superior to conventional methods of sensory evaluation and data leveraging for two reasons: they quantify the likelihood of a specific flaw and quantify the likelihood of where in your production the flaws occurred.

Our current probabilistic models are very strong and can identify individual flaws to within 95% accuracy and some even to 99%. From this, we can back-trace to where an issue could have occurred, but in doing so we assume that flaw compounds taste and act independent of each other. So far we have discovered that our models still work well given this assumption, however we aim to further strengthen these models by purposely contaminating beer with more than one flaw and with flaws that usually occur together as a result of a particular production issue. From this data we aim to teach our algorithms the different production issue “fingerprints.”

In future work, Analytical Flavor Systems aims to further expand the power and scope of human sensory by utilizing similar methods that are commonplace in analytical chemistry techniques of quantification. By building individualized models of each of our clients’ products, our algorithms will be able to predict the chemical concentration of flaw compounds using sensory data alone. By changing the concentration of a known flaw in a beer, we can create a calibration curve from which we can predict approximate concentration. This concept can be thought of in very simple terms. In analytical chemistry you have an instrument, say a UV-vis. You run known standards at different concentrations from which you get an output from the instrument measuring concentration. You plot the data and perform the appropriate regression. This mathematical function can then be used to quantify the concentration of a known sample. This same concept can be applied to sensory: you have known standards at different concentrations and the human nose and palate are the “instrument” from which they report an output measured in intensity of flavor attributes across 24 dimensional flavor space. We then create the appropriate regression to model the data to produce a calibration curve, which we can use to predict the concentration of a known flaw in beer.

Conclusion

Sensory can be an incredibly powerful tool in beer quality control, but that data is usually not leveraged beyond a “go, no go” or a “like, or dislike” decision. Yet, the reality is that human sensory data, if collected on a complete sensory system such as Gastrograph, can be interpreted though an array of analyses and applied at each of the critical control points. Through the purposeful contamination of beer, Analytical Flavor Systems has developed models to flag and predict the likelihood of common flaw compounds in beer. These probabilities can be used to determine where in production these issues could have occurred via probabilistic density clustering in graphical databases that contain the “textbook” knowledge of beer production issues and resulting chemical flaws. This analysis of human sensory data is far more informative and actionable than the current scope of most sensory-based quality control programs and offers real benefits to breweries through better quality control and through actionable quality assurance improvements.