Beautiful Future: How Deschutes Uses Artificial Intelligence & Machine Learning to Brew Better Beer

Barley, hops, water, yeast, and… the cloud?


Ask any brewer and they’ll admit that while beer has likely been around since the dawn of civilization, we’re all still learning new ways to brew it more efficiently, creatively, and quickly. But balancing the brewer’s art with modern approaches to automation, measurement, and decision making requires brewers to toe a fine line. Take the personality out of the process, and you sacrifice the “craft” in craft beer. Ignore the best tools available, and you waste precious resources that could be better spent on the creative side of the brewing equation.

From their outpost on the eastern edge of the Cascades in Bend, Oregon, Deschutes Brewery has tackled this problem in a forward-thinking way, embracing their brew team’s passion for tech and programming. Through their operational technology team, they’re using a cutting-edge approach to brewing technology aimed at saving time and money, making higher-quality beer, and in turn freeing up company resources for an aggressive innovation program. The equation, at its core, is pretty simple: Produce the same amount of beer in less time, while maintaining or improving the quality of the beer along the way, and you’ll have more resources for the intentional play that leads to new beers that drinkers love.

Heading the charge is Brewmaster Brian Faivre (pictured below), a computer science major with a coding background. He turned to brewing for a career but never lost that passion for programming. About five years ago, as machine learning and artificial intelligence concepts were filtering out of the most advanced research institutions and becoming more widely available, Faivre grabbed a book on the subject and started researching.

“I realized, we could totally do this,” Faivre says. “A lot of these machine-learning concepts were now accessible to us, and were more mainstream. Open-source software was growing more and more accessible. So, we asked ourselves, ‘If we’re going to pick something, what would be beneficial?’”

Brewing is generally a process based on measurement and decision making. The technology for constant real-time measurement of fermentation parameters in a tank is extraordinarily costly, so most brewers large and small use manual measurements by their cellar staff to drive the decision-making process.

“In a handful of the transitions—fermentation, free-rise, diacetyl rest, and cooling—a brewer or a lab tech has to go out, get a sample, prep the sample, and do that analysis,” Faivre says. “If the fermentation hasn’t quite reached our expected parameters, then we wait—we always err on the side of caution due to the quality concerns there. But that extra time reduces our potential capacity. If we’re now spending six, eight, 12, or 24 hours longer per fermentation, it just adds up.”

Measuring every hour or two isn’t feasible, due to the labor involved in measurement, so naturally the cellar team errs on the side of caution. If the next measurement takes place a half or full day later, that might not seem like much—but when you consider the number of times that tank is turned per year, and how many tanks the brewery uses for fermentation, the impacts get big, quickly.

The Deschutes operational technology team knew that they could use machine learning to improve their approach to predicting when fermentations would actually finish.

“The concept was to look at apparent degree of fermentation, take a measurement to see where we’re at, compare to past data, and develop a future prediction,” Faivre says. “Ideally a brewer could get a prediction, given the time we’re at right now, and know that in this many hours we should be able to go into free-rise, and in this many hours we should be able to bung the tank.”

As a fermentation progresses, the machine-learning system would adjust and continue to learn. But to develop an accurate and verifiable predictive ability requires training the system on reliable data. Enter Senior Data Analyst Kyle Kotaich, a physics-major-turned-production-brewer who joined Deschutes back in 2014, and who has been a crucial part of the machine learning project.

“The beginning of any AI or machine-learning project is making sure you have a good data structure, and it’s accessible by all the tools you want to use,” Kotaich says.

They started by developing a core data structure, then fed in years of time-stamped measurement data from thousands of fermentations. As they plugged in the data, a curve took shape. They were able to construct an algorithm that could predict the future progress of a fermentation based on the current time and measurements. While not all fermentations are the same, as new measurements are fed into the system, the algorithm can accurately adjust its prediction for the end of fermentation. Rather than a past process of giving the fermentation another 12 to 24 hours in order to make sure it has finished, the predictive system narrows that window down to minutes.

One of the biggest challenges has been convincing the experienced professionals in the cellar that the prediction from the machine-learning system can be trusted.

“The buy-in process for bringing AI into a traditionally creative process can pose a challenge,” Kotaich says. “Some say brewing is an art, but it does have a strong foundation in science. And what we started doing was using our knowledge of the science and the process. We put a lot of time and energy into validating—training the model with historical data, and data it hasn’t seen before. But being able to present it to people statistically, in a way they can understand that the data presented by the algorithm is just as good—if not better—than somebody going and taking a measurement. That, in itself, was a bigger project than developing the algorithm.”

Today, cellar operators at Deschutes have such a high level of confidence in the algorithm that they typically allow the software to trigger next steps in the brewing process. Occasionally, they’ll encounter an anomaly that requires double-checking, but that’s grown more and more rare.

Brewery Pi

Deschutes’ approach to software and technology also has paid dividends for the broader brewing community, as the brewery has released some of the tech they’ve developed into the public sphere through the Brewery Pi program. An inexpensive approach to brewery data tracking and visualization, Brewery Pi runs on $50 hardware and is thoroughly extensible, so that brewers without a programming background can set it up for their own needs.

Conceptually, it’s pretty simple: a system that allows brewers to define the parameters they want to track and visualize; it then automatically charts and graphs those data points over time, so that brewers can see how those measurements are developing.

Most businesses that develop proprietary systems keep that advantage for themselves, but Deschutes is committed to sharing the technology and learning with brewers everywhere.

“There are a ton of people in brewing who can talk about how to make an IPA recipe, but less so about these things, so we thought it was a great place for us to contribute,” Faivre says.

The Bottom Line

Machine-leaning prediction has helped the brewery improve efficiency, but how does that translate into beer quality and creativity? Most beer consumers aren’t concerned with how efficiently or cost-effectively a brewery makes their beer—they want high-quality beer, and they want new and exciting beers. The machine-learning project has certainly impacted quality, ensuring that the beer they brew is tight to the specifications they’ve developed for each brand. But the other impact—freeing up resources that could then be invested in innovation—is also significant.

“We have six new 1,000-barrel fermentors sitting on the backside of the building, that are probably 70 percent installed,” Faivre says. “But we pulled the plug on that project, because this machine-learning system freed up all the fermentation capacity that we need. This project has really paid off as we push our capacity limit. Say we gain four additional fermentations per fermentor per year, and we have about 40 of those—it really adds up.”

After the first two to three months of using the system, Kotaich reviewed the impact and found that it had reduced total fermentation time by 206 hours. Now, the brewery has been able to reduce planned tank residency time by about 36 hours per fermentation. It’s easy to see how that will continue to pay off in the future.

Rather than investing in more tanks, the brewery has instead been able to invest heavily in the innovation process—but more on that in the next installment of this series. For now, Faivre is excited that the brewery has built a team that can harness new technology to solve problems efficiently.

“We’re gaining the knowledge internally, and we’re right at the spot where everything is coming together to do more interesting research and work,” he says. “We’ll see where we can leverage some of these learnings to continue exploring the space. I love the fact that it’s helping to learn and understand more of what we do when we make beer.”