AI Beer: Artificial intelligence brews beer
Adrian Minnig, technical director at the Swiss microbrewery MN Brew, is particularly mindful of tradition. But recently he has begun utilizing artificial intelligence to develop recipes. In this interview, he talks about his experiences and the first attempts at brewing AI beer.
MN Brew – craft beer brewery from Switzerland
Mr. Minnig, could you first tell us about your brewery MN Brew?
Our brewery is quite small. While we’re a team of just six employees, we only work part-time for the brewery. But it’s not even a part-time job really. We actually run the brewery as a hobby, and all of us have separate full-time jobs. However, in terms of the time requirement, it is a pretty ambitious hobby. Combined, we are currently investing roughly two full-time positions in the brewery.
We started in 2015 after having a crazy idea. One member of our team originally wanted to make whisky. Then, he found out he’d first have to produce beer, or at least something similar to beer. So he started with a 30-liter home brewing system. As a result, more and more people joined the team.
Adrian Minnig is the technical director at the microbrewery MN Brew, Rothenburg, Switzerland, where he puts “electric ideas” to work. When Minnig isn’t in the brewery, he works as a logistics engineer in a military maintenance company.
We built our first brewing system from stainless steel pots, and then automated it. But soon after – in 2017 – we bought a professional 500-liter brew house and four brewing tanks. This remains our setup today. And we really do everything in the brewery, from brewing and filling to labeling, all in our 120 square meters. If things are going well for us, then we’ll brew once weekly. Of course, last year it was a bit less due to the coronavirus.
AI beer brewed out of a college project
Let’s discuss “Deeper,” the beer you brewed using artificial intelligence. How did a regional craft beer hobby brewery come up with this project?
One of our team members is a technical employee at the Lucerne University of Applied Sciences and Arts. Through him we met Kevin Kuhn, the initiator of the AI project. Kuhn knew that we often worked with the college, especially on art projects for our labels. He wanted to know if we saw any opportunity to collaborate on brewing beer using artificial intelligence. Then, the working group from the college stopped by to inspect the brewery and equipment and find out how the brewing process works.
Wait, so the team from the college developed beer brewing software using artificial intelligence without knowing the details of how beer brewing works in practice?
The AI working group students were just looking for applications for their software. During their studies, they developed a food recipe generator. This project was about assembling a menu. They even had the Lucerne University canteen chef prepare a complete meal according to an AI recipe. That worked brilliantly, and the idea came to them one evening over a beer: “If it works so well for food, then it’ll work for drinks, too!” They chose beer because they enjoy it most. During their research, they found beer recipes online. They initially developed their AI for beer based on what they had interpreted from these recipes.
So, the first AI draft was perhaps a bit naive. These were essentially malt and hops figures in percentages. They didn’t yet realize that temperatures and times are also essential in brewing.
Did things then become clearer for them in the brewery?
That’s right. The college team went back to the laboratory after visiting us, modifying and developing their AI software. Now, “brewable” recipes have emerged based on the composition of the raw materials. The AI now accounts for things like the maximum ratios of specialty malts and the cooking times for hops. A remaining major challenge for version 2.0 is incorporating the resting temperatures and times.
And then you got to take over this second AI draft – with the exception of the mash program – and get started right away?
As brewers, we still needed to specify a few parameters. For example, we have set the target alcohol content in our brewing software. The AI hadn’t exactly worked out a completely finished recipe. But we were essentially able to implement the recipe according to the AI specifications with only a few adjustments.
AI still seems to be in its infancy here. We also checked whether the AI was producing recipes that worked for different beer styles and recipes. Looking at the first 100 recipes that the AI created for us, after around 80 we could say: “Yes, they look good. We can implement them straight away!”
What kind of beer is the AI beer “Deeper”?
We decided on a Session IPA because we wanted to stay below 6% ABV. It’s a typical American IPA: fruity, light in color and very refreshing. The AI actually suggested lactose as an ingredient, which is perhaps one of the truly exciting aspects of the recipe. We would have thought of that ourselves. As a non-fermentable sugar, the lactose adds some residual sweetness.
Are you already planning more brews together with Lucerne University?
We want to brew an AI beer once or twice a year. One brew – 500 liters. The goal shouldn’t be to dictate to the AI. Instead, we brew whatever recipes it produces. It’s exciting for our customers to know that we will always have something new in our range that they have never drunk before.
Tens of thousands of recipes as an AI database
On what basis does the AI work? Abstract rules or concrete recipes?
The AI processed every recipe on https://www.brewersfriend.com/. Home and hobby brewers utilize this platform to exchange and share their recipes. In total, the AI has analyzed and incorporated over 160,000 recipes.
So you could think of it like a hive mind that is forming the basis for the AI recipes.
Yeah, you could say that. Then again, we also found that some recipes still contain errors. So you have to be very careful. After all, there is no real quality assurance on the platform. When someone posts a recipe, no return value is provided on brewersfriend.com about whether the result is good or bad.
Would it be possible for the AI to check this and provide the hobby brewer with a prognosis as to whether his recipe idea is good or bad?
We had a similar idea: A database stores beer recipes generated by the AI. If you brew and taste the beer, you can give the AI your feedback. This feedback then enables the AI to learn or to develop even better recipes.
But to go one step further: Perhaps at some point you could ask what the best Pilsner recipe for the region of Southern Germany looks like. As the taste there is probably different than in Switzerland, for example.
Clear recommendation for brewers
What’s your take-away on developing recipes with artificial intelligence?
We’ve brewed three “Deeper” brews so far, and the beer is selling very well! I’d recommend that every brewer embark on the AI beer adventure at least once. If only because it opens up the possibility of discovering ingredients that you might not have thought of. I already mentioned the AI’s idea of adding lactose. It also suggested two different hops with aroma profiles we didn’t know at all before. So, the AI encouraged us to look into these and discovered new varieties for us.
Mr. Minnig, thank you very much for the interview!
Want to share your developments and innovations in the beverage industry to an international specialist audience? Then we invite you to take part in the next drinktec from September 12 to 16, 2022 in Munich.
This article is powered by BRAUWELT.