Mikkeller said in our interview that a robot can also brew a beer (which caused a mini-controversy, e.g. here). But, can a robot, in the sense of an artificial being with artificial intelligence, also design a recipe? Lucerne University of Applied Sciences, Jaywalker Digital and MN Brew wanted to find out. And so did we.
Among the most valuable companies in the world are companies that make money with or more money because of data. Data doesn’t lie and data is better than intuition. And data is everywhere: not only photos, shares, likes and posts are data, even dialogues from TV shows are data. The latter can be used, for example, to have an artificial intelligence (or AI) write new dialogues for new episodes. The AI will figure out and use the most common words and word combinations. The results are currently bullshit, but eventually the results will be good or good enough.
Beer recipes are data
Beer recipes are also data that can be analyzed and used to create new recipes: Combinations can be analyzed and based on them an own recipe can be developed. IntelligentX has done this before and there is also a project at the Lucerne University of Applied Sciences and Arts (HSLU) where artificial intelligence wrote a beer recipe. This recipe was then brewed by MN Brew in Emmen and the result is solid.
But first to the process, which is described in detail here. In project of Marc Bravin (Algorithmic Business Research Lab of Lucerne University of Applied Sciences) and Kevin Kuhn (Jaywalker Digital), beer type, malt, hops, dryhopping, boiling time and “additional ingredients” are considered as parameters.
First, the program had to learn which types of malt and hops could be combined. Only then could it generate a recipe for a beer that is brewable and tastes good. To do this, it analyzed 157,633 beer recipes and identified 315 malt and 1,648 hop varieties as raw materials. Marc Bravin explains, “The neural network we trained for our «Brauer AI» has its origins in text processing. It is a so-called transformer network. Such networks can remember sequences that have already been generated, in our case recipe ingredients, with pinpoint accuracy.” Thus, the Artificial Intelligence was able to recognize patterns that were repeated in beer recipes and eventually the AI wrote a recipe.
315 malts and 1,648 hop varieties? There aren’t that many hop varieties. Adrian Minnig of MN Brew explains: “The algorithm recognized identical hop varieties with, for example, different alpha values, spelling mistakes in the analyzed recipes, etc. as a new variety if they could be found in the text. In the case of malt, there was also a challenge with packaging units and manufacturer names.” They would have pointed out to the developers why, in a development step, the hops that were within an alpha value range were grouped together.
Exciting intellectual exercise
The HSLU model is a fun, certainly interesting, but above all intellectual exercise that is particularly or exclusively exciting from a theoretical point of view. Neither the approach nor the idea of an optimized beer recipe is wrong. No, this one is exciting and the follow-up will certainly provide interesting results. Rather – presumably to reduce complexity – beer recipe-relevant paramaters are omitted here: No beer recipe is complete without yeast. Temperature control during brewing and fermentation will determine whether a beer is bad or great. An (incomplete) recipe alone does not make a good beer.
Brewer Adrian says, “The goal of the project was to see if a computer could provide new input to the brewer. From the beginning, it was out of scope for the generator to do everything – i.e. also take fermentation management into account. Instead, we wanted to provide brewers, probably primarily from the hobbyist sector, with a way to generate ideas. So not to build a machine to replace the brewer and his knowledge.”
Recipe already the best solved challenge
Understandable. But, we’ve never met a brewer who desperately needed recipes or complained about not having enough recipes. If anything, the recipe aspect is one of the best solved challenges in the brewing process. And AI cannot fulfill the requirement to develop the very best recipe, but arguably only an optimized recipe. This then probably corresponds to the lowest common denominator-unless there is a feedback loop where recipes are linked to a rating and thus raw material combinations are evaluated (something IntelligentX already does). With many unfit beer recipes out there, the question is whether all 157,633 recipes were a good source of data.
But of course we wanted to know how the beer tastes. Adrian has thankfully provided us with a few bottles: The beer has a nice, clear color. It’s quite pleasant to drink, though slightly sour and somewhere between mineral and metallic. Sparkling to well carbonated. There’s nothing the AI screwed up or did particularly well. Which can be considered a success at this time of development.
This all sounds very negative now, and it’s not really meant that way. We wouldn’t be surprised if AI is used to optimize recipes in the future – but then probably by large breweries rather than craft breweries. The work of HSLU and Jaywalker Digital is an exciting one and everyone involved has certainly learned a lot. The knowledge gained here can potentially be adapted to other fields. And it would be really exciting to try a beer optimized in all respects and all parameters.
That’s why we’re happy to toast the continuation of this project.