It's all in the blend

To better understand the role of machine learning as distiller, we first need to understand what gives whisky its distinct character. Whiskies aren’t just differentiated by their different ingredients, but also by the charred wooden casks they’re stored in. Rather than mere containers, the casks themselves play a vital role in giving each blend its unique flavor.

When whisky is first distilled, it’s a clear liquid that can have an elegant or a smoky character. To get the rich aroma, flavour and color we are used to seeing, this clear form, known as new make, needs to spend at least three years (usually much longer) in wooden casks. This is the maturation phase, where the all-important flavour infusion takes place.

Over time, whiskies slowly begin to take on the colour, aroma and flavours from the casks in which they’re stored, which also includes the flavours and aromas of their previous contents, such as bourbon, sherry, wine or other styles of spirit. “From these casks, we can generate hundreds of thousands of different whiskies,” states Angela D’Orazio, Master Blender of Mackmyra.

Master Distillers can spend their whole lives meticulously tasting, tweaking and experimenting to create the best flavours possible, turning acts of chemistry into a form of art – and this is where Mackmyra wanted machine learning to work its magic.

The process

Humans have always selected the different blends of ingredients and casks to create near-infinite flavor combinations. To do so, we first explored all current generative models but due to poor performance, ended up creating a proprietary generator-discriminator model that was designed to explore new spaces and generate unique recipes, but we also wanted to make the best possible whisky. We used previous recipe data, tasting notes, ratings of previous recipes, expert reviews, customer reviews and cask information – internal ratings, cask types, filling stages, volumes and alcohol levels to make our model understand what’s Mackmyra whisky. Then we created a framework which can innovate in this space, creating new whiskies that are unique but ultimately taste excellent.

This was not only faster than a person carrying out the process manually, but thanks to the algorithm’s ability to sift through and calculate a vast amount of data, new and innovative combinations that would otherwise never have been considered, could be found.

It’s important to stress, however, that this solution is not designed to replace a Master Blender. “The work of a Master Blender is not at risk,” Angela states. “While the whisky recipe is created by AI, we still benefit from a person’s expertise and knowledge. We believe that the whisky is AI-generated, but human-curated. Ultimately, the decision is made by a person.”