Champagne Predictions with Deep Learning Model! Data Gets the Answer

What makes vintage champagne great? The answer remains in the deep learning model

We begin to learn about these wines and how to study them, because the more you know about your wines, the deeper you go into the house's knowledge and skills," said Denis Bunner, Bollinger's deputy head winemaker. "People were talking about climate change and how it would affect it, and I started to think that if I research some very hot years in the past, and analyze vintages from those years, I will understand better what will happen in the future." From there, the project of champagne predictions with a deep learning model heightened.

Working with a deep learning predictions datasets specialist and a mathematician, Denis Bunner started to develop a deep learning model that combined historical meteorological datasets; soil and plant conditions for each grape variety; timings of major seasonal events like de-budding, blossoming, ripening, and harvesting; and lab testing (for factors such as sugar levels and acidity) of the grapes and the wine itself. Bunner had the advantage of designing on Bollinger's own painstaking archiving, including meteorological data from local weather stations and information from the industry body (and his former employer) the Comité Interprofessional du Vin de Champagne (CIVC).

Earlier in 2021, Bollinger's winemakers were qualified for getting their first taste of La Grande Année 2014, a prestige fizz that had been kept for aging in the champagne house's cellars since it was blended. La Grande Année, Bollinger's flagship vintage champagne, is created only in years when the broad quality is deemed adequately high and enjoys seven years of aging under cork before it's launch. Ahead of opening up the 2014 vintage, questions remained over just how strong a year it really was, given a roller-coaster growing season that marked out record-breaking heat in June followed by a cold, wet summer that slowed grape maturation. Furthermore, for a champagne house well known for its forthright pinot noir character, it was a vintage that distinctly favored chardonnay.

But for Bunner, the answer was clear-cut even before the bottles were opened. Having spent two years combing through a mountain of historical data surrounding the interactions of terroir, vines, climate, and wine quality over the seasonal cycle, he was convinced that 2014 would be a home run, despite some of his colleagues' hesitancy. "I told them, 'No, it is really going to be a great vintage.' I was trusting deep learning for champagne predictions, and all the parameters were aligned," he says.

In Champagne, Bunner, the chefs de cave are accountable not just for constructing the complex blends that ultimately result in a finished wine, but for supervising the entire cycle of wine production, from first growth in the vineyards through to harvest, pressing, fermentation, and beyond—they are the guardians of quality, and the brains of the creative process.

However, Bunner's original conclusions were made not with the tastebuds, but by applying a deep learning model to a subject still often regarded, even by its practitioners, as an alchemical craft more than a science-led discipline. The jumping-off point was a determination to recognize the probable impacts of climate change on wine growing and quality. The deep learning predictions bank that enabled this was Bollinger's Wine Library, a unique collection of vintages going back generations that had been assembled from stock lying for years in the far reaches of the house's cellars. Several years back, a six-year project began to gather, taste, assess, and restore the hundreds of dust-encrusted bottles that now make up this liquid archive.

"We were looking at the maturation and ripening seasons of the vines, dividing each stage to find a relationship between the growing process and the final quality," Banner mentions. "We can see week by week exactly what the right scenario should be, and the farther you go from it, the more risk you have." Deep learning prediction is a model that can predict the strength of a season even as it unfolds, by adding an element of statistical prediction that is otherwise a slow, intuitive one that manifests in the taste of months. it happens. In the normal run of things, Banner says, about 4,000 barrels need to be tasted after the first fermentation before a clear view of the quality of the vintage can be achieved. "We may have an idea when we harvest grapes during harvest, but to be sure, we have to wait for that flavor," he further mentioned.