Artificial intelligence may not be able to appreciate the taste of a glass of red wine, but apparently it can tell you where it came from. And by doing so, modern technology might confirm the classic concept of terroir.
A team of scientists used machine learning and gas chromatography to successfully analyze 80 wines and identify which seven Bordeaux châteaus produced them. Even more astonishing, when they mapped the wines’ chemical signatures as XY coordinates on a map, the wines clustered correctly according to their Left and Right Bank origins. This breakthrough research validates the concept of terroir and has far-reaching implications.
“This study demonstrates the remarkable potential of gas chromatography analysis to explore fundamental questions about the origin and age of wine,” said the lead authors, Alexandre Pouget, a computational neuroscientist at the University of Geneva, and research enologist Stéphanie Marchand of the Institute of Vine and Wine Science (ISVV) at the Université de Bordeaux. They published their findings in the journal Communications Chemistry.
Combining Wine Passion With Science Know-How
The impetus behind the project was a question that had plagued Pouget for decades: Could he use machine learning to study wine? Born and raised in Paris, he has had a lifetime passion for wine. As a young neuroscientist in the late 1980s, he studied the brain with machine learning, a type of artificial intelligence that identifies patterns in large data sets.
His work took him to California, New York and Geneva, but the question of whether he could use machine learning to analyze wine lingered—until six years ago when he and his wife visited Veuve Clicquot for a private tasting. He mentioned his idea, and the winemaking team pointed him in the direction of the ISVV in Bordeaux and to Marchand, who has specialized in the chemical analysis of wine since 2008. The project then got its start as a humble “Saturday experiment” in the ISVV’s chemistry lab, said Marchand.
The idea was to use gas chromatography (GC) and electron ionization mass spectrometry paired with the computational powers of machine learning to identify specific wines by their molecular signature. “In this particular paper we asked two main questions: Can we identify the estate, independent of vintage? And can we identify the vintage, independent of the estate?” said Pouget.
Gas chromatography has been used for years to analyze wine. It’s a straightforward process that vaporizes a chemical mixture, in this case, wine, and records the various molecular components in the form of a chromatogram. “The measurements look like an electrocardiogram with peaks but it’s not repetitive,” said Pouget. Each peak is made of multiple points, and when you get a peak, it means there is more of that substance. Further analysis is done to determine what the peaks are. “It’s a lot of work and for a variety of reasons, they need to manually calculate how big the peak is.”
As a neuroscientist, Pouget has used machine learning to identify patterns in large data sets related to the brain. Handling such large data translated to the chemical analysis of wine, because when a wine is put through GC, the resulting data set is vast—30,000 different points on the chromatogram.
Machine learning makes analysis much faster and easier as the algorithm is designed to know where to look in the data. It also removes the human tendency to second guess. Some of the peaks are tiny, so a researcher will make arbitrary decisions on which might be important. “But with machine learning the algorithm will go in there and pay attention to the things we want it to,” said Pouget. “The algorithm will figure out which part to look at depending on the question you ask.”
Perfect Identification
Marchand approached seven châteaus, all classified-growth estates who have chosen to remain anonymous, with three on the Right Bank in Pomerol and St.-Èmilion and four on the Left Bank. She selected 12 vintages from the period of 1990 to 2007. While the surface area of estates has changed over time, particularly in the Médoc, none had changed during the period under study. Pouget did not know the appellations of the châteaus prior to seeing the results.
The team trained the algorithm with the chromatograms of 73 wines. Then they ran seven wines through as a test to see if the algorithm could identify the château that produced each wine. They ran the experiment 50 times, each time retraining the algorithm on a different set of 73 wines, with seven wines kept aside for the test. The machine did a perfect job, correctly identifying the estate 100 percent of the time. When it came to vintage, the algorithm was able to correctly identify the vintage 50 percent of the time.
At first the researchers thought the results could be picking up on the blends of grape varieties at the estates. “We thought that the machine might be picking up differences inherent in the grape varieties—but it was not that at all,” said Marchand. In fact, the entire chromatogram was “informative with respect to geographic location and age, thus suggesting that the chemical identity of a wine is not defined by just a few molecules but is distributed over a large chemical spectrum,” according to the study.
“For the seven châteaus, it’s amazing that they do have a specific identity,” said Pouget. “To me, it’s wonderful news.”
A Terroir Map
But that wasn’t all. Their research had revealed something they had not expected.
The algorithm summarizes the 30,000 measurements for each wine’s chromatogram in two numbers—this is called dimensionality reduction—and those two numbers are XY coordinates. So all of the chromatograms can be placed on a graph or map. If the chromatograms are similar to one another, they will end up similar to one another on the map.
When the researchers did this, the wines from the same estate but different vintages clustered together. For a different estate, you get another cluster somewhere else. “Now we knew that there was something common in the chemical fingerprint for a given estate across the vintages,” said Pouget.
Then he noticed that using the XY coordinates to map, three wines or three clusters were on one side and four were on the other. “Stéphanie had not told me what those wines were. When I saw that, I thought—oh my goodness, did she give me Right and Left Bank?” said Pouget. “I’ll never forget when we had the Zoom call. I showed her the figure and I saw her jaw drop. I said, Stéphanie, is this Right Bank and Left Bank?”
Indeed, they were. The clusters reflected the geography of Bordeaux’s appellations. This means that, chemically speaking, there is similarity between wines that are geographically closer to one another. “Visually, it’s stunning,” said Pouget. “You take a chemical fingerprint of a wine and project it onto a piece of paper and you recover a map of Bordeaux—which is amazing stuff. This is where I got super excited.”
Potentially this means the researchers could take a wine that was not part of the study and see if it appears on the coordinate map close to other wines from the same appellation. The results have encouraged the team to broaden the scope of their inquiry.
“We’ve already begun gathering samples of Pinot Noir from around the world,” said Marchand. She’s also curious to turn her research towards Syrah. And in Bordeaux, Marchand would like to broaden the research to include wines from modest estates sold at consumer-friendly prices. As gas chromatography has been used by wine scientists for decades, the available database is extensive.
Fighting Fraud and Adapting to a Changing Climate
The research also has possible uses for law enforcement and winemakers adapting to climate change.
Being able to accurately identify a wine’s producer could help fight counterfeiting. “The criminals have powerful tools. We need to have them too,” said Marchand. Scientists at the ISVV already work with French investigators.
In terms of climate change, we know that changing weather patterns impact the chemical makeup of a wine. This new combination of machine learning and chemical analysis is capable of showing a winemaker how the typicity of their wine, as defined by the chemical fingerprint, is changing.
Marchand isn’t worried about a machine replacing a winemaker, but believes it can inform winemakers in the choices they are making and how to adapt to climate change if they want to maintain their wine’s historic identity.
Other possible applications for this breakthrough in wine chemistry and machine learning are still being determined. “We’re at the basic science stage right now,” said Pouget. “To turn it into practical application, I’m saying two to four years of work.”
And it all started with a wine-loving neuroscientist’s desire to unravel some of the mysteries of wine. “And so here I am, 30 years later,” said Pouget.
Stay on top of important wine stories with Wine Spectator’s free Breaking News Alerts.