by Bob Holmes
One of the first really promising steps in this direction came a decade ago, when world-famous chef Heston Blumenthal of the Fat Duck restaurant was experimenting with salty ingredients in desserts, and discovered that white chocolate and caviar make a great flavor combination. The pairing was so bizarre—yet so delicious—that Blumenthal mentioned it to a colleague at an industrial flavor company. A little work soon revealed that both members of this unlikely pairing are rich in a compound called trimethylamine, which has a fishy flavor.
This got Blumenthal thinking. If a shared flavor molecule accounts for the success of this odd pairing, maybe similar “molecular rhymes” might point us to other surprising flavor combinations. The idea makes some intuitive sense. As we’ve seen, chefs often balance similarity and contrast in building their dishes—and since flavor is all about molecules, then similar flavors should share flavor molecules. As Blumenthal pursued these molecular similarities, he came up with a whole kitchenful of brilliant, unexpected matches: liver and jasmine, which share sulfur compounds; carrots and violets, which share a molecule called ionone; pineapple and blue cheese; snails and beets.
In the years that followed, Blumenthal’s insight sparked a whole gastronomic movement. Going by the name of “food pairing,” it makes these molecular rhymes the centerpiece of combining foods. There’s even a commercial service (foodpairing .com) that, for the price of a monthly subscription, will let professional chefs and enthusiastic amateurs start with any ingredient and follow a web of molecular similarities to find other foods with supposedly complementary flavors.
A Canadian sommelier named François Chartier has begun to investigate the pairing of wine with food along the same lines, based on chemical similarities between ingredients in the food and aroma compounds in the wine. For instance, Chartier suggests pairing a rosemary-scented lamb stew with a dry Riesling wine, to take advantage of the citrusy, floral-smelling molecules in that wine, which echo those in the rosemary. This “molecular sommellerie” was novel enough to earn Chartier’s book on the subject, Taste Buds and Molecules, a prize as the “best innovative food book in the world” at the 2010 Gourmand World Cookbook Awards.
You’d think, with all this excitement, that food scientists would be eager to sink their teeth into molecular food pairing to see if it really works. But very few have actually done so—and even fewer have published their results in the scientific literature. (Foodpairing, Inc., the company selling food-pairing ideas to chefs, hasn’t released any evidence to back up its approach.)
The obvious test would be to have people rate how well pairs of ingredients go together, and see whether those that share more flavor molecules receive higher ratings. Wender Bredie, a Danish food scientist at the University of Copenhagen, did exactly that a few years ago, using fifty-three different pairs of ingre-dients ranging from cinnamon and apple to cinnamon and garlic, malt and cocoa to malt and blue cheese. They found that the number of flavor molecules in common made absolutely no difference to the rated pleasantness. “I have never seen an experimental study with a correlation that was so low,” recalls Bredie. (It’s worth noting that Bredie’s study—like an earlier study by a different group that reached much the same result—was only presented at a scientific conference, not published in a scientific journal. That means the research has not been vetted by other experts, so the conclusions should be regarded as preliminary.)
Bredie’s group did find one interesting result, though. Pairs with fewer molecules in common tended to be perceived as more novel than those that shared more molecules—a feature at least a few high-end chefs might like to make use of, though novelty is not the same as pleasantness. “With high-end restaurants, you want something unique and surprising for your customers,” says Bredie. “And it doesn’t have to be nice. If you go to Noma”—the New Nordic restaurant in Copenhagen that was rated number one in the world for several years—“the foods aren’t very nice. You go there, and you have a fantastic experience. But if you ask the customers, ‘Is this something you’d really like to eat more frequently?’ they’d probably say no.” (When Bredie says that Noma’s dishes “aren’t very nice,” he means that the ingredients and techniques are often unusual and challenging, such as the moss cooked in chocolate that is featured on the menu as I write this.)
A second approach would be to look at the flavor combinations that people actually use, and count the molecules their ingredients have in common. If these real combinations are more likely to share flavor molecules than random sets of ingredients are, that would be evidence that molecular rhymes really do make combinations taste better. The data are out there: the Internet age has provided a huge treasure trove of real flavor combinations, in the form of online recipes, and anyone with a few hundred dollars to spare can subscribe to a database that lists all the flavor compounds in any given food ingredient. The big challenge is making sense of the tangled web of recipes, ingredients, and flavor molecules.
Enter Sebastian Ahnert. By day a theoretical physicist at the University of Cambridge and by night an enthusiastic amateur cook, Ahnert has exactly the skill set necessary to pick apart the problem. A few years ago, he and his colleagues downloaded more than fifty-six thousand recipes from three online recipe archives (Epicurious, Allrecipes, and a Korean database called Menupan) and studied their molecular overlaps. Real recipes, they found, had a slight tendency to share more flavor molecules in common than random collections of ingredients—but only for North American, Western European, and Latin American cuisines, where common ingredients like milk, eggs, butter, and wheat share overlapping flavor profiles. Asian recipes actually shared fewer molecules than random ingredients, because their most common ingredients, such as soy sauce, scallion, ginger, and rice, have largely nonoverlapping flavors. When Ahnert left those most common ingredients out of his analysis, he found no evidence at all to support the food-pairing hypothesis.
Ahnert’s analysis made a big splash when it was published in a topnotch scientific journal, but he wasn’t satisfied. Recipes aren’t an ideal starting point, because some ingredients—flour and eggs come to mind—are often included more for structural reasons than because they contribute important flavors. So Ahnert went back to the drawing board. This time, instead of recipes, he used ingredient pairings recommended by well-known chefs, which he found in a best-selling book called The Flavor Bible by Karen Page and Andrew Dornenburg. He found that chef-recommended pairs share more flavor molecules than randomly paired ingredients—and the pattern gets stronger when he counts only the most abundant flavor molecules, or those with the most food-related odors.
So maybe there’s something to food-pairing theory after all. Not everyone is convinced yet, particularly because as I write this, Ahnert has not yet published his most recent reanalysis. But even if foods that go well together do tend to share flavor molecules, that’s not the same as saying that foods that share flavor molecules necessarily go well together. This molecular-rhyming approach is probably an idea generator at best.
To explore a really high-tech approach to finding unusual, exciting flavor combinations, I visited IBM’s Thomas J. Watson Research Center in Yorktown Heights, New York. The company has a long history of taking on the biggest challenges in artificial intelligence, and the Watson Research Center is where it all happens. IBM’s Deep Blue supercomputer made headlines back in 1997 when it beat world chess champion Garry Kasparov in a six-game match. Then, in 2011, Deep Blue’s successor, Watson (named, like the research center itself, after the company’s longtime president from the first half of the twentieth century), beat two human champions at the quiz game Jeopardy. With Watson’s win under their belts, IBM’s researchers started looking for new applications for their expertise. Why not, they asked, turn Watson’s immense powers to the kitchen? After all, cooking is both highly creative and a familiar, everyday activity that millions of people do regularly. And Watson’s computing prowess should pay off big time, because the computer could learn more recipes, mo
re ingredients, more techniques than any human ever could, just like it had for Jeopardy trivia. Let’s do it, the Watson team decided.
The Thomas J. Watson Research Center sits in forested hills just off the Taconic State Parkway, less than an hour’s drive north of midtown Manhattan. The main building—a vast, three-story curving facade, designed by renowned architect Eero Saarinen—looms over visitor parking, and the main entrance, sheltered by a thrusting, flaring overhang, leads into a 1960s-futuristic foyer. It’s all very high concept, very expensive looking, and very formal. Just what you’d expect from IBM, the corporation long notorious for its severely conservative dress code.
In such a setting, Florian Pinel comes as a distinct shock. What you first notice about the French-born software engineer is not his broad face, blue eyes, or unruly mop of stringy brown hair. It’s the four stainless-steel studs that pierce the corners of his mouth and lower lip, and the longer, blade-shaped fin that emerges from the center of his lower lip, just above his chin. Dressed in jeans and a ratty shirt, instead of IBM’s traditional white shirt and tie, Pinel looks more at home in a restaurant kitchen than in an IBM conference room.
Appearances aren’t deceiving in this case—Pinel is indeed comfortable in the kitchen. While working at IBM, he spent his weekends studying at New York’s respected Institute of Culinary Education, earning a chef’s ticket in 2005. For a while after that, he worked Saturday nights as a line cook in a Manhattan restaurant, just for the thrill of it. “That was a big rush,” he recalls. Eventually, though, he gave that up and focused on cooking at home in his newfound leisure time. When Watson came along, he was ready.
How do you teach a computer how to cook? Not the way you’d teach your kid, by having him or her stand by your elbow and watch. And not the way Pinel learned, in culinary school. Instead, you feed the computer data. Lots and lots of data. Food chemists have identified the key flavor chemicals in most ingredients, and psychologists have measured how pleasing we find each of them. Cyberspace is full of recipes that show how people all over the world cook: which ingredients they use, and how they combine them. Pinel and his team input all this information into Chef Watson’s memory bank. From this mass of data, the computer chef extracted patterns: the sets of ingredients that were likely to go well together, and the sequence of steps to use in combining those ingredients. (A big help for the latter was getting access to Bon Appétit magazine’s archive of nine-thousand-plus recipes, all tested and carefully edited into a standard format.)
Anyone with a computer can consult Chef Watson for free (at least as of this writing) at ibmchefwatson.com. You simply type in an ingredient or two, and Chef Watson suggests a few other ingredients you might try. Once you’ve settled on a core set of four ingredients—and, optionally, specified a style such as French, summer, or vegetarian—you can choose from a list of suggested recipes, complete with measurements and cooking techniques. It’s that simple.
Behind the scenes, though, a lot is going on. To come up with its recommendations for ingredients that work together, Chef Watson looks for ingredients that are already used together in existing recipes somewhere in the world, or sets of ingredients that share several flavor chemicals like Heston Blumenthal’s white chocolate and caviar. But for Pinel, finding these sets of ingredients alone isn’t enough to make Watson truly creative. “We think there are two things that make something creative,” he says. “It has to be novel, and it has to be valuable.” For a recipe, “value” equals deliciousness—something that Chef Watson could estimate from its knowledge of which flavor chemicals people like best, and from its calculations of chemical overlaps. And novelty was an easy one—Watson just calculated how similar a recipe’s ingredients are to those used in other recipes. Tomatoes, garlic, oregano? Not very novel. Asparagus, pig’s feet, and Indian spices? You bet. For each ingredient combination, Chef Watson gives you a “synergy” score—essentially a composite of compatibility, pleasantness, and surprise. A high synergy score means Watson is confident of its choice of ingredients, says Pinel. “This is going to work well, and it’s not going to be trivial, either.”
For a foodie, software like this is endlessly intriguing, and it’s easy to get sucked down the rabbit hole, browsing one idea after another. But it doesn’t take much experimentation to realize that for all of Chef Watson’s knowledge and computing power, it lacks the finely honed kitchen instincts of a Michelin-starred chef. Instead, it’s more like your brilliant-but-loopy buddy who blurts out whatever thought happens to pass through his mind, no matter how bizarre. I tested Chef Watson in late January, right around Robbie Burns Day, the great Scottish haggis-and-whisky fest. Since the traditional accompaniment, “neeps and tatties” (turnips and potatoes, for the uninitiated), is traditionally boring, I thought I’d see whether Chef Watson had any better ideas. I started with the key ingredient, turnips, and specified that I wanted a Scottish recipe. Then—to continue the Scottish theme—I added another of that country’s favorite foods: beer.
The chef’s suggestion: “Scottish Turnip Meatball,” a veal/turkey meatball served in a sauce of chili powder, the Indian spice mix called garam masala, turnip, avocado, and clam juice. It sounds like a bizarre jumble, and I almost didn’t try it. But in the spirit of research, I finally subjected my family to it one night—and, to our surprise, it worked pretty well. The creamy unctuousness of the avocado made just the right counterpoint to the turnip’s bitterness, and the garam masala and clam juice added a subtle depth to the flavor profile. In fact, it was good enough that we served it to dinner guests a few weeks later. Maybe there really is something to this whole molecular food-pairing thing.
Want something to cook for your Super Bowl party? Just specify “Super Bowl” in the “Pick a Style” field, and Chef Watson suggests some ingredients to start with: raisins, garlic, chocolate chips, and endive. The synergy score is off the charts, well over 90 percent, so for some reason Chef thinks this is a great combination. I think I’ll try another spin of the wheel.
Watson’s next suggestion: pork belly, shallot, ginger, and white pepper. Ah, more promising. One of the suggested recipes for that combination is Superbowl Pork Belly Bolognese, which sounds plausible. But the recipe itself is over the top, calling for ground chicken breast, ground pork belly, and ground chicken wings (bones in, I wonder?), as well as chorizo sausage and a quarter cup of horseradish.
To drink? Why not a Cauliflower Bloody Mary, made with Pernod and ouzo, not vodka or gin, and swapping out the seasoned tomato juice for a puree of cauliflower, shiitake mushrooms, and onion. Garnish with grape wedges, Chef suggests—a peculiarly fussy take on the usual lime wedge, and certainly one I wouldn’t have thought of on my own. (Weird recipe steps like that show up frequently, because Watson bases its procedures on existing Bon Appétit recipes, substituting similar ingredients as needed. In this case, probably, Watson decided to substitute grapes for lime because both are fruits with plenty of acid, and it just borrowed the technique verbatim. That likely also explains the ground chicken wings in the pork belly bolognese. You’ll soon find your own humorous combinations.)
I’m poking fun here, because Chef Watson leaves some low-hanging fruit. But even its goofiest ideas often have a kernel of real inspiration. When I started with Italian sausage and broccoli, it suggested a recipe adapted from a braised brisket dish: rub the sausage with a seasoning mix, “working the paste into all the cracks,” then put the sausage in a casserole “fat side up.” Clearly Watson doesn’t understand the difference between a brisket and a sausage. But that night as I lay in bed, I realized that a spice rub might add a nice touch to something as prosaic as a grilled bratwurst, or even a hot dog. Good idea, Chef. And maybe, with the right drink, a garnish of grape wedges wouldn’t be such a bad idea after all.
The jury’s still out on whether Chef Watson is a major step forward in culinary creativity or just an amusing sideline. So far, the app is generating about fifty thousand ingredient pairings a month, says Pinel—which so
unds like a lot until I realize that I’ve probably done fifty today, all by myself. Some users just look for suggestions of ingredient pairs, and build or adapt their own recipes; others click on complete recipes. The next step, says Pinel, is to add nutritional information to the mix, so that Chef Watson can double as Dietician Watson.
With menu chosen and kitchen chemistry properly deployed to marshal just the right flavor molecules, there’s one more step that a savvy cook can take to boost the flavor of a meal: serve it properly. We’ve seen that presentation can legitimately be considered as part of the flavor experience: Changing the color, shape, or weight of the plate or bowl can make the food taste sweeter or more bitter. Charles Spence, the psychologist behind that study, took that notion still further in another experiment. Working with Charles Michel, a professional chef, Spence presented volunteers with a salad containing identical ingredients plated in one of three ways. Some diners got an ordinary tossed salad, some got a salad with each ingredient stacked neatly in separate piles, and some got a salad that was dramatically plated in a splash of colors and shapes that resembled a Wassily Kandinsky painting. The diners who got the Kandinsky salad found it both more pleasing aesthetically and also tastier than those who got the boring versions. For any cook, at home or in a restaurant, an attractive presentation is more than just window dressing—it makes the food itself more flavorful.
We can apply the same principle to wine: drinking it from an elegant glass should make it taste better. There are functional reasons for this as well as the psychological ones. A large, tulip-shaped glass that tapers inward toward the lip offers more space for volatile molecules to gather above the liquid, enhancing the aroma. Studies have verified that the same wine does indeed taste more pleasant from a glass like this than from a straight-sided water glass. On the other hand, there’s little hard evidence that you get any extra flavor boost from having a different glass shape for Bordeaux-style wines than for Burgundies, as some high-end crystal makers suggest. I recommend spending your money on the wine, instead. (I asked one expert in the oenology department at the University of California, Davis, who’s done some of the wine glass studies, what kind of wine glasses she uses. “Whatever I get free from the winery,” she replied, with a laugh.)