Scientific endeavors rarely proceed without unforeseen complications. Before Erik Severin took his baby blue Mazda glc wagon down to Monterey Park, for example, he inquired about fish at his local Ralph’s supermarket. He was told that if his experiment required something that would rot with a pungency worthy of the cliché, he would do best with a saltwater fish, which Ralph’s did not have in stock. So Severin, a graduate student at Caltech, drove to Monterey Park, where he knew of a Chinese grocery by the unlikely, albeit appropriately Californian, name of Shun Fat. From bins of saltwater fish he chose a small kingfish, paid 69 cents, and brought it back to the Noyes Laboratory at Caltech.
Severin installed the kingfish in a glass jar with two glass tubes attached. He arranged for a stream of air to pass through one tube, circulate around the kingfish, absorb its aroma, and proceed out the other tube, over what is technically referred to as an array-based sensor apparatus, although Severin and his labmates call it the prototype electronic nose, or sometimes the noselet. I set the fish on a heating element and let it rot for days, says Severin. Fortunately, his sense of smell is not as keen as some. People would come in, their eyes would start tearing up, and they’d say, ‘My God, this is awful.’
The experiment’s denouement, according to Severin, was a qualified success: the electronic nose could easily distinguish between kingfish fresh and kingfish rancid, but not between, say, two-day rot and three-day rot. Still, Severin and his colleagues have high hopes for the nose. Since their noselet experiments, they’ve built a more sophisticated smelling device--the supernose. Using computerized switches, the supernose can automatically control details like the concentration of the aromas, how fast they flow by, and which odor follows which. I used to have to stick close to the noselet, changing the jars, adjusting valves, and so on, says Severin. I had trouble finding five minutes to run to the bathroom. Now, with the supernose, we set it up, press a button, and leave the nose running for days. Someday, the researchers expect, their supernose will be able to tell a merlot from a Beaujolais, a Heineken from a Budweiser, or anything that human noses, or even dog noses, can do effortlessly.
Until half a dozen years ago, such an olfactory accomplishment seemed safely beyond the reach of technology. With the Caltech achievement, though, the science of electronic noses appears to be in for a paradigm shift. The supernose points the way to an absurdly inexpensive nose on a chip--an electronic sniffer that will cost no more than the average kingfish. Such noses will serve as sensors and aromatic arbiters in uses so widespread that they may someday do for odors what the computer chip has done for pretty much everything else.
The Caltech nose is the change-of-life child of a chemist named Nate Lewis, who became famous in his field not just for the preternaturally early development of his chemical insight, but also for his ability to bring to bear on any single project a merger of disparate scientific talents, effecting a kind of cross-pollination of ideas and then publishing the final word on that particular research. His reputation attracted the brightest graduate students and postdocs in the country, whom he would unleash on problems of his choosing, further enhancing his reputation.
By 1992, however, Lewis was showing signs of a midlife crisis. He had spent the first 15 years of his career, as he puts it, trying to coax semiconductors into acting like artificial leaves to produce stored chemical energy from sunlight. Now he was contemplating 30 more years spent doing the exact same thing. He was ready to try something new.
Lewis had his epiphany walking along a beach in Ventura, California, after a conference. I was thinking, he says, that we know how people touch; we know how people see; we know how they hear. We don’t really know anything about how they taste or smell. That’s pretty interesting. And in science it’s good to choose a problem hardly anyone knows anything about.
The olfactory system in humans is still the subject of much debate. Simply put, what we smell is nothing more than stray molecules of substances, known as odorants, that waft through the air and settle into molecular receptors inside our nostrils. The odorants induce an electric signal in the receptors that is carried through neurons into the olfactory bulb, which sends signals into a brain structure known as the olfactory cortex, which passes them along until they eventually reach the hippocampus, a primitive brain structure associated with memory in mammals. (That association may be why the smell of lilacs on a spring morning can evoke with such passionate insistence the uncomplicated days of youth.)
All of this, however, goes far beyond the problem that Lewis, or anyone else hoping to model the nose, had to worry about. His main question was first electrochemical and then, perhaps, computational: How do odorants induce an electric signal in the neurons in the nose itself, and how is that signal decoded?
A typical approach of chemistry to detecting and identifying odorants would be to build what are known as lock-and-key receptors. Odorant molecules would fit into receptors like keys into locks, with each molecule specific to its own receptor and vice versa. But because a mind- bogglingly vast variety of molecules stimulate our nostrils, such an approach would be hard for researchers to implement. And it didn’t seem nature could work this way, either. It’s probably true, says Lewis, that male dog noses, for instance, have lock-and-key receptors for pheromones, which are sex attractants. But it’s unlikely that dogs develop lock-and-key receptors for cocaine when they’re trained to smell it out at airports. So how do they do that? Then there was the problem of complicated odors: What happens with a scent like Coca Cola’s, a blend of perhaps 100 different odors that add up to a scent similar to but distinct from Pepsi’s? Could a nose really be expected to have 100 different lock-and-key receptors for each variant odor in each brand of cola?
As Lewis figured it, the nose had to generate some olfactory signal, even for molecules it had never encountered before. The brain and nervous system might respond by realizing they were faced with a new smell entirely, or something that smelled like almonds, or brussels sprouts, or lilacs--new, but not entirely new.
The only way the nose could send such complex, coherent messages (or the only way Lewis could envision it doing so) was if it was engaged in a game of pattern generation and recognition. Receptors would not be specific but promiscuous, so to speak. Each receptor would respond to each odorant molecule with a signal. The signals would vary from receptor to receptor and from odorant to odorant, so each incoming odorant would generate its own particular pattern of signals from all the different receptors. This pattern would be transmitted to the pertinent parts of the brain, where the process of recognition would take over. (Aha, lilacs!) This scenario had the added benefit of assigning the learning problem to the brain, where it seemed to belong, rather than to the nose.
Now Lewis had only to find chemical substances that could be used as electrochemical sensors to mimic the promiscuous reactivity of the receptors. Each odorant would induce a reaction in each chemical sensor, and that reaction would induce a stronger or weaker electric signal. Taken together, a collection of these chemical sensors, each different, would transmit a pattern specific to whatever odorant had set them off.
Imagine, for example, an artificial nose made with four chemical sensors, confronted by a succession of odors. Fresh fish might induce a strong signal in sensors one and three, a weak signal in sensor four, and not much reaction at all in sensor two. Rotten fish might induce a similarly strong signal in sensor one, little or nothing in sensors three and four, and it might send sensor two off the charts. So the hypothetical nose would have little trouble telling fresh fish from rotten fish. Then along comes an expensive designer cologne, which coincidentally generates nearly the identical reaction in all four sensors as fresh fish (depending on the fish). Now the nose would do well to have a fifth or sixth sensor, so that the odds that the cologne would invoke a different pattern from fresh fish would be greatly increased. Given enough of these sensors, the electronic nose might generate a different pattern for every possible odorant, including those that had never come its way before.
What chemicals, wondered Lewis, should he use to make those sensors? For years he had been collaborating with a Caltech chemist named Bob Grubbs, an expert in materials technically known as organic conducting polymers, and less technically as plastics that conduct electricity. Unlike most metals, these plastics can be easily dissolved into a liquid form and then painted onto substances. They’re metallic, but they don’t react like metal, says Lewis. You can dissolve them and spin them wherever you want or paint them on.
Lewis figured he might be able to use these conducting polymers to make his electronic receptors. A polymer molecule is just a long chain of single molecules known as monomers--for instance, polyacetylene is a long chain of single acetylene molecules. In a conducting polymer, electrons move freely down these chains. When two chains abut, electrons can hop from one to the other. The fewer points there are at which the chains touch, the more resistance there is to an electric current.
If stray molecules--odorants, for instance--were to waft over the polymers, some would nestle in the gaps between polymer chains and thus be absorbed into the material the way water is absorbed into a sponge. Moreover, different molecules would be absorbed differently by different conducting polymers, just as a sponge will absorb a lot of water but only a little gasoline. The absorption would cause the polymer to swell like a sponge, pushing the chains apart and decreasing the number of spots where electrons could hop from one chain to another. This would show up as an increase in resistance to the flow of electricity. And that change in resistance could be measured easily. After a few seconds the odorants would float out of the gaps between polymer chains the way they had floated in.
The key to Lewis’s scheme was that a specific odorant would cause different conducting polymers to swell to different extents. Grubbs had already made a handful of conducting polymers for Lewis by chemically attaching different molecules to the backbone of the polymer. This wasn’t particularly easy to do, and the number of different conducting polymers they might eventually make was limited, but they provided what Lewis called four or five tweaks that would react in different ways to any passing odorant.
Chemists have some intuition in these things, says Lewis. We know like likes like, for instance; so we know something with lots of water in it will like to interact with something else that has water in it. And if it has lots of benzene in it, then it doesn’t like water, it likes benzene. Things in between are going to be in between. So we knew what our first four or five guesses were going to be. We want something that has a charge on it; something that has a benzene in it; something that has grease in it; something that has water in it. We can make the conducting polymers pretty different and see what’s going to happen. And maybe if we can make these polymers different enough, then if we give them a vapor, they’ll swell, and they’ll become more resistive and they’ll be differently more resistive, so we’ll get our pattern.
This was a brilliant idea, one of those eureka moments in science. It had a downside, however, as often happens with brilliant ideas. It had been done. One night, Lewis did a computer search of the scientific literature to discover whether any of his peers or predecessors had had similar thoughts. British scientists, he found, had already written books on electronic noses made of six or seven different conducting polymers. In fact, as he learned later, several firms were planning to sell them commercially, for upwards of $50,000 a nose.
So we didn’t have the invention of an electronic nose, says Lewis, and we didn’t have the first way to think about doing it from a resistor or electrical measurement. That was all out there. What they did have was Lewis, who still believed there was something interesting to be done with electronic noses, and his students, who had ideas of their own.
When Lewis had first returned from Ventura, he had approached Michael Freund, one of his postdocs, and said, Let’s build a nose. Freund began his nose project with the conducting polymers from Grubbs but was a little leery of them because they have an annoying sensitivity to air. In the real world, they fall apart and lose their conductivity within a few hours, leaving researchers no choice but to work with them in an airless environment, such as a vacuum chamber. So Freund made one sensor from Grubbs’s polyacetylene, just to see if it would work, and then considered a second kind of conducting polymer, known as polypyrrole. Polypyrrole did fine in air but didn’t come in multiple variations like the polymers Grubbs had already designed.
Then Freund had his eureka moment. All he had to do was take his polypyrrole and mix it with various different kinds of insulators, as if he were swirling different sauces--strawberry or butterscotch or fudge--into vanilla ice cream. Insulators are substances that resist the flow of electricity; like conducting polymers, they will absorb different chemicals at different rates. Mixing them into the polypyrroles was a simple way to mimic the different conducting polymers. Now he would have a single conducting polymer with different insulators swelling in it.
Freund put his idea to work in the very first noselet. After mixing his insulators with his polypyrrole, he would paint a little swath of each variation on a glass slide, connect electrodes to each side of the swath, run a current through, and then detect how the resistance changed when he passed vapors over it. His first experiments were with various solvents from around the lab--methanol, ethanol, acetone, benzene, tetrahydrofuran, and so forth. It was clear, says Freund, that they responded differently to different solvent vapors. Some of the sensors’ resistances would go up, some would go down, some would go down and up, or up and down. But if you looked at them all, you would see one pattern with ethanol, another with methanol.
Freund also showed various patterns to Lewis without telling him what they were. I very quickly trained Nate to distinguish between different solvents from their different patterns, he says, so I knew a computer would be able to do it.
Lewis, Californian that he was, wanted to see whether his incipient nose could distinguish between wines and other alcoholic beverages. Freund gave it a whirl, first with beer, but the bubbles from the head stymied him. Whenever he tried to pass air through the beer, it would bubble up, overflow the vessel, and clog his gas-flow pipes. Freund moved on to wines and liquors, which the noselet had little trouble telling apart, although it failed to differentiate between individual wines. And then Freund left to take a position at Lehigh University before he perfected the apparatus. Nate was talking about fish when I left, says Freund, so I’m glad I got out of there.
The Caltech nose still had a major evolutionary step to go through before it could change the face of artificial smell research. Lewis was talking to Grubbs about how to make the sensors even more variable. One of the two--neither remembers who--realized that despite all the work on conducting polymers, they didn’t really need them after all. Anything would work, says Lewis. All they had to do was mix any conductors with any insulators, and if the insulator would swell, then it would change the resistance differently to different odorants. By using nonconducting polymers for their insulators, they could even make the mixtures conveniently paintable.
You can take little particles of carbon as the conducting part, or little balls of gold, or little balls of silver, says Lewis. This time you can think of the insulators as different flavors of ice cream, with various conductors mixed in like nuts or chocolate chips. It can be almost any conductor in any insulator. So now you start thinking about how many different things can be made this way. Our first arrays now have 17 off- the-shelf polymers as insulators. But 17 probably isn’t enough. We want a million different sensors. Now we think we can do that. It’s not hard to envision a computer chip with these little wells, and you have the sensors in the wells, and a set of wires going in and out, and you measure the resistance. So reading a million resistances on a chip--no problem. The British had the idea right. We just took it much further by broadening what could be used as a sensing material.
Even 17 sensors, however, would allow for an unfathomably large number of possible patterns--too many to analyze easily. Or as Lewis says, Sometime after we made the polypyrrole nose, we realized we had more signals than we knew what to do with.
Fortunately, the problem was not hard to solve, because Lewis, Freund, and company were lucky enough to be working at a place like Caltech. They went to talk to Caltech biophysicist John Hopfield, the father of a system of computing known as a neural network. At the time, Hopfield was working on computer programs that might simulate what the brain does in processing signals from the olfactory bulb.
Hopfield had his researchers teach Lewis’s chemists how to run the necessary computer programs. Neural networks are artificial- intelligence programs wired rather like the interconnected neurons of human brains. Like people, they can learn as time goes by, so they can be trained to recognize patterns. The network would just sit there saying, ‘What does this smell like? What does that smell like?’ explains Lewis. It would take the input from all the different sensors, recognize how the patterns evoked by different odorants were similar and how they differed, and then record them for future use.
The more odors the network sniffed--the more patterns that came its way--the more it would recognize. Eventually it would learn which patterns are similar, suggesting that the smells were similar. You can train the software in a neural network to take the patterns and find the differences, says Lewis. You just need enough sensors sending enough signals so that no two things have the same pattern and seem to smell alike.
By early 1996, Lewis had created one of his signature teams to flesh out the electronic nose. It helped that he had access to Caltech’s remarkable intellectual resources. He set his electrochemists to work with biologists, neural network experts, computer scientists, chip designers, and even some physicians in Galveston, Texas, who participated through electronic mail.
Lewis has great dreams for his nose.
For starters, his nose on a chip should cost less than a dollar to manufacture and sell, which is a reasonable markdown from the $50,000 plus of existing electronic noses. That one dollar should buy the neural network hardware on a chip, the signal conditioning on a chip, and the million sensor dots. The Microdevices Laboratory at Caltech’s Jet Propulsion Laboratory (jpl) is currently building Lewis his miniature nasal device. Once we get the elements right, says Lewis confidently, the rest of it should not be very stressful.
With the nose in hand, uses should come by the snoutful. The world is already full of designer sensors, such as smoke detectors and breath analyzers, that look for a single chemical. These can be considered one-trick noses. The chemical in question reacts with another chemical on a film, prompting a reaction that changes either the electrical or optical properties of the film, and that in turn sets off an alarm. But a carbon monoxide sensor won’t smell a Freon leak from the refrigerator or a methane leak from the stove, and a breath analyzer won’t tell you if your house is on fire. Lewis’s nose, says Minoo Dastoor, who manages environmental and biomedical technology at jpl, is a fundamental technology advance. Instead of looking at the response of one specific chemical to one individual sensor, you’re looking at the response of one specific chemical to a whole array of sensors.
The Caltech nose can be trained to recognize what’s normal--the everyday smells, or what scientists would call the baseline--and then alert users to anything different. Lewis foresees people using his sensors for all kinds of sniff tests: Cadillac wants the leather in its various cars to smell the same. It doesn’t need to know what’s in the leather, it just wants it to smell like it did yesterday. Give this thing the pattern, and it will tell you whether the leather smells right. Cheese makers want their cheese to smell the same, too. Or perfume. Or brake fluid, which smells funky when it goes bad. Train the nose to recognize good brake fluid and it will signal you if it gets out of sorts. The same nose, trained differently, might smell a bad batch of gasoline in your fuel tank or warn people if they’re getting carbon monoxide inside the car.
Lewis and Dastoor have convinced nasa to fly the artificial nose on the space shuttle, probably in 1998. nasa is seriously interested, says Dastoor, in incorporating the nose in the life support systems of the planned space station. As Lewis says, These space stations seem to stink when something goes wrong. I hear the shuttle can smell like crazy, and the Russian space station, mir, is worse. They don’t know if the vapors are bad for astronauts or not. And besides, he says, the British already have a nose flying on mir, which I’m continually reminding nasa. This fact has not escaped me, and I hope it doesn’t escape nasa either.
What intrigues Lewis even more than the commercial applications is the science that can be done with his nose. In his search through Caltech for people who could help him understand the olfactory system, he met up with a computational neurobiologist named Jim Bower, and the two bonded instantly. Bower makes biologically realistic models of mammalian nervous systems and has a particular interest in the olfactory system.
I spent most of the last ten years, says Bower, trying to convince people that the olfactory system recognizes odors based on tremendous receptor complexity, that what you really want to do is not detect specific features that you know about to begin with, but to sample as broadly as possible. That the main objective of the nose is to generate some signal no matter what the stimulus. I spent ten years trying unsuccessfully to persuade people in my field to think about it this way. And then I have a lunchtime conversation with this chemist, and he’s already seen for himself that this is how it has to be.
Bower and Lewis are collaborating, using Lewis’s nose to make predictions about how humans and rats might smell. The idea is to measure when the nose has trouble distinguishing between two odors, and then seeing whether humans and rats have the same trouble. If the electronic nose turns out to be drastically different from the human or rat nose, the researchers can readjust it by retraining the neural network. If we can predict on the output of Nate’s nose what odors are hard to distinguish, says Bower, it means we’re starting to understand something about the complexity of the human olfactory recognition problem.
The experiments may also help Lewis understand what it is about a molecule that determines its odor. Suppose you could look at a molecule, he says, and compute anything you want: How wide is it? Where are its electrons? How will it swell a polymer? You still can’t say what it will smell like. But once the researchers start understanding the patterns generated by the molecules in the electronic nose, they can start comparing those with the various features of the molecules. They can find molecules with similar physical features and give them to humans, rats, and the electronic nose, to see whether different noses perceive them the same way.
The possibilities are endless, in part because of the ridiculous simplicity of the electronic nose once it reached its final incarnation. Indeed, Lewis says one of his friends, an astrophysicist from the University of California at Berkeley, insisted that Lewis’s vaunted nose was just a high school experiment. Lewis responded, I know, and I’m proud of that. Lewis and the astrophysicist, however, may even have been overestimating the complexity. It’s so easy, says Severin, that a junior high school student recently heard Nate lecture, and he went and did a science project on it.
