From smell detection to smell generation
In 'Creating a stink in the name of science,' BBC News looks at the odor recorder developed in Japan and which was widely publicized last July. The device uses about 15 sensors to analyze the smell of an object and about 100 chemicals to reproduce it. So far, the system, which learns by using a neural network, can only recognize some fruits, such as apples, oranges or bananas. When it can identify more scents, it could be used in a large number of applications, from the fragrance industry to online shopping, and from computer games to movies and television. But read more...
Let's start with the current limitations of the prototype.
At the moment the prototype can only reproduce certain fragrances including apples, bananas, oranges and lemons. But synthesising billions of different smells is still problematic, admits [Takamichi Nakamoto, professor at the School of Engineering at the Tokyo Institute of Technology.] There is not a small number of "primary" odours from which all others can be created.
So how this odor recorder work? As you can see in the figure below, the odor recorder device has three components: an array of quartz crystal microbalance (QCM) sensors, the odor blender itself and a PC to control the whole system. (Credit: Takamichi Nakamoto's laboratory)
For additional information, you can visit Takamichi Nakamoto's laboratory and read these two pages about the odor recorder and the experimental setup.
Here are more details about the principle of the odor recorder (see reference below).
First, the target odor to be recorded is introduced into a sensor array and its output pattern is memorized. Then, the responses of the sensors to the blended odor, made up of multiple component odors, are measured and are compared with those to the target odor. The recipe of the target odor is obtained from that of the blended odor in the case that the sensor-array output pattern of the blended odor agrees with that of the target odor. Otherwise, the recipe of the blended odor is iteratively modified so that the sensor-array output pattern of the blended odor can approach that of the target odor using adaptive MIMO (multi-input multi-output) feedback control theory. The recipe of the target odor is obtained after the convergence.
Let's now return to BBC News to discover how the device 'learns' about various odors.
The device does not necessarily mix the right recipe on its first attempt. Instead it learns the correct ingredients list by comparing the analysis of the target scent with more and more refined concoctions. The machine "learns" recipes using a neural network, a collection of computer processors which function in a similar way to a simple animal brain.
Anticipating future demands, Nakamoto and his team have already designed a computer game.
It is a relatively simple cooking game, where users wearing a mask add virtual ingredients to an onscreen frying pan to rustle up a Japanese curry. As you add the butter and onions, realistic scents are pumped through the mask. In turn, garlic, meat and spice aromas complete the overall dish.
This looks as a funny game. But if you want more serious information about this research work, you can read an article published by Chemical Senses under the name "Study of Odor Recorder for Dynamical Change of Odor" (Volume 30, Supplement 1, Pages 254-255). Here are two links to the abstract and to the full paper (PDF format, 2 pages, 224 KB). The quote about the principle of the odor recorder displayed above comes from this article.
Sources: Jonathan Fildes, BBC News, October 13, 2006; and various websites
You'll find related stories by following the links below.