Here we examine what is commonly called a Raman image and discuss how it is rendered. We consider a Raman image to be a rendering
as a result of processing and interpreting the original hyperspectral data set. Building on the hyperspectral data rendering,
we demonstrate the use of Raman imaging for the characterization of thin-film and ion-implanted silicon (Si) structures. A
Raman image does not self-reveal solid-state structural effects and requires either the informed selection and assignation
of the as-acquired hyperspectral data set by the spectroscopist or the use of statistical software applications. High spectral
resolution allows us to clearly resolve the substrate Si Raman scattering from that of the polysilicon film. Consequently,
high spectral resolution allows us to strongly resolve (structurally, not spatially) or contrast the substrate Si and polysilicon
film in Raman images; that is, the spectral resolution contributes to the chemical or physical differentiation of spectrally
similar materials in thin-film structures.
The primary goals of this column installment are to first examine in more detail what is commonly called a "Raman image" and
to discuss how it is "rendered," and second to demonstrate the use of Raman imaging for the characterization of thin-film
and ion-implanted silicon (Si) structures. The Raman images of Si and their rendering from the hyperspectral data sets presented
here should encourage the reader to think more broadly of Raman imaging applications of semiconductors in electronic and other
devices, such as microelectromechanical systems (MEMS). Raman imaging is particularly useful for revealing the spatial heterogeneity
of solid-state structures in semiconductor devices. Here, test structures consisting of substrate silicon, silicon dioxide,
polycrystalline silicon, and ion-implanted silicon are analyzed by Raman imaging to characterize the solid-state structure
of the materials.
Before we begin our discussion on Raman imaging of silicon structures, we need to carefully consider and understand what actually
constitutes a "Raman image." Consider the black and white photograph, perhaps the most basic type of image with which we are
familiar. One can think of it as a three-coordinate system, with two spatial coordinates and one brightness (independent of
wavelength) coordinate. Progressing to a color photograph, we now have a four-coordinate system by adding a chromaticity coordinate
to the original three of a black and white image. That chromaticity coordinate in the color photograph is the key to thinking
about hyperspectral imaging in general and Raman imaging in particular. Whether color discrimination is due to the cone cells
in our eyes, the wavelength sensitive dyes in color photographic film, or the color filters fabricated onto charge-coupled
device (CCD) sensors, a wavelength selection is imposed on the image. That color discrimination may not faithfully render
a color image that corresponds to the true color of the object; for example, people who are color blind may not see the true
colors of an object. In our daily lives, it is the combination of shape (two spatial coordinates), brightness (intensity coordinate),
and color (chromaticity coordinate) that we interpret in the images we see to draw meaning and respond to the objects from
which they originate. In that sense, image interpretation comes naturally to us and we begin to do it from a very early age
without giving it much analytical thought.
When applying these same basic imaging principles of the four-coordinate system to hyperspectral imaging, the interpretation
of the spectral (chromaticity) and intensity (brightness) coordinates are no longer "natural" and require mathematical thinking
for proper interpretation. This mathematical thinking by the spectroscopist comes in the form of direct selection of spectral
band position, width, shape, and resolution from closely spaced bands. The intensity coordinate is most often interpreted
as the value at one particular peak or wavelength relative to that of another. However, even the simple application of an
intensity coordinate requires the removal of background light unrelated to the spectroscopic phenomenon of interest. Furthermore,
there are band-fitting operations and width and shape measurements that can be made and plotted against the two spatial coordinates
to render an image. One can also apply any number of statistical and chemometric tools found in most scientific software.
Therefore, whether the spectroscopist is directly engaged in the selection of spectral features and processing of the hyperspectral
data set or simply uses statistical software operations, the spectral image is still the rendered product from a group of
mathematical functions operating on the data.
Now let's apply these hyperspectral imaging considerations to Raman imaging. We are all too familiar with the fluorescent
background that often accompanies Raman scattering. In some instances, the spectroscopist will digitally subtract the fluorescent
background before rendering a spectrum that consists almost exclusively of Raman data. By analogy, we might expect the same
practice to hold if we wish to call an image a "Raman image" rather than a "spectral image" (one that includes data from all
light from the object). For example, if the signal strength in a hyperspectral image detected at a particular Raman shift
consists of 90% fluorescence and only 10% Raman scattering, it would not be correct to call that a Raman image. However, if
the fluorescent background and any contribution from a light source other than Raman scattering is first removed, then we
have a Raman image. Having done that, we are not necessarily finished rendering our Raman image. If we wish to spatially assign
chemical identity or another physical characteristic to the image, the data must be interpreted either through a spectroscopist's
understanding of chemistry and physics or through the statistical treatment of the data. The spectral image data as acquired
are not self-selecting or -interpreting. It is only after applying the spectroscopist's judgment or statistical tools that
one renders a color coded Raman image of compound A, compound B, and so on.
I hope that you can see (pun intended) that Raman imaging is not an entirely simple matter or as intuitive as photography.
Rather, it requires careful attention to the mathematical operation on the data and interpretation of the results. A Raman
image is rendered as a result of processing and interpreting the original hyperspectral data set. The proper interpretation
of the data to render a Raman image requires an understanding or at least some knowledge of the processes by which the image