Decision making in spectroscopy

Kevin D Judge, University of Rhode Island


Vibrational spectroscopy provides a great deal of information regarding the chemical properties of a substance; however spectral features associated with different chemical substances are not always obvious. Through the aid of advanced computer algorithms, pattern recognition techniques and chemometric models, complex and subtle relationships between a material and its spectrum can be identified and utilized to draw useful conclusions. The ability of computers to make inferences from spectral information was demonstrated in several applications. ^ Spectral imaging provides a substantial amount of data, with each pixel of an image representing a spectrum. For images of chemical mixtures, spectra of pure components can be discovered through the use of chemometric tools. Using such methods, spectral images can be exploited to obtain chemical maps, indicating the spatial locations of certain chemical substances. Artificial differences in the form of noise and background interference make it difficult to accurately compare the contents of each pixel. Background removal and noise reduction techniques minimized these undesirable effects, increasing the clarity of the image and each individual spectrum. This was shown using several images of biological samples, including certain types of bacteria. ^ Bacillus anthracis, a specific type of bacteria, has been a recent concern as a result of Anthrax and the threat of bioterrorism. Not surprisingly, bacteria samples of other species of Bacillus have similar chemical properties as is evident in their infrared spectra. Spectroscopy allows for the analysis of the total cellular composition of bacterial samples. The power to distinguish between relatively less harmful genera and species of bacteria suggests the potential to detect Anthracis and other hazardous biological substances. Classification on the genus and species level was demonstrated using statistical models and artificial neural networks. ^ Artificial neural networks (ANN) are models that mimic neurons in the brain. They have the capability to realize complex, nonlinear relationships. A major disadvantage to ANNs is that the reasoning behind their results is not intuitive, giving it a "black box" quality. ANNs were employed to act as expert systems in the interpretation of both infrared and Raman spectra. Spectral features correspond to the presence of different bonds in a chemical's structure. Further analysis of the ANNs resulted in an explanation to what was learned, creating factors that illustrated the characteristic spectral trends associated with different functional groups. Bayes theorem was applied to calculate a probability for each decision. ^

Subject Area

Chemistry, Analytical

Recommended Citation

Kevin D Judge, "Decision making in spectroscopy" (2007). Dissertations and Master's Theses (Campus Access). Paper AAI3276987.