Computational Optimization of Chemical Vapor Detector Arrays

Abstract

Arrays of broadly responsive, chemically sensitive detectors have been used for many years as a means of detecting a wide array of vapors. These systems have been used in fields ranging from analysis of wines and coffees to land mine and nerve agent detection to disease diagnosis. Despite their successes, these systems have been plagued by problems, namely a lack of sensor diversity, miniscule libraries of previously-recognized analytes, significant sensor drift, and weak signal processing capabilities compared to the mammalian olfactory system. This work details progress toward the alleviation of those problems with regard to arrays of polymer/carbon-black composite chemiresistor detectors developed at Caltech. Specifically, it has been determined that larger sensor arrays allow the suitable recognition of more analytes, and a greater chance of successful discrimination between a given analyte others to which it is similar. Additionally, new classes of percolative, low carbon-black sensors have been developed that yield far higher sensitivities and stronger responses than traditional sensors, as well as responses that are exponential with concentration. Such sensors allow for recognition of analytes using lower precision electronics than was previously realizable. A method for calibrating the system with few analyte exposures has also been developed from an analysis of the correlations between sensor/analyte response changes with time over groups of analytes and sensors. Further work has allowed algorithmic optimizations to assign functional group identities and certain physiochemical information such as molar volume and octanol/water partition coefficients to analytes that are completely unknown to the system, using a model built upon other known analytes. Additionally, a comparison of linear and nonlinear classifiers is performed to identify data characteristics that might be more suited to linear classifiers such as Fisher's Linear Discriminant or nonlinear ones such as Artificial Neural Networks. These improvements to chemical vapor detector arrays and the processing of their data allow the extraction of more useful information and the minimization of time spent training and calibrating the system. By constructing more appropriate sensor arrays, establishing non-comprehensive but extensive analyte response libraries, choosing useful algorithmic classifiers, and performing timely and mimimal calibration, the utility of detector systems can be maximized while minimizing maintenance.