Joshua L. Hertz, Baranidharan Raman, Kurt D. Benkstein and Steve Semancik



Arrays of chemical sensors, using a variety of transduction techniques, have been extensively studied because of the data-rich signals they produce.  The unique response patterns acquired from such arrays can be considered as chemical fingerprints to particular analytes in order to create an “electronic nose.”  Typically, the fingerprints are first determined by training a sensor (i.e., identifying unique aspects of measurements taken during exposure to known analytes).  In the ideal case, a sensor is trained on all possible analytes that the sensor will encounter in its lifetime.  Unfortunately, such a training protocol is not feasible in most cases, since the set of chemicals that will be encountered is not fully known.  Therefore, a sensor should not only identify known analytes but also possess the capability to predict overlapping characteristics of unknown gases.  Furthermore, for practical use, pre-trained recognition methods must be able to cope with variability in sensor response due to various sources of aging.  Such aging potentially leads to misalignment between sensor measurements during operation and the chemical fingerprints registered during training, thereby impeding recognition of analytes.


Correct recognition of a specific analyte requires sensors to detect, in some way, an aspect of molecular features unique to that analyte. On the other hand, generalization to unknown chemical species requires detection of features that are common across a desired class of analytes. The opposing nature of these constraints suggests that achieving both capabilities in a single device will require a multi-step chemical identity resolution process.  This is in opposition to the conventional one-step (“all-at-once”) procedure that has been used in most sensing studies to date.  The inspiration for a multi-step approach is the processing known to occur in biological systems.  There, the combinatorial input from a large population of olfactory receptors is transformed by neurological signal processing such that initially coarse odor representation is increasingly refined over time to become more odor-specific.


In a chemical sensor, this new approach to odor recognition initially discriminates between broad chemical classes and then subsequently uses additional data for finer discrimination of sub-classes and, eventually, specific compositions.  Here, we apply such a hierarchical, bio-inspired approach to a temperature-controlled chemiresistor microarray and demonstrate that not only is a small set of training analytes sufficient to allow generalization to novel chemicals, but also that the scheme can be easily adapted to provide robust categorization despite aging.  We also show that the viability of the technique arises from the ability to find repeatable chemical interactions between a device and specific compositional features of target analytes.