Optical Filter Selection for High Confidence Discrimination of Strongly

Aug 4, 2015 - Optical filter-based chemical sensing techniques provide a new avenue to develop low-cost infrared sensors. These methods utilize multip...
0 downloads 0 Views 5MB Size
Download PDF
Article pubs.acs.org/ac

Optical Filter Selection for High Confidence Discrimination of Strongly Overlapping Infrared Chemical Spectra Kevin J. Major,*,† Menelaos K. Poutous,† Kenneth J. Ewing,‡ Kevin F. Dunnill,† Jasbinder S. Sanghera,‡ and Ishwar D. Aggarwal† †

Department of Physics and Optical Science, UNC Charlotte, Charlotte, North Carolina 28223, United States Optical Sciences Division, US Naval Research Laboratory, 4555 Overlook Avenue SE, Washington, DC 20375, United States

‡

ABSTRACT: Optical filter-based chemical sensing techniques provide a new avenue to develop low-cost infrared sensors. These methods utilize multiple infrared optical filters to selectively measure different response functions for various chemicals, dependent on each chemical’s infrared absorption. Rather than identifying distinct spectral features, which can then be used to determine the identity of a target chemical, optical filter-based approaches rely on measuring differences in the ensemble response between a given filter set and specific chemicals of interest. Therefore, the results of such methods are highly dependent on the original optical filter choice, which will dictate the selectivity, sensitivity, and stability of any filter-based sensing method. Recently, a method has been developed that utilizes unique detection vector operations defined by optical multifilter responses, to discriminate between volatile chemical vapors. This method, comparative-discrimination spectral detection (CDSD), is a technique which employs broadband optical filters to selectively discriminate between chemicals with highly overlapping infrared absorption spectra. CDSD has been shown to correctly distinguish between similar chemicals in the carbon−hydrogen stretch region of the infrared absorption spectra from 2800−3100 cm−1. A key challenge to this approach is how to determine which optical filter sets should be utilized to achieve the greatest discrimination between target chemicals. Previous studies used empirical approaches to select the optical filter set; however this is insufficient to determine the optimum selectivity between strongly overlapping chemical spectra. Here we present a numerical approach to systematically study the effects of filter positioning and bandwidth on a number of three-chemical systems. We describe how both the filter properties, as well as the chemicals in each set, affect the CDSD results and subsequent discrimination. These results demonstrate the importance of choosing the proper filter set and chemicals for comparative discrimination, in order to identify the target chemical of interest in the presence of closely matched chemical interferents. These findings are an integral step in the development of experimental prototype sensors, which will utilize CDSD.

H

Mammalian color vision utilizes three broadband overlapping filters in the eye cone cells. Each filter transmits power over a set spectral range to the brain, enabling discrimination of over 20 million different colors in a cluttered background.14 Biomimetic detection approaches that replicate human color vision have been reported for the discrimination of nerve agent simulants,15 and the detection of glucose in blood.16−18 Each of these systems employed the use of broadband overlapping optical filters operating in the near-infrared spectral region. Other filter-based detection approaches include programmable correlation radiometry19−21 and multivariate optical elements.22−25 Programmable correlation radiometry is a spectral comparative-radiometry technique which uses correlation spectroscopy and synthetic spectra as the basis for remote detection of chemical species. A strength of this technique is that it replaces the reference cell of a conventional correlation

igh-selectivity detection of toxic and energetic materials using spectroscopic techniques requires the generation of well resolved spectral data, both of the target materials and environmental background.1,2 Once acquired, the spectral data is evaluated using multivariate analytical methods such as principle component analysis,3−8 partial least-squares regression analysis,9−11 or linear discriminant analysis12,13 to detect spectral features indicating the presence of the target chemical. Successful detection requires that the spectral data is of sufficient resolution and bandwidth, such that the identification algorithm can separate and identify the target chemical spectral signatures. As the degree of overlap between the target chemical and environmental background spectral signatures increases, spectral data with greater resolution are required to enable high confidence detection. This approach has limited utility due to intrinsic infrared line broadening at ambient temperatures, as well as increased complexity and cost in highresolution spectroscopic systems. A different approach is to develop sensors based on biological systems that are capable of discriminating between closely overlapping spectral bands. © XXXX American Chemical Society

Received: May 6, 2015 Accepted: August 4, 2015

A

DOI: 10.1021/acs.analchem.5b01723 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

Figure 1. Representative FTIR absorption spectra for hexanes (blue), heptane (red), pentane (green), and acetone (black) in the C−H stretch region (2800−3100 cm−1). The dashed lines indicate the chosen initial filter positions (2879, 2939, 2965 cm−1), which dictated the separation between the filter peaks for the present study. Each spectra is normalized to the maximum absorption value for that chemical within this wavenumber range.

spectrometer with a programmable diffractive optical element, whose function is to produce the desired spectrum or to modulate the incoming light to spectrally discriminate between different species absorption. This method is capable of spectral detection of a given species, provided it has the “programmable” spectrum available. Multivariate Optical Elements use multilayer interference filters whose transmission spectra represent the features of a spectral regression vector within a given spectral range. Strengths of this approach are the ability to generate compound specific filters enabling a low cost and small footprint for a detector based on this technology. Neither approach has been extended into the mid-infrared wavelength range where more selective and intense absorption bands are available for high selectivity detection of chemicals. Recently, a unique method describing the interaction of a biomimetic three filter mid-infrared system with the fundamental C−H absorption bands (2800−3100 cm−1) for different hydrocarbon vapors was reported.26−28 The method, comparative-discrimination spectral detection (CDSD), utilizes higher-dimension operations between vectors, generated by the change in power through each filter set as the concentration of the chemical changes, to develop a unique identifier for a particular chemical, by comparison to a base set. The capability of a three filter system to discriminate between three highly overlapping mid-IR spectra (fuel oil, hexanes, acetone) was demonstrated using CDSD. This technique has similarities to the previously discussed methods found in the literature. Specifically, it requires that filtered data be used similarly to these referenced systems. CDSD however uses a unique approach, which both expands the dimensionality of the configuration-space, as well as, employs comparative relationships between the filter responses of the various chemicals in a given set. Unlike sensing techniques, such as programmable correlation radiometry or multivariate optical elements, CDSD is not attempting to resolve spectral differences based on the optical filter selectivity characteristics. On the contrary, this technique uses low-resolution, large-bandwidth, overlapping spectral filters to construct the representative chemical vectors, in order to explore relationships between vectors, as well as

their commonly constructed surfaces and volumes. Rather than using complex optical elements, the biomimetic CDSD approach relies on individual chemical responses to simple Gaussian shaped optical-filters. Therefore, this method can be viewed as a new approach to analyze data collected in systems similar to the described near-IR photometers. For any filter-based sensing approach, no matter the complexity of the filters involved, the selection of these filters is paramount to develop sensors with high selectivity. Recently, the effect of the near-IR filter bandwidth, degree of filter overlap, and angle between the response curves for analyte and interferent absorbance bands, were studied for a multifilter system.29 The results supported the conclusion that the optimization of a multifilter photometer is nontrivial, and that the number of constituents in the background and degree of spectral overlap of the constituents must be taken into account. This development is an extremely important consideration for such techniques. In our previously reported work, the optical filters used were chosen empirically, based upon the relative position of each filter to the IR absorption bands, without optimization of the system’s response to different filter parameters choices. While this approach worked well for first principle studies, we needed to explore the filter set choices further. The present work investigates the impact that the optical filter bandwidth and center wavelength position (relative to the target absorption band) have on the selectivity of a biomimetic three filter mid-infrared system for four volatile chemical vapors: pentane, hexanes, heptane, and acetone. During the course of the studies presented here, optimum filter sets (given initial boundary conditions) are compared to less “successful” filter sets to elucidate necessary requirements for filter selection to utilize this methodology for the implementation of future filter-based (nonspectral analysis) systems.

■

EXPERIMENTAL SECTION Materials and Data Collection Method. Laboratory grade hexanes (mixture of hexane isomers), and acetone, along with HPLC grade pentane and heptane were purchased from B

DOI: 10.1021/acs.analchem.5b01723 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

Figure 2. Temporal evolution of the FTIR spectra for (a) pentane, (b) hexanes, (c) heptane, and (d) acetone, flowing through the test apparatus at a rate of 0.1 mL/min. Each sample is injected at t = 0 min and spectra are collected every 10 s for 20 min. The absorbance values on the z-axis are set independently for each chemical to better illustrate the decay dynamics. Maximum absorbance values are (a) 2.5, (b) 8.0, (c) 2.5, and (d) 1.5, respectively.

Fisher Scientific and used as received. Additionally, a 50/50 mixture of hexanes/heptane was prepared by combining 1 mL of hexanes and 1 mL of heptane into a glass vial using volumetric pipettes (Eppendorf). For data collection, samples were injected (2.5 μL) through a septum using a sealed 10 μL syringe into a t-joint which was connected to an MKS Airgard spectrometer equipped with a 10 m path length gas cell using cajon fittings and Siltek treated stainless steel tubing. Dry laboratory air was flowed through the sample chamber at 0.1 mL/min. Spectra were collected every 10 s, starting 2 min before the injection at t = 0 s and continuing for an additional 20 min after injection. By collecting data points before injection, we were able to establish the baseline for the system with no chemical present. All raw data was collected and analyzed as absorption spectra. After each experiment, the system was thoroughly purged with dry laboratory air to remove any remaining chemicals. Each experimental run was repeated three times. Infrared Spectral Range Selection and Analysis. Similar to our previous work,26−28 we chose to study the spectra of these chemicals in the C−H stretch region between 2800 and 3100 cm−1 (3.6−3.2 μm). Within this region there is a strong overlap of the C−H stretching absorption bands for all chemicals studied, as observed in Figure 1. This presents an ideal spectral region for determining the effects of varying filter parameters in a biomimetic multifilter system. Spectral analysis was conducted using Matlab. A routine was written to import the absorbance spectra versus time data for each sample run, and save that data as a multipage matrix. This allowed us to vary specific aspects of the filter parameters and calculate the response to each chemical within a given filter set. To conduct CDSD analysis, we are first required to determine which spectra would be utilized for a training set, to build our base vectors and their geometrical operational relations.

Training spectral sets were chosen by examining the IR absorbance spectra shown in Figure 2 for each chemical flowing through the system, so that we could focus on well-defined spectral signatures to build the detection vectors. The “training spectral set” was chosen independently for each of the four chemicals studied. The use of “early time” spectra, meaning spectra collected in the initial stage of the experiments after injection, was adequate for quickly evaporating chemicals (i.e., acetone), while use of “later time” collected spectra was preferred for relatively slowly evaporating chemicals (i.e., heptane). These choices are particular to our experimental setup, which uses flowing finite amount of vapors, and therefore is dynamically adjustable by air flow. Filter Walk Experiment. The main focus of these studies was to determine the relative selectivity of various three-filter sets. This required the development of a method to study the discrimination capabilities of a large number of filter sets, in a systematic fashion. A significant issue with studying the interaction of a three-filter system with a chemical, while varying both the filter center wavelengths and fwhm, is the interpretation of the resulting data. In our methodology all filters have ideal Gaussian-shaped spectral profiles, characterized by the filter center wavelength (CWL) peak position (CWL1, CWL2, CWL3) and full width at half-maximum (FWHM1, FWHM2, FWHM3), each of which can be independently varied. Analysis of variable filter sets requires the reduction of the number of independent variables. This is accomplished using a two-step process. First, the fwhm is set to be equal for all filters (equal band filters). This reduces the number of independent variables to four (CWL1, CWL2, CWL3, FWHM1,2,3). Second, the relative peak position of each filter with respect to the other two filters is held constant (equidistant filters). This process results in two independent variables (CWL1, FWHM1,2,3), as CWL2 and CWL3 will be C

DOI: 10.1021/acs.analchem.5b01723 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

Figure 3. Experimental design layout for the filter walk experiments. The full extent of the experiment is presented on the left; the expanded view shows the CWL1 Position region from 2890 to 2900 cm−1, in steps of 2 cm−1, and fwhm from 32 to 40 cm−1, in steps of 2 cm−1. This maps the corresponding values for each filter set as they are varied throughout the experiment. The rightmost layout shows the design variables superimposed on the number of test spectra correctly identified for Hexanes, with a value of 10 as successful detection for all test spectra, and the value 0 as the inability to detect.

Figure 4. Normalized base vector plots for a filter set with (a) CWL1 = 2879 cm−1 and (b) CWL1 = 2910 cm−1, and CWL2, CWL3 spaced at identical peak separation. The fwhm of all filters is 10 cm−1 for both filter sets. The choice of peak location values results in less (a) or more (b) distinct vectors.

between our filter sets, no matter the absolute spectral peak position of each filter. We evaluated a “filter walk” across the region of interest, and contrasted it to fwhm variations of all filters. Thus, we compared the equidistant filter choices to spectral peak positions and, the equal-band parameter to bandwidth variation. The spacing between filters CWL1−CWL2 was fixed at 60 cm−1, and between CWL1−CWL3 at 86 cm−1, based on the relative distances between the peaks of interest as previously described (and denoted in Figure 1). All filters were “walked” across the spectral range of interest from a starting value of CWL1 at 2750 to 3000 cm−1 in steps of 2 cm−1. The fwhm of all filters was varied from 2 to 200 cm−1 in steps of 2 cm−1. This generated a matrix of 12 600 unique filter sets for

dictated by CWL1. This approach requires that an initial choice for the filter spacing be made. In order to make a rational choice for our filter spacing, we chose to base our initial filter selection criteria on the three-chemical set of pentane, hexanes, and heptane. In the region of interest, we observe three distinct absorption peaks for this chemical set. This allowed us to make an initial choice for our filter set by centering each of the three filters on one of the three main absorption peaks, as denoted by the dashed lines in Figure 1. Thus, the separation between CWL1-CWL2 and CWL1-CWL3 were fixed based on the center wavelength of these absorption peaks. This ensures that we will study filter sets where the filters are centered on the absorption peaks, and also defines a consistent separation D

DOI: 10.1021/acs.analchem.5b01723 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

The technique requires first to generate well-defined individual base vectors for the “target” chemical or chemicals to be identified. Using a set of n-optical (for our current studies, n = 3) transmission filters of finite but limited band, the total detectable power within the filter bandwidth is “collected” by a set of n-detectors, resulting in a list of values representing the total integrated response of the target signatures under the optical filter limits. This list of filter response values is then used to construct n-dimensional vectors of unit magnitude, within the n-spectral filter orthonormal configuration space, which is equivalent only to the relative response of the target signature to the optical filter band-pass. For the studies presented here, all filter response values are not directly collected through physical filters, rather they are calculated as the integrated total response between numerical filters with collected FTIR spectra of the target chemicals. This allows for the study of the discrimination ability of a large number of filter sets, which would otherwise be impossible from a practical standpoint. For a group of m-chemical targets (for our current studies m = 3, as we only examine three out of the four total chemicals at any given time), there are m normalized vectors of ndimensions, each uniquely representing a target chemical. These are called base vectors. This procedure is similar to fingerprint pattern identification.30−32 The m-base vectors define a polygon in the n-dimension configuration space. Any test set of filter response values that is collected through the existing n-optical filters, can be expressed as a normalizedmagnitude test vector in the same configuration space. If the test vector is a linear combination of any, or all, of the predetermined base vectors, it has definite geometrical relations with the base vector polygon’s sides and surface normals. More specifically, if the test vector is only a linear combination of the existing base vectors, then it lies within the polygon volume. Further, it has projections to all surface normals of the polygon’s sides. If the test vector lies outside the polygon volume, it may be expressed as a linear combination of components internal and external to the polygon. However, it will not have projections to all side surface normals, as it is not possible to be orthogonal to all of them. These orthogonality criteria serve as the discrimination rules for the test vectors representing each chemical. The test vectors are compared to each polygon side surface normal, and based on the comparison results a “truth” table is assembled. The truth table has three possible value entries: 0, 1, and ≫1. In the case of a 0 outcome, the test vector does not have a component within the polygon side defined by the particular base vectors. In the case of a 1 outcome, the test vector may be composed of a linear combination of the two base vectors acting as the edges of the polygon side. For an outcome ≫1, there is no possibility that the test vector is made by any component of the base set. The results of the truth table are further compared within the possible outcomes, to discriminate for cases of contradiction in result. At the end of the comparative step between the table outcomes, a likelihood of detection is the output. The operations reduce the possibilities to a binary outcome, of “yes” or “no,” concerning the presence of any of the m-target chemicals in the test material. In this fashion, CDSD analysis was conducted individually on each of the 10 test measured spectra for each chemical in the set. Adding up the corresponding binary outcome values for “yes” (1) and “no” (0) across our 10 test spectra provides the number of test spectra correctly identified. We save these values in three individual matrices in the position corresponding to the

study as outlined in Figure 3. This experiment acted as an ideal test case to begin to unravel the filter interaction with this chemical set. By moving the filters as described, at least one filter started or ended completely off the region of interest. Additionally the variation of the fwhm started as narrow as some of the specific features observed in the given IR spectra, to nearly as broad as the full IR spectral region of interest. Generation of Base and Test Vectors. Utilizing the selected 10 “training spectra,” we then generated a unique set of base vectors for each filter set under examination. For a given filter set, the overlap integral between each filter (F1, F2, F3) and each chemical spectrum was calculated, thus defining a 3dimensional vector in configuration space (F1, F2, F3). We performed this calculation for each of the 10 training spectra from each chemical. The resulting vectors were normalized, and an average base vector for each chemical was generated for each given filter set. Examples of base vectors for all four chemicals are shown in Figure 4, for two different filter sets with CWL1 at 2879 and 2910 cm−1, with FWHM1,2,3 equal to 10 cm−1 in both sets. The base vector for each chemical is dependent on the combined response between the overlap of each chemical’s IR absorption spectrum and optical filter, and thus these vectors change direction for different filter sets. The relative separation between the base vectors is the defining factor in discrimination between the different chemicals. Figure 4 indicates that for both of these filter sets, the acetone vector exhibits significant separation from all three alkane vectors. Additionally, comparison of Figure 4b (filter set with CWL1 = 2910 cm−1) shows that the separation between the alkane base vectors is greater than in Figure 4a(filter set with CWL1 = 2879 cm−1). This indicates that heptane is more readily discriminated from hexanes and pentane for this filter set. As our detection method is comparative, the relative interplay between the three chosen chemical detection vectors is what determines the discrimination capability of the operations. Such effects will be explored further and related back to these base vector plots in the following discussion section. Using the base vectors sets, we computed the results of CDSD analysis, on the spectral set collected for each chemical using the 10 test spectra that immediately follow the 10 training spectra in experimental data collection. A full detailed description of this technique is presented elsewhere,28 a brief overview is presented here for the reader. CDSD Data Analysis Technique. CDSD is an identification technique which utilizes the output of a biomimetic optical-filter based system to generate a binary response to the presence (1) or absence (0) of a particular chemical. Discrimination of chemical vapors is based on the operation of a set of optical filters on the infrared spectra of different chemical vapors. This operation results in the generation of unique sets of filter response values for each chemical of interest. Chemical vapors studied in the current work consist of three hydrocarbons as well as acetone, all of which exhibit very similar IR absorption spectra (Figure 1). Discrimination with this technique is made based on similarities between an established set of filter response values and the unknown filter response values to be tested. It is used to discriminate between very similar spectral signatures of unmixed chemicals within a limited spectral band, and allows the estimation of the chemical components in a mixture or, the presence of unmixed chemicals in a spectrally “cluttered” background environment. E

DOI: 10.1021/acs.analchem.5b01723 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

Figure 5. Number of test spectra correctly identified for pentane (a), hexanes (b), and heptane (c) as the comparison set. The color scale represents 100% correct detection (10 out of 10 test spectra correctly identified) as red and 0% correct detection (0 out of 10 test spectra correctly identified) as blue.

Figure 6. Number of test spectra correctly identified for acetone (a), hexanes (b), and heptane (c) as the comparison set. The color scale represents 100% correct detection as red (10 out of 10 test spectra correctly identified) and 0% correct detection (0 out of 10 test spectra correctly identified) as blue.

Specifically, the development of this method for filter selection is a necessary step to constructing highly selective filter based sensors. Previously, CDSD has been used to evaluate the presence or absence of chemical species, within sets of noninteracting mixtures or in isolated cases, in the presence of common atmospheric conditions, such as humidity. These results have been reported.27,28 The study presented here focuses on the exploration of the configuration of the filter-derived space, in order to determine selection rules for optical filter choices, which will lead directly to the realization of functional experimental apparatus. Therefore, the results presented rely on unmixed chemical data as well as an example test case with a 50/50 mixture of hexanes/heptane.

filter set under consideration. We then move to the next filter set (as described in Figure 3) and repeat the above analysis procedure. First, calculate base vectors based on the overlap integral of this filter set and the 10 training spectra for each chemical, then conduct CDSD analysis on the 10 test spectra, determine the number of test spectra correctly identified, and store this value in the corresponding matrix. After completing this analysis for all 12 600 filter sets, we then have full matrices with complete results for a given three chemical set. These results are then plotted as contour maps for each individual chemical, where the x-axis represents CWL1 (recalling that CWL2 and CWL3 are set to be dependent on CWL1), the yaxis represents the fwhm of all filters (FWHM1,2,3), and the height of the contours correspond to the number of test spectra correctly identified, from none (0) to all (10). The matrices are designed and presented in this manner, such that clear comparison between discrimination of various filter sets could be realized. Since this data was all collected using unmixed chemicals, and only differ in experimental collection duration times, the only change in the spectra between the training and test sets is a reduction in the overall signal intensity (due to the evaporation and flow of the chemicals through the system with time). Therefore, when studying these matrices, a clear distinction can be identified between filter sets that provide strong discrimination and those sets that provide little or no discrimination.

■

RESULTS AND DISCUSSION Minor changes in the relative positions of the three color receptors in the mammalian color vision can result in color blindness.33,34 Therefore, it can be expected that the mid-IR biomimetic system described herein will exhibit similar relationships between the filter positions and fwhm. Additionally, the key point of our CDSD analysis, which correlates directly to biomimetic color vision, is that both are comparative techniques. Thus, in addition to the filter choice, the examined chemical system also plays a key role in the discrimination. For example, if one is interested in selecting a filter set which discriminates between hexanes and heptane, within our F

DOI: 10.1021/acs.analchem.5b01723 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

Figure 7. Number of test spectra correctly identified for acetone (a), hexanes (b), and pentane (c) as the comparison set. The color scale represents 100% correct detection as red (10 out of 10 test spectra correctly identified) and 0% correct detection (0 out of 10 test spectra correctly identified) as blue.

collected data set, one could choose to compare these two chemicals to either pentane (very similar spectral signature) or acetone (relatively different spectral signature). Likewise, other possible comparative combinations of three chemical sets could be studied. We chose to examine three distinct comparative combinations; pentane−hexanes−heptane (PHT), acetone− hexanes−heptane (AHT), and acetone−hexanes−pentane (AHP). By studying these specific chemical combinations we examined a data set where all chemicals have strongly overlapping infrared signatures (PHT), as well as two sets where the IR spectra of one of the chemicals was less strongly overlapping (AHT and AHP). This allows us to examine whether a third chemical that was relatively different, or very similar generated the greatest discrimination between two similar chemical spectra. For all three chemical sets, the same set of filter walks (i.e., the same 12 600 filter sets) were examined as outlined in the Experimental Section. Comparative Combination Studies with Varying Filters. The first illustrative result from these studies is the effect of the filter fwhm in combination with the PHT chemical set presented in Figure 5. This experiment demonstrated that when all three chemical spectra were strongly overlapping, broader filter sets (fwhm > 20 cm−1) showed no ability to discriminate across all three chemicals, no matter where the filters were positioned. This means that the broad filters have the same (or very similar) responses to chemicals with highly overlapping spectra. This leads to poorly defined separation between the individual chemical base vectors, as the angle between all three vectors will be essentially zero, and thus no discrimination between the chemicals will occur. These results are in relatively stark contrast to both experiments where acetone replaced one of the highly overlapping chemicals. For instance, in Figures 5 and 6, replacing pentane with acetone, much broader filter sets (up to a fwhm = 100 cm−1 at CWL1 = 3000 cm−1) are capable of discriminating between each of the three chemicals. While it is not surprising that a large number of filter sets (of varying both CWL1 position and fwhm) show clear discrimination of acetone, since acetone’s spectral signature is relatively different, the fact that broader filter sets could discriminate hexanes and heptane in this case is significant. These results indicate that the volume of the polygon in the 3-dimension configuration space defined by the AHT chemical system is greater than the corresponding

polygon volume for the PHT chemical set. By expanding the polygon volume, the capability to discriminate between hexanes and heptane, each exhibiting strongly overlapping IR absorption bands, is significantly enhanced to include the use of much broader optical filters. This concept was further explored and confirmed by examining the AHP chemical set as presented in Figure 7. In this case, there are a more limited number of filter sets with all 10 test spectra correctly identified (100% correct detection). This can be justified as the H−P separation will be less than the H-T separation due to the increased similarities between component spectra (Figure 1). Comparing the AHP set to the original PHT set, however, shows that replacing the more similar heptane with the less similar acetone again increases the number of filter choices which will work to discriminate between all three chemicals, and also allows for the use of broader filters. In this instance, filter choices with CWL1 = 2932 or 2934 cm−1 and a fwhm = 26 cm−1, demonstrate 100% correct detection for all three chemicals for the spectra studied. For the original set, the broadest filters which would meet these requirements were 16 cm−1 fwhm (at a CWL1 position of 2892−2896 cm−1). Additionally, a significantly larger number of broader filter choices are able to provide 100% correct detection for hexanes in the system with acetone, compared to the set where heptane is the third chemical of interest. All of these results indicate that if one is interested in using the CDSD method to distinguish between two very similar chemicals (and the third chemical is a matter of choice), choosing a chemical that has a relatively different spectral signature such that the vector volume is increased, will yield a higher confidence result. This relates directly back to the biomimetic properties of our system as it utilizes a comparative technique. Therefore, the relationship between each set of chemicals is a key component for choosing a proper filter set for adequate discrimination. These results are further confirmed by comparing the absolute number of filter sets with 100% correct detection, within our 12 600 studied filter sets for each of the three sets of base vectors. Table 1 demonstrates that the number of available filters is dramatically increased by using acetone as a base vector (AHT and AHP sets) than with the alkanes alone (PHT). Considering each filter’s peak spectral position, the studies presented indicate that with fixed spacing between F1−F2 and F1−F3, there is common discrimination by filter choices with CWL1 values above 2900 cm−1 for all chemical sets studied. G

DOI: 10.1021/acs.analchem.5b01723 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

centered directly on the absorption peaks, does not provide discrimination between these chemicals. For the slightly shifted filter choices, a measurable separation between the hexanes and pentane vectors is apparent, as the filters are able to detect between the subtle differences in these very strongly overlapping spectra. Thus, the latter filter choice allows for discrimination between these chemicals, as confirmed with the results from the number of test spectra correctly identified. Construction of Combined CDSD Plots for Filter Selection. All work presented above utilized as-received, unmixed single chemicals. This is a requirement to determine which filter choices are optimal using the CDSD method. Any sensing applications based on this method, however, will be expected to distinguish not only if a single target chemical is present, but also when multiple chemicals in a measured set are present. Therefore, we wanted to study how well selected filter choices were able to distinguish a 50/50 mixture of hexanes/ heptane, to correlate our selection maps developed using single chemicals with a more complicated test scenario. To do so, we first needed to select filter choices with high probability of correctly identifying the presence of each of the three chemicals. The individual chemical selection maps are inadequate for this purpose as they only show the results for each single chemical. Therefore, the plots for the individual chemicals in the PHT (Figure 4) and AHT (Figure 7) chemical sets were combined such that the resulting contour plots directly show the likelihood of correct identification for all three components, with respect to the changing filter parameters. The combined plots were generated by multiplying the results of the single chemical matrices together, for each individual filter set. This operation results in values representing the confidence of correct identification of a test chemical in the three chemical set (CCI). These results for the PHT and AHT sets are presented in Figure 8. In these combined confidence plots, the height of the contours represent the CCI for each three-chemical set and is defined from 0 to 1000. If any one chemical shows no correct detection for a given filter set, the CCI value will be zero, no matter the results for the other chemicals. Likewise, a maximum

Table 1. Number of Filter Sets with 100% Correct Detection for Each of the 4 Target Chemicals Using PHT, AHT, and AHP Base Vectors, Respectively pentane hexanes heptane acetone

PHT

AHT

AHP

104 108 491 N/A

N/A 3445 3099 9430

513 1943 N/A 6049

This result means that our initial hypothesis to place the filters centered (CWL1 = 2879 cm−1) on the spectral absorption peaks for greatest discrimination was incorrect. With this spacing, filter choices with center wavelengths offset from the chemical absorption peak perform better. By placing the filter spectral peaks at these values, it appears that the subtle differences between the closely overlapping spectra generate larger variances in the filter response, leading to greater separation. Further, considering filter choices at higher wavenumbers (CWL1 > 2900 cm−1), for the chemical sets that replace one of the initial chemicals with acetone, show that positioning the filters in this region allows for broad filters to be utilized. This is an important finding as we are most interested in studying filters with large fwhm since they are more readily commercially available, and should also provide greater sensitivity due to greater signal throughput. To further demonstrate and confirm these findings, we examined the vectors calculated from both our original filter choice position (CWL1 = 2879 cm−1) as well as when the filters are positioned slightly off of the center of the spectral peaks (CWL1 = 2910 cm−1), as previously described in Figure 4. We note that in both cases the acetone vector unsurprisingly has a large separation from the three alkane vectors. This distance is what allows for greater discrimination using acetone as the third target chemical because of the greater difference in comparison between acetone and any of the alkanes, in contrast to comparison between any pair of alkanes. Examining the separation between the alkane vectors shows that for CWL1 = 2879 cm−1, the hexanes and pentane vectors become indistinguishable from one other. This is why a filter choice

Figure 8. Combined confidence plots for pentane, hexanes, heptane set (a) and for the acetone, hexanes, heptane set (b). The color scale represents 100% correct detection (CCI = 1000) as red and 0% (incorrect) detection (CCI = 0) as blue. The positions of the markers (×) indicate the specific values for the filters in the experiments of Figure 9. H

DOI: 10.1021/acs.analchem.5b01723 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

Figure 9. Binary “yes/no” signal detection CDSD results for a measured 50/50 mixture of hexanes/heptane over time, using (a) hexanes, pentane, heptane discrimination vectors with CWL1 = 2910 cm−1 and (b) hexanes, acetone, heptane discrimination vectors with CWL1 = 2910 cm−1. Results for filter sets corresponding to the markers in Figure 8, with fwhm = 10 cm−1 (top) and fwhm = 40 cm−1 (bottom) are presented for each case.

CCI of 1000 is only found for filter sets which have 100% correct detection of the test set (values of 10 on the individual chemical plots) across all three chemicals. Using these combined confidence plots in Figure 8, we then chose two filter sets from each three-chemical set to examine with our 50/50 hexanes/heptane mixture. The positions of these filter sets are denoted by the markers (×) in Figure 8. For the PHT system we used filter sets with CWL1 = 2900 cm−1 and fwhm of 10 and 40 cm−1. For the AHT system we used filter sets with CWL1 = 2910 cm−1 and fwhm of 10 and 40 cm−1. These filters were specifically chosen so that we could explore the response for the 50/50 hexanes/heptane mixture to similar filter choices between the two chemical sets, while operating in regions where we expect to achieve good discrimination based on the combined confidence maps. Mixture Discrimination Results using Selected Filters. Detection results for the 50/50 hexanes/heptane mixture are presented for the four selected filter sets in Figure 9. Results presented in Figure 9 are binary in nature; a response of 1.0, 0.9, and 0.8 for hexanes, pentane, and heptane or acetone respectively represent detection of each chemical using the CDSD method. These values were chosen such that the “on” detection values were offset for clarity and have no other meaning. Therefore, in Figure 9, if a chemical is detected, the CDSD value reported for each chemical will have the assigned values, if there is no detection, then the value for that chemical will be zero. Because these are binary “yes/no” confidence detection plots, the shape of the curves show no chemical dynamics. Once enough chemical is present in the system, the signal will turn “on”. The signal will remain “on” as the chemical evaporates and flows through the system, so long as enough vapors remain present for clear discrimination. The results in Figure 9 confirm our previous findings. Mainly, we can see that both of the filter selections in Figure 9b relating to the AHT chemical set show a clear detection of the 50/50 hexanes/heptane mixture for the majority of the experiment. For times before 200 s we observe that the signal is saturated and thus our method detects no chemical present, as the signal is out of range. From t = 200 s to t = 750 s (with fwhm = 10 cm−1) and from t = 200 s to t = 590 s (with fwhm = 40 cm−1) our method detects both hexanes and heptane present, with no

false positive signal (acetone) throughout the length of the experiment. In both cases the system detects heptane for a longer time due to the relative evaporation rates between the two flowing chemicals. Compared to the results from the PHT discrimination set in Figure 9a, we observe clearly longer detection times for the studied fwhm, by exchanging the closely overlapping pentane with acetone. Additionally, in Figure 9a, we observe a false detection of pentane near the end of the detection time for both filter choices. As the chemical evaporates and moves through the FTIR cell, the resulting signal weakens. CDSD then incorrectly identifies the residual hexanes as pentane because of the strong overlap of the two chemical spectra (Figure 1). Such false positives can be removed by further adjusting our acceptable error in measurement. These results demonstrate that the combined confidence plots are a valuable means to select filter sets with high probability of detection for a given chemical set.

■

CONCLUSIONS We have shown high confidence discrimination of strongly overlapping infrared chemical spectra using carefully selected broadband optical spectral filters. We demonstrated a systematic exploration of the relative effects of optical filter peak position and fwhm on chemical detection, within a given set of strongly overlapping spectral signatures. We have shown that discrimination with this methodology is clearly dependent not only on the optical filter characteristics, but also on the given chemical set under study. Specifically, we have shown that with filter spacing based on the relative position of spectral features, filter choices with center positions offset from these features provide the greatest discrimination. Additionally, in the case of two very strongly overlapping chemical spectra (i.e., hexanes and pentane), we demonstrated that a relatively spectrally different chemical (i.e., acetone) provides a greater number of potential filter choices for discrimination than a more closely overlapping chemical spectrum (i.e., heptane). These results provide an important basis for understanding chemical detection using optical filter-based CDSD. Importantly they suggest that our three filter system is behaving similarly to biomimetic human color vision. The insight gained I

DOI: 10.1021/acs.analchem.5b01723 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

hardware system development. This system will allow for the direct collection of optically filtered data, without the need for a spectrometer. Once fully evaluated, the prototype will allow for the determination of the achievable discriminatory limits of this technique, and thus allow for comparison to more traditional sensing approaches.

from these tests will be applied to even more complicated chemical sets and mixtures, specifically those with more than three chemicals present. Knowing that there is a benefit to utilizing a third relatively different chemical, to distinguish between two strongly overlapping chemical signatures, can be directly applied to even more complex chemical systems. Since our method is comparative, we can use this effect to our advantage and design both filter and chemical sets based on specific target chemicals of interest. The work presented here increases the understanding of filter-based chemical sensing using CDSD, thus illustrating the advantages and disadvantages of the technique. As with many filter-based sensing techniques, one of the main benefits is the potential to develop inexpensive, lightweight chemical sensors with no moving parts or spectral scanning. The CDSD method, however, allows for the use of low resolution “off-the-shelf” transmission filters, as long as these filters are properly selected, as illustrated in the work presented in this paper. Furthermore, CDSD expands the configuration-space and thus benefits from the presence of interferent chemicals (that are part of the initial training set) as it will compare effectively to the desired target chemical. This also relates directly to the main weakness of this approach, as the algorithm must be properly trained for the desired chemical set or no detection information can be obtained. Unlike traditional spectral-scanning methods, sensors based on this approach will only yield n-coordinate values (where n = number of detectors) rather than full spectra, and thus it is only appropriate to provide information on the presence or absence of chemicals “known” to the system, i.e. those which it has been trained to detect. A key point, is that though these studies illuminate some very important aspects of the relative effect of filter positioning, overlap, and bandwidth on chemical detection using a biomimetic multifilter optical system, it is not possible to study all potential filter combinations. As both the relative position and fwhm of each filter is an independent variable, there is an infinite number of potential filter sets that could be explored. Additionally, we utilized only Gaussian shaped filters, though in theory there is no limit to the type of filter line shape that could be applied to this method. Because of this fact and the results from these initial tests, we are continuing to work on additional means to study how the relationship between filter choices and chemical sets affect the discrimination of various chemical systems. Such studies may further illustrate which vector parameters need to be maximized to achieve the greatest discrimination between chemicals. The method presented in this paper, however, is a clear means to explore the relationship between any two independent optical transmission filter variables for any filter-based sensing method. Therefore, this approach can be utilized to narrow the scope of potential filter choices which will allow for more appropriate optimization models and for the selection of filters from those which are commercially available. The information gleaned from these studies is therefore an important step to determine potential optical filter choice combinations to study in direct-filtering devices which use this biomimetic approach to chemical detection. We are currently in the process of completing the evaluation of an experimental prototype system based on this methodology. The filters selected for this prototype were chosen using the approach presented in this paper. These methods narrowed the scope of potential filter choices during the design phase of the instrument and were thus a fundamental step in the functional

■

AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. Fax: (704) 687-8241. Notes

The authors declare no competing financial interest.

■ ■

ACKNOWLEDGMENTS This work was supported by funding from the Office of Naval Research (Award Number: N000141310208). REFERENCES

(1) Janni, J.; Gilbert, B. D.; Field, R. W.; Steinfeld, J. I. Spectrochim. Acta, Part A 1997, 53, 1375−1381. (2) Steinfeld, J. I.; Wormhoudt, J. Annu. Rev. Phys. Chem. 1998, 49, 203−232. (3) De Lucia, F. C.; Gottfried, J. L.; Munson, C. A.; Miziolek, A. W. Spectrochim. Acta, Part B 2007, 62, 1399−1404. (4) Daniel, N. W.; Lewis, I. R.; Griffiths, P. R. Appl. Spectrosc. 1997, 51, 1868−1879. (5) Blake, T. A.; Kelly, J. F.; Gallagher, N. B.; Gassman, P. L.; Johnson, T. J. Anal. Bioanal. Chem. 2009, 395, 337−348. (6) Banas, K.; Banas, A.; Moser, H. O.; Bahou, M.; Li, W.; Yang, P.; Cholewa, M.; Lim, S. K. Anal. Chem. 2010, 82, 3038−3044. (7) Botti, S.; Almaviva, S.; Cantarini, L.; Palucci, A.; Puiu, A.; Rufoloni, A. J. Raman Spectrosc. 2013, 44, 463−468. (8) Mahoney, C. M.; Fahey, A. J.; Steffens, K. L.; Benner, B. A.; Lareau, R. T. Anal. Chem. 2010, 82, 7237−7248. (9) Gottfried, J. L.; De Lucia, F. C., Jr.; Munson, C. A.; Miziolek, A. W. J. Anal. At. Spectrom. 2008, 23, 205−216. (10) Saleem, A.; Canal, C.; Hutchins, D. A.; Davis, L. A. J.; Green, R. J. Anal. Methods 2011, 3, 2298−2306. (11) Castro-Suarez, J. R.; Pacheco-Londono, L. C.; Velez-Reyes, M.; Diem, M.; Tague, T. J.; Hernandez-Rivera, S. P. Appl. Spectrosc. 2013, 67, 181−186. (12) Germain, M. E.; Knapp, M. J. J. Am. Chem. Soc. 2008, 130, 5422. (13) Hughes, A. D.; Glenn, I. C.; Patrick, A. D.; Ellington, A.; Anslyn, E. V. Chem. - Eur. J. 2008, 14, 1822−1827. (14) Wyszecki, G.; Stiles, W. S. Colorimetry. In Color Science Concepts and Mehods, Quantitative Data and Formulae, 2nd ed.; John Wiley & Sons: New York, 1982. (15) Ewing, K. J.; Sanghera, J.; Aggarwal, I. D.; Block, M. J. Proc. SPIE 2004, 5585, 98−108. (16) Amerov, A. K.; Sun, Y.; Arnold, M. A.; Small, G. W. Proc. SPIE 2001, 4363, 1−10. (17) Helwig, A. M.; Arnold, M. A.; Small, G. W. Appl. Opt. 2000, 39, 4715−4720. (18) Sodickson, L. A.; Block, M. J. Clinical Chemistry 1994, 40, 1838−1844. (19) Allen, J. J.; Sinclair, M. B.; Pfeifer, K. B. Advanced Polychromator Systems for Remote Chemical Sensing, LDRD Project 52575 SAND2004-6700; Sandia National Laboratories: Albuquerque, NM, 2005. (20) Sinclair, M. B.; Butler, M. A.; Ricco, A. J.; Senturia, S. D. Appl. Opt. 1997, 36, 3342−3348. (21) Sinclair, M. B.; Pfeifer, K. B.; Butler, M. A.; Senturia, S. D.; Deutsch, E. R.; Youngner, D.; Cabuz, E.; Hocker, G. B. In Moems and Miniaturized Systems IV; El Fatatry, A., Ed.; SPIE-Int Soc Optical Engineering: Bellingham, WA, 2004; pp 37−47. (22) Myrick, M. L.; Soyemi, O.; Li, H.; Zhang, L.; Eastwood, D. Fresenius' J. Anal. Chem. 2001, 369, 351−355. J

DOI: 10.1021/acs.analchem.5b01723 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry (23) Myrick, M. L.; Soyemi, O. O.; Schiza, M. V.; Farr, J. R.; Haibach, F.; Greer, A.; Li, H.; Priore, R. Proc. SPIE 2002, 4574, 208−215. (24) Nelson, M. P.; Aust, J. F.; Dobrowolski, J. A.; Verly, P. G.; Myrick, M. L. Anal. Chem. 1998, 70, 73−82. (25) Soyemi, O.; Eastwood, D.; Zhang, L.; Li, H.; Karunamuni, J.; Gemperline, P.; Synowicki, R. A.; Myrick, M. L. Anal. Chem. 2001, 73, 1069−1079. (26) Aggarwal, I. D.; Major, K.; Poutous, M.; Ewing, K.; Sanghera, J. In Advanced Photonics; Optical Society of America: Barcelona, 2014. (27) Major, K. J.; Ewing, K. J.; Poutous, M. K.; Sanghera, J. S.; Aggarwal, I. D. Proc. SPIE 2014, 9072, 907212. (28) Poutous, M. K.; Major, K. J.; Ewing, K. J.; Sanghera, J.; Aggarwal, I. Appl. Spectrosc. 2015, 69, 305−313. (29) Tarumi, T.; Amerov, A. K.; Arnold, M. A.; Small, G. W. Appl. Spectrosc. 2009, 63, 700−708. (30) Moayer, B.; Fu, K. S. Pattern Recognition 1975, 7, 1−23. (31) Kawagoe, M.; Tojo, A. Pattern Recognition 1984, 17, 295−303. (32) Chong, M. M. S.; Tan, H. N.; Jun, L.; Gay, R. K. L. Pattern Recognition 1997, 30, 1475−1488. (33) Blackwell, H. R.; Blackwell, O. M. Vision Res. 1961, 1, 62−107. (34) Carroll, J.; Neitz, M.; Hofer, H.; Neitz, J.; Williams, D. R. Proc. Natl. Acad. Sci. U. S. A. 2004, 101, 8461−8466.

K

DOI: 10.1021/acs.analchem.5b01723 Anal. Chem. XXXX, XXX, XXX−XXX