# Fileset

[c8mh01169a.pdf](https://mdr.nims.go.jp/filesets/88c07a54-8dba-47c8-bed3-74cb278d4f28/download)

## Creator

[Kosuke Minami](https://orcid.org/0000-0003-4145-1118), [Gaku Imamura](https://orcid.org/0000-0002-3130-7190), Takahiro Nemoto, [Kota Shiba](https://orcid.org/0000-0001-7775-0318), [Genki Yoshikawa](https://orcid.org/0000-0002-9136-8964)

## Rights



## Other metadata

[Pattern recognition of solid materials by multiple probe gases](https://mdr.nims.go.jp/datasets/5ea1413e-8f06-41c1-b6c3-e4f335cf9729)

## Fulltext

Pattern recognition of solid materials by multiple probe gases580 | Mater. Horiz., 2019, 6, 580--586 This journal is©The Royal Society of Chemistry 2019Cite this:Mater. Horiz., 2019,6, 580Pattern recognition of solid materials by multipleprobe gases†Kosuke Minami, *ab Gaku Imamura, ab Takahiro Nemoto,a Kota Shiba ab andGenki Yoshikawa abcA pattern recognition-based chemical sensor array is an efficientapproach to discriminating odours or a complex mixture of gaseousmolecules. In such an approach, solid materials are coated onsurfaces of sensors as probe receptors, and gaseous moleculesare exposed to those sensors as targets. Here, we propose thereverse approach, that is, gaseous molecules as probes and solidmaterials as targets, leading to pattern recognition of solid materials.Using a nanomechanical sensor as an example of a sensing platform,we have demonstrated that this approach can discriminate polymerswith different molecular weights as well as those having slightlydifferent functional groups evaluated through detailed classificationusing a support vector machine in addition to principal componentanalysis and linear discriminant analysis. Classification of those targetsolid materials with 100% accuracy has been achieved with somespecific combinations of probe gases. Since any kind of gaseousmolecule and any kind of chemical sensor can be utilized as theprobe and sensing platform, respectively, this study will open a newhorizon for comprehensive analysis of solid materials through apattern formed by the gas–solid interaction.Chemical sensor arrays have attracted significant attention asa powerful tool for detecting, discriminating and identifyingtarget analytes, especially various odours composed of a com-plex mixture of gaseous specimens. A large variety of chemicalsensor arrays have been utilized, including quartz crystalmicrobalances (QCM), conducting polymers (CP), field-effecttransistors (FET) and nanomechanical sensors.1–5 In a chemicalsensor array, sensing signals are obtained by measuringphysicochemical interactions induced by the sorption of targetanalytes in sensing materials designed to respond to a widerange of chemical classes. Since such a multidimensional data-set obtained by the chemical sensor array contains much infor-mation, multivariate analyses and machine learning can beeffectively applied to discriminate and identify each specimen.Although a wide range of applications have been demonstratedin various fields, such as food, agriculture, medicine and environ-mental science,5–11 these pattern recognition-based analyses arebasically limited to gaseous analytes.In this study, we propose a reverse approach, that is, patternrecognition of solid materials. As the sensing signals of chemicalsensors are based on the interaction between gases and solids,a sensing element and a target analyte should be exchangeable(i.e. solid materials as target analytes and gaseous molecules assensing probes), leading to the pattern recognition of solida Center for Functional Sensor & Actuator (CFSN), National Institute for MaterialsScience (NIMS), 1-1 Namiki, Tsukuba, Ibaraki 305-0044, Japan.E-mail: MINAMI.Kosuke@nims.go.jpb International Center for Materials Nanoarchitectonics (MANA), National Institutefor Materials Science (NIMS), 1-1 Namiki, Tsukuba, Ibaraki 305-0044, Japanc Materials Science and Engineering, Graduate School of Pure and Applied Science,University of Tsukuba, Tennodai 1-1-1 Tsukuba, Ibaraki 305-8571, Japan† Electronic supplementary information (ESI) available: Additional sensingsignals, additional PCA and LDA analyses, and additional results of SVM andMahalanobis distances. See DOI: 10.1039/c8mh01169aReceived 19th September 2018,Accepted 20th November 2018DOI: 10.1039/c8mh01169arsc.li/materials-horizonsConceptual insightsA conventional analysis of solid materials generally focuses on thespecific physical/chemical parameters of the materials. Here, we demon-strate a novel concept – that is, analysis of solid materials through their‘‘pattern’’ formed as a result of dynamic interaction between solid andmultiple probe gases. Since gaseous molecules diffuse into solid materialswhile interacting with most atoms inside, resultant dynamic responsescontain much information stemming from various physical/chemicalinteractions. Thus, the pattern provides a high-resolution fingerprintof the solid material, reflecting not only its intrinsic material propertiesbut also its myriad properties (e.g. detailed surface/internal/interfacialstructures and distribution of thickness/morphology-dependent physicalproperties), which are usually difficult to be fully covered by conventionalapproaches. Since any kind of gaseous molecule can be utilized as a probe,this approach possesses unlimited possibilities to differentiate solidmaterials and their properties. Moreover, the target is not limited to asimple material but includes a complex mixture of functional materialsand various thin-film devices as long as gaseous molecules can interact.In contrast to conventional materials science, which usually focuses onlyon a certain aspect, this concept provides a novel insight in terms of acomprehensive ‘‘pattern’’, which contains much information includingproperties inaccessible with existing approaches.MaterialsHorizonsCOMMUNICATIONOpen Access Article. Published on 19 December 2018. Downloaded on 3/19/2019 4:26:47 AM.  This article is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported Licence.View Article OnlineView Journal  | View Issuehttp://orcid.org/0000-0003-4145-1118http://orcid.org/0000-0002-3130-7190http://orcid.org/0000-0001-7775-0318http://orcid.org/0000-0002-9136-8964http://crossmark.crossref.org/dialog/?doi=10.1039/c8mh01169a&domain=pdf&date_stamp=2018-12-03http://rsc.li/materials-horizonshttp://creativecommons.org/licenses/by-nc/3.0/http://creativecommons.org/licenses/by-nc/3.0/http://dx.doi.org/10.1039/c8mh01169ahttps://pubs.rsc.org/en/journals/journal/MHhttps://pubs.rsc.org/en/journals/journal/MH?issueid=MH006003Kosuke MINAMIThis journal is©The Royal Society of Chemistry 2019 Mater. Horiz., 2019, 6, 580--586 | 581materials. To demonstrate this new approach, we focus on ananomechanical sensor as an example of a sensing platform.A nanomechanical sensor detects mechanical informationderived from the interactions between gaseous molecules andsolid materials with high sensitivity. Since it has been confirmedthat almost all kinds of solid materials including organic smallmolecules, polymers and inorganic nanoparticles provide somesignals as a result of the gas–solid interaction,12–20 a nano-mechanical sensor is an ideal platform to examine various kindsof solid materials. We have demonstrated successful discrimina-tion of polymers having different molecular weights as well asthose composed of different monomers by means of patternrecognition. Furthermore, detailed analysis using support vectormachine (SVM)-based classification models has revealed thatonly 2 or 3 selected probe gases can identify solid specimenswith high classification accuracy. Since any gas species includingthe complex mixture can be utilized as a probe to increase thevariety of signal patterns, this approach is expected to provideunlimited resolution of patterns of solid materials depending oneach purpose.As an initial proof-of-concept, we performed identificationof 4 different polymers through pattern recognition using nano-mechanical Membrane-type Surface stress Sensors (MSS).21,22We selected polystyrene (PS) and poly(4-methylstyrene) (P4MS) asa set having similar chemical structures, and polycaprolactone(PCL) and poly(vinylidene fluoride) (PVF) as a set with ahydrophobic nature (Fig. S1; see also the Supplementary Text,ESI†). Each polymer was dissolved in DMF and deposited ontoeach channel of MSS by inkjet spotting. Twelve different vapoursare used as probes to acquire signals for each gas–solid inter-action (details are provided in the Experimental section). Uponexposure to each vapour, the polymers exhibited unique responsesin terms of their intensity as well as their shape (Fig. 1a; see alsoFig. S2, ESI,† for all signal responses), reflecting the differences inchemical and physical affinity between each polymer and vapour.For the obtained dataset, we conducted unsupervised and super-vised analyses, namely principal component analysis (PCA) andlinear discriminant analysis (LDA), respectively. Multiple para-meters were extracted as feature sets from each decay curve ofeach normalized signal response (Fig. 1b; details can be foundin the Experimental section).23,24 With all the features from the12 vapours, the 4 different polymers can be clearly distinguishedby forming well-separated clusters in the principal componentspace (Fig. 1c; see also Fig. S3, ESI,† for the PC1–3 and PC2–3planes). On the PC 1–2 plane, PS and P4MS form clusters closeto each other, reflecting their similarity in the chemical andphysical affinity to each probe gas. With the LDA as shown inFig. 1d, each polymer was clearly classified without any overlaps,demonstrating the feasibility of the present approach to dis-criminating solid materials by pattern recognition. It should benoted that several small sub-clusters can be found in eachcluster (Fig. 1c). Since each small sub-cluster corresponds toFig. 1 Identification of polymers by pattern recognition using probe gases. (a) Typical signal responses of MSS. See also Fig. S2, ESI.† (b) Schematics ofthe methods for feature extraction from each normalized signal response. (c and d) PCA and LDA score plots of 4 polymers using 12 different probegases. PS(350k), polystyrene, Mw = 350 000 (red); P4MS, poly(4-methylstyrene) (black); PVF, poly(vinylidene fluoride) (green); PCL, polycaprolactone(blue). N = 11.Communication Materials HorizonsOpen Access Article. Published on 19 December 2018. Downloaded on 3/19/2019 4:26:47 AM.  This article is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported Licence.View Article Onlinehttp://creativecommons.org/licenses/by-nc/3.0/http://creativecommons.org/licenses/by-nc/3.0/http://dx.doi.org/10.1039/c8mh01169a582 | Mater. Horiz., 2019, 6, 580--586 This journal is©The Royal Society of Chemistry 2019each polymer layer coated onto each channel of the MSS(11 channels for each polymer species), the differences betweenthese sub-clusters are regarded as the coating reproducibilityof each polymer layer. Thus, this approach is proved to haveenough resolution to discriminate such minute differences inthe quality of coatings as well as the different materials.We also developed machine learning models based on aSVM classifier with a non-linear kernel.25 The 36 feature sets(3 parameters from each probe gas) of 132 samples (33 samplesfrom each polymer) were used in building an optimal SVMmodel and its validation. Eighty percent of the samples(105–106 samples) were used for the training dataset. Aftertuning the hyperparameters of a radial basis function (C and g),the remaining 26–27 samples were used for validation of theSVM model. To calculate identification accuracy, 5-fold crossvalidation was adopted.26 All combinations of each probe gaswere calculated to create SVM models. The number of trainedSVM models was 4095 (= 212 � 1). The details of the SVMclassifier can be found in the Experimental section. Identifi-cation accuracies depending on the combination of the probegases are shown as a dot plot in Fig. 2a, and the calculatedresults of average accuracy obtained from the combinationswith selected probe gases are shown in Fig. S4, ESI.† By the SVManalysis, the feature set from the 12 probe gases can clearly classifyeach polymer with 100% identification accuracy. Remarkably,almost a quarter of all combinations of the probe gases resultedin 100% identification accuracy with 2 to 12 kinds of probe gases.The usage rates of all the probe gases are summarized in Table 1.These results clearly indicate that the appropriate selection ofprobe gases depending on the target solid samples leads tohighly accurate and efficient identification (Fig. 2b). For exam-ple, in the present case, the specific combinations of two probegases (i.e. [ethyl acetate, ethanol], [ethyl acetate, toluene] and[ethyl acetate, chloroform]) achieved 100% identification accu-racy, while another combination [ethyl acetate, propionic acid]resulted in the worst accuracy with 75.5 � 14.1%. Based onthese results, we conducted PCA again for visual recognitionusing the best and worst combinations of the two probe gases.As expected, most of the clusters were well-separated in the bestcombinations, while those in the worst combination denselyoverlapped, especially between PS and P4MS (Fig. 2c–f). It isassumed that high pattern recognition accuracy can be achievedby a combination of probe gases discriminating polar PCL fromothers and ones discriminating PS, P4MS and PVF from eachother (Table 1; see also Fig. S5, ESI†). It should be noted thatclear separation of clusters does not necessarily lead to highclassification performance.27To evaluate further the applicability of the pattern recogni-tion of solid materials, we demonstrated the identification ofthe molecular weights of polymers. Two additional polystyreneswith different molecular weights, PS(35k) and PS(280k), werealso coated onto separate MSS channels in the same manner.Fig. 2 SVM classification for the identification of the molecular weights of polymers. (a) Dot plots of classification accuracy calculated by SVM with5� 2 cross validation as a function of numbers of combination of probe gases (n) used for the calculation. Histograms show the number of combinations(top) and classification accuracy (right). (b) Dot plots of classification accuracy with the combinations of 2 probe gases. The best and worst casesin accuracy are shown beside the plot. (c–f) PCA score plots for each combination of two probe gases, which result in the best or worst accuracy asfollows: the best combinations ethyl acetate/ethanol (c), ethyl acetate/chloroform (d) and ethyl acetate/toluene (e); and the worst combination ethylacetate/propionic acid (f).Materials Horizons CommunicationOpen Access Article. Published on 19 December 2018. Downloaded on 3/19/2019 4:26:47 AM.  This article is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported Licence.View Article Onlinehttp://creativecommons.org/licenses/by-nc/3.0/http://creativecommons.org/licenses/by-nc/3.0/http://dx.doi.org/10.1039/c8mh01169aThis journal is©The Royal Society of Chemistry 2019 Mater. Horiz., 2019, 6, 580--586 | 583Using these MSS channels, their responses to the probe gaseswere measured. The same feature sets were extracted and com-bined with the dataset of previously measured PS(350k) andP4MS. The PCA and LDA were conducted using the 12 probegases. Although the PCA score plots resulted in some misclassi-fication, especially between PS and P4MS, the LDA providedclear discrimination of polystyrenes including PS and P4MS interms of molecular weight (Fig. 3). According to a previousstudy,13 a response of a nanomechanical sensor is stronglyaffected by the physical properties of a receptor layer, includingYoung’s modulus. The Young’s moduli of polystyrene thinfilms used in this study are reported to be in the range from3.4 to 3.9 GPa.28 Thus, it is found that the current patternrecognition approach can discriminate materials with such anarrow range of Young’s moduli. The SVM classification wasalso performed with all combinations of the 12 probe gases.As shown in Fig. 4 and Table 1, 312 combinations (7.6%) canidentify the differences in molecular weight with 100% accuracy,Table 1 Summary of SVM classification. (Usage rate columns) Usage rate of each probe gas achieving 100% accuracy in the identification of polymersand molecular weight in all combinations. (Accuracy columns) Identification accuracy using one feature set from each probe gas. Each identificationaccuracy is shown as a mean value � standard deviation of accuracy obtained in cross validationProbe gasUsage rate AccuracyPolymer (%) Mol. weight (%) Polymer (%) Mol. weight (%)Water 82.1 53.5 58.6 � 6.3 65.4 � 15.7Ethanol 59.1 51.0 82.0 � 7.7 78.4 � 16.91-Hexanol 34.6 51.9 78.2 � 8.8 82.0 � 10.8Hexanal 56.1 62.2 91.4 � 9.7 86.8 � 12.0n-Heptane 51.6 67.9 85.8 � 8.3 89.1 � 8.7Methylcyclohexane 46.2 63.1 87.0 � 3.2 78.3 � 3.7Toluene 58.5 27.2 84.5 � 8.0 91.5 � 3.9Ethyl acetate 70.9 78.5 94.0 � 3.8 80.7 � 15.8Acetone 52.8 97.4 83.7 � 11.8 88.3 � 6.6Chloroform 65.6 14.7 99.3 � 1.4 71.1 � 13.8Aniline 46.2 49.7 60.2 � 8.8 77.9 � 6.5Propionic acid 60.4 53.8 84.5 � 15.7 78.8 � 10.5No. of 100% combinations 1021 312Total no. of combinations 4095 = (212 � 1)Fig. 3 Identification of the molecular weights of polymers. (Top left, top right, bottom left) PCA score plots of the 4 polymers coated on MSS using the12 different probe gases. (Bottom right) LDA score plots of the polymers. PS(350k), polystyrene, Mw = 350 000 (blue); P4MS, poly(4-methylstyrene),Mw = 72 000 (black); PS(280k), polystyrene, Mw = 280 000 (red); PS(35k), polystyrene, Mw = 35 000 (green).Communication Materials HorizonsOpen Access Article. Published on 19 December 2018. Downloaded on 3/19/2019 4:26:47 AM.  This article is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported Licence.View Article Onlinehttp://creativecommons.org/licenses/by-nc/3.0/http://creativecommons.org/licenses/by-nc/3.0/http://dx.doi.org/10.1039/c8mh01169a584 | Mater. Horiz., 2019, 6, 580--586 This journal is©The Royal Society of Chemistry 2019while the largest feature set extracted from all 12 probe gasesresulted in a lower accuracy of 95.0 � 0.10% (see also Fig. S6,ESI†).27,29 It should be noted that a specific combination of twoprobe gases, i.e. [chloroform, aniline], achieved 100% accuracy(Fig. 4b). Furthermore, even one specific probe gas, i.e. toluene,achieved 91.5� 3.9% identification accuracy (Table 1), and mostPCA score plots of each probe gas provide rough discriminationof the differences in molecular weights (Fig. S7, ESI†). Theseresults indicate that the pattern recognition can also be effec-tively applied to solid materials to identify each analyte even withsimilar chemical and physical properties by choosing a couple ofappropriate probe gases. As demonstrated in a previous study,15these patterns should be correlatable with other material para-meters, such as Young’s moduli, leading to quantitative pre-diction of such parameters using machine learning-basedregression analyses.This approach will also be effective in the industrial phase.In the industrial phase, for example, it is quite important toassess the quality of sensor products, especially the coatingquality of receptor materials. As a proof-of-concept, we assessedthe coating quality of the receptor layers of MSS through thispattern recognition-based approach. Sensing signals recordedfrom 11 different sensors are analysed by PCA, followed by aquality evaluation based on Mahalanobis distances.30–32As shown in Fig. S8 and S9 in the ESI,† it is possible toquantitatively assess the coating quality of PVF and PCL,respectively. Thus, this approach will provide various possibilitiesin the industrial phase.ConclusionsIn this study, it has been demonstrated that solid specimenscould be identified by pattern recognition, achieving 100%accuracy in the case of some specific combinations of probegases. As proved by means of SVM as well as PCA and LDA, evenslight differences in material properties such as molecularweights could also be discriminated through this approach.Since any kind of gaseous or volatile molecules can be poten-tially utilized as a probe for this pattern recognition-based solidmaterial identification, this approach possesses unlimited pos-sibilities to differentiate solid materials. The potential targetsolid materials of this concept include inorganic nanoparticles,functional organic materials and biomolecules such as pep-tides, proteins and nucleic acids. It should also be noted thatthis concept is not limited to nanomechanical sensors butcan be expanded to a variety of chemical sensors. Moreover,this approach will also be effective in the industrial phase,including the assessment of receptor coatings on sensor pro-ducts as demonstrated in this study. Therefore, the presentedconcept of the pattern recognition-based analysis of solid materialswill open a new horizon for chemical sensors and materialsscience.ExperimentalMaterialsPolystyrene (Mw = 35 000) (PS(35k)), polystyrene (Mw = 280 000)(PS(280k)), polystyrene (Mw = 350 000) (PS(350k)), polycaprolactone(PCL), poly(4-methylstyrene) (P4MS), and poly(vinylidene fluoride)(PVF) were purchased from Sigma Aldrich, and used in this study.N,N-Dimethylformamide (DMF) as a solvent to prepare polymersolutions for inkjet spotting was purchased from Wako PureChemical Industries. Ethanol, 1-hexanol, hexanal, n-heptane,methylcyclohexane, toluene, ethyl acetate, acetone, chloroform,aniline and propionic acid (analytical or higher grade) used asprobe gases were purchased from Sigma-Aldrich, Tokyo ChemicalIndustry, and Wako Pure Chemical Industries. All chemicals wereused as purchased. To obtain water vapour, MilliQ water wasused (Merck MilliPore).Experimental procedureTo identify each polymer material, we coated polymer layersonto an MSS chip by inkjet spotting. The detailed fabricationprocess of the MSS chip has been provided in previousreports.21,22 An inkjet spotter (LaboJet-500SP, MICROJETCorporation) equipped with a nozzle (IJHBS-300, MICROJETCorporation) was used. Each polymer was dissolved in DMF(1 mg mL�1), and the resulting solutions were deposited ontoeach channel of the MSS. The injection speed, volume of adroplet and number of inkjet shots were fixed at B5 m s�1,Fig. 4 SVM classification for the identification of the molecular weights ofpolymers. (a) Dot plots of classification accuracy calculated by SVM with5 � 2 cross validation as a function of the number of gases (n) used for thecalculation. Histograms show the number of combinations (top) andclassification accuracy (right). (b) Dot plots of classification accuracy with2 probe gases. The best and worst cases in accuracy are shown besidethe plot.Materials Horizons CommunicationOpen Access Article. Published on 19 December 2018. Downloaded on 3/19/2019 4:26:47 AM.  This article is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported Licence.View Article Onlinehttp://creativecommons.org/licenses/by-nc/3.0/http://creativecommons.org/licenses/by-nc/3.0/http://dx.doi.org/10.1039/c8mh01169aThis journal is©The Royal Society of Chemistry 2019 Mater. Horiz., 2019, 6, 580--586 | 585B300 pL, and 300 shots, respectively. A stage of the inkjetspotter was heated at 80 1C to dry DMF. Each polymer wascoated at least 2 different channels to investigate the coatingquality: PS(35k), N = 3; PS(280k), N = 2; PS(350k), N = 11; PCL,N = 11; P4MS, N = 11; PVF, N = 11.The coated MSS chips were mounted in a Teflon chamber,which was placed in an incubator (Incubator-1) with a con-trolled temperature of 25.0 � 0.5 1C. The chamber was con-nected to a gas system consisting of two mass flow controllers(MFCs), a mixing chamber, a purging gas line and a vial for asolvent liquid in an incubator (Incubator-2) with a controlledtemperature of 15.0 � 0.5 1C. The vapour of each solvent wasproduced by bubbling of carrier gas. Pure nitrogen gas wasused as carrier and purging gases. The total flow rate was keptat 100 mL min�1 during the experiments. The concentrations ofthe 12 different solvent vapours were controlled using MFC-1 atPa/Po of 0.1, where Pa and Po stand for the solvent’s partialvapour pressure and saturated vapour pressure, respectively.Before measuring MSS signals, pure nitrogen gas was intro-duced into the MSS chamber for 1 min. Subsequently, MFC-1(sampling line) was switched on/off every 10 s with a controlledtotal flow rate of 100 mL min�1 using MFC-2 for 5 cycles (Fig. S2,ESI,† for full signal responses). Data were measured with a bridgevoltage of �0.5 V, and recorded with a sampling rate of 10 Hz.The data collection program was designed using LabVIEW(National Instruments Corporation).Pattern recognition procedurePrincipal component analysis (PCA) and linear discriminantanalysis (LDA) were utilized for reducing the dimensionality ofthe datasets. By projecting the data onto a lower-dimensionalspace, one can visually recognize the coated materials accordingto the cluster separation. To identify each coated material andevaluate its accuracy, we developed classification models basedon a support vector machine (SVM) with a radial basis functionkernel. To assess the quality of sensor products, we used theMahalanobis–Taguchi system (MTS). Feature sets were extractedfrom each decay curve of a normalized MSS signal measuredwith the 12 different gases (Fig. 1b). Three different slopes mij(tn)were extracted from the i-th channel with the j-th gas accordingto the following equation: mij(tn) = [Iij(t0) � Iij(t0 + tn)]/tn, whereIij(t) and t0 denote the signal output at time t and the time whenthe signal response starts to decay, respectively. In this study, wechose 3 time points for tn; tn = 0.5, 1, 1.5 (s). Three sets ofparameters [mij(0.5), mij(1), mij(1.5)] were extracted from the last3 signal responses from 40 to 100 (s) out of the 5 signalresponses in each measurement, because the latter signalresponses can provide more reproducible signal responses thanthe former ones, which exhibit initial fluctuations associatedwith mixing of sample gases and pre-adsorbed gases.PCA and LDA were adopted using scikit-learn packages forPython. PCA projects data onto lower dimensions so that thevariance of the first principal component (PC1) becomes thelargest. Successive principal components are determined tomaximize the variance under the constraint that the (n + 1)-thprincipal component is orthogonal to the n-th component.In contrast to PCA, LDA projects data onto lower dimensionsto maximize the cluster separation; LDA maximizes the dis-tance between the classes and minimizes the variance in thesame class. Classification models based on a non-linear SVMwere developed using scikit-learn packages for Python. To opti-mize and evaluate the models, we employed 5 � 2 cross valida-tion. The whole datasets were first split into 5 datasets, of which4 datasets were used as training datasets, and the remaining1 dataset was used as a test dataset. The training datasets werefurther split into 2 sub-datasets. Based on these sub-datasets,the hyperparameters of the SVM (i.e. C and g) were optimized.This validation process was repeated for all the combinations ofthe 5 datasets for evaluating the classification accuracy of themodels.The MTS was adopted by scikit-learn packages for Python.To evaluate the Mahalanobis distance, the data were projectedonto a lower-dimension space (PC1–2 plane) by PCA. Then, theMahalanobis distances of each plot on the PC1–2 plane werecalculated by MTS.Conflicts of interestThere are no conflicts to declare.AcknowledgementsWe thank Yuko Kameyama, Keiko Koda and Eri Sakon(WPI-MANA, NIMS) for coating of the receptor layers and datacollection. This work was supported by the MSS alliance; JSTCREST (JPMJCR1665); a Grant-in-Aid for Scientific Research (A),18H04168, MEXT, Japan; a Grant-in-Aid for Young Scientists,18K14133, MEXT, Japan; the Public/Private R&D InvestmentStrategic Expansion Program (PRISM), Cabinet Office, Japan;the Leading Initiative for Excellent Young Researchers, MEXT,Japan; the Center for Functional Sensor & Actuator (CFSN), NIMS;and the World Premier International Research Center Initiative(WPI) on Materials Nanoarchitectonics (MANA), NIMS.References1 K. J. Albert, N. S. Lewis, C. L. Schauer, G. A. Sotzing,S. E. Stitzel, T. P. Vaid and D. R. Walt, Chem. Rev., 2000,100, 2595–2626.2 U. Lange, N. V. Roznyatovskaya and V. M. Mirsky, Anal.Chim. Acta, 2008, 614, 1–26.3 A. Lv, Y. Pan and L. Chi, Sensors, 2017, 17, 213.4 H. Bai and G. Shi, Sensors, 2007, 7, 267–307.5 J. Gutiérrez and M. C. Horrillo, Talanta, 2014, 124, 95–105.6 A. Wilson, M. Baietto, A. D. Wilson and M. Baietto, Sensors,2009, 9, 5099–5148.7 A. D. Wilson, M. Baietto, A. D. Wilson and M. Baietto, Sensors,2011, 11, 1105–1176.8 A. D. Wilson, Proc. Technol., 2012, 1, 453–463.9 G. Konvalina and H. Haick, Acc. Chem. Res., 2014, 47, 66–76.10 R. A. Potyrailo, Chem. Rev., 2016, 116, 11877–11923.11 J.-W. Yoon and J.-H. Lee, Lab Chip, 2017, 17, 3537–3557.Communication Materials HorizonsOpen Access Article. Published on 19 December 2018. Downloaded on 3/19/2019 4:26:47 AM.  This article is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported Licence.View Article Onlinehttp://creativecommons.org/licenses/by-nc/3.0/http://creativecommons.org/licenses/by-nc/3.0/http://dx.doi.org/10.1039/c8mh01169a586 | Mater. Horiz., 2019, 6, 580--586 This journal is©The Royal Society of Chemistry 201912 J. K. Gimzewski, C. Gerber, E. Meyer and R. R. Schlittler,Chem. Phys. Lett., 1994, 217, 589–594.13 G. Yoshikawa, Appl. Phys. Lett., 2011, 98, 173502.14 K. Shiba, T. Sugiyama, T. Takei and G. Yoshikawa, Chem.Commun., 2015, 51, 15854–15857.15 G. Imamura, K. Shiba and G. Yoshikawa, Jpn. J. Appl. Phys.,2016, 55, 1102B1103.16 K. Shiba, R. Tamura, G. Imamura and G. Yoshikawa, Sci.Rep., 2017, 7, 3661.17 I. Osica, G. Imamura, K. Shiba, Q. Ji, L. K. Shrestha, J. P.Hill, K. J. Kurzydłowski, G. Yoshikawa and K. Ariga, ACSAppl. Mater. Interfaces, 2017, 9, 9945–9954.18 H. T. Ngo, K. Minami, G. Imamura, K. Shiba and G. Yoshikawa,Sensors, 2018, 18, 1640.19 K. Minami, K. Shiba and G. Yoshikawa, Anal. Methods, 2018,10, 3720–3726.20 K. Shiba, R. Tamura, T. Sugiyama, Y. Kameyama, K. Koda,E. Sakon, K. Minami, H. T. Ngo, G. Imamura, K. Tsuda andG. Yoshikawa, ACS Sens., 2018, 3, 1592–1600.21 G. Yoshikawa, T. Akiyama, S. Gautsch, P. Vettiger andH. Rohrer, Nano Lett., 2011, 11, 1044–1048.22 G. Yoshikawa, T. Akiyama, F. Loizeau, K. Shiba, S. Gautsch,T. Nakayama, P. Vettiger, N. F. de Rooij and M. Aono,Sensors, 2012, 12, 15873–15887.23 J. Yan, X. Guo, S. Duan, P. Jia, L. Wang, C. Peng, S. Zhang,J. Yan, X. Guo, S. Duan, P. Jia, L. Wang, C. Peng andS. Zhang, Sensors, 2015, 15, 27804–27831.24 S. M. Scott, D. James and Z. Ali, Microchim. Acta, 2007, 156,183–207.25 V. N. Vapnik, Estimation of dependences based on empiricaldata; Empirical inference science: afterword of 2006, Springer,2006.26 C. M. Bishop, Pattern recognition and machine learning,Springer, 2006.27 T. Nowotny, A. Z. Berna, R. Binions and S. Trowell, Sens.Actuators, B, 2013, 187, 471–480.28 J. M. Torres, C. M. Stafford and B. D. Vogt, Polymer, 2010, 51,4211–4217.29 M. Pardo and G. Sberveglieri, Sens. Actuators, B, 2007, 123,437–443.30 P. C. Mahalanobis, Proc. Natl. Inst. Sci. India, 1936, 2, 49–55.31 G. Taguchi and R. Jugulum, The Mahalanobis-TaguchiStrategy: A Pattern Technology System, John Wiley & Sons,Inc., 2007.32 A. Adebanji and S. Nokoe, Effects of Mahalanobis Distanceand Prior Probabilities: On the Performance of the Linear andQuadratic Discriminant Functions: A Monte Carlo Approach,LAMBERT Academic Publishing, 2011.Materials Horizons CommunicationOpen Access Article. Published on 19 December 2018. Downloaded on 3/19/2019 4:26:47 AM.  This article is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported Licence.View Article Onlinehttp://creativecommons.org/licenses/by-nc/3.0/http://creativecommons.org/licenses/by-nc/3.0/http://dx.doi.org/10.1039/c8mh01169a