Functional Nanoparticles-Coated Nanomechanical ... - ACS Publications

Aug 15, 2018 - Graduate School of Frontier Sciences, The University of Tokyo, 5-1-5 Kashiwa-no-ha, Kashiwa, ... through this study that various inform...
0 downloads 0 Views 3MB Size
Download PDF
Article Cite This: ACS Sens. XXXX, XXX, XXX−XXX

pubs.acs.org/acssensors

Functional Nanoparticles-Coated Nanomechanical Sensor Arrays for Machine Learning-Based Quantitative Odor Analysis Kota Shiba,*,†,‡ Ryo Tamura,*,‡,§,∥ Takako Sugiyama,‡ Yuko Kameyama,‡ Keiko Koda,‡ Eri Sakon,‡ Kosuke Minami,‡ Huynh Thien Ngo,‡ Gaku Imamura,†,‡ Koji Tsuda,§,∥,⊥ and Genki Yoshikawa†,‡,#

Downloaded via KAOHSIUNG MEDICAL UNIV on August 16, 2018 at 01:52:31 (UTC). See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

†

Center for Functional Sensor & Actuator (CFSN) and ‡World Premier International Research Center Initiative (WPI), International Center for Materials Nanoarchitectonics (MANA), National Institute for Materials Science (NIMS), 1-1 Namiki, Tsukuba, Ibaraki 305-0044, Japan § Research and Services Division of Materials Data and Integrated System, National Institute for Materials Science, 1-2-1 Sengen, Tsukuba, Ibaraki 305-0047, Japan ∥ Graduate School of Frontier Sciences, The University of Tokyo, 5-1-5 Kashiwa-no-ha, Kashiwa, Chiba 277-8561, Japan ⊥ Center for Advanced Intelligence Project, RIKEN, 1-4-1 Nihombashi, Chuo-ku, Tokyo 103-0027, Japan # Materials Science and Engineering, Graduate School of Pure and Applied Science, University of Tsukuba, 1-1-1 Tennodai, Tsukuba, Ibaraki 305-8571, Japan S Supporting Information *

ABSTRACT: A sensing signal obtained by measuring an odor usually contains varied information that reflects an origin of the odor itself, while an effective approach is required to reasonably analyze informative data to derive the desired information. Herein, we demonstrate that quantitative odor analysis was achieved through systematic material designbased nanomechanical sensing combined with machine learning. A ternary mixture consisting of water, ethanol, and methanol was selected as a model system where a target molecule coexists with structurally similar species in a humidified condition. To predict the concentration of each species in the system via the data-driven approach, six types of nanoparticles functionalized with hydroxyl, aminopropyl, phenyl, and/or octadecyl groups were synthesized as a receptor coating of a nanomechanical sensor. Then, a machine learning model based on Gaussian process regression was trained with sensing data sets obtained from the samples with diverse concentrations. As a result, the octadecyl-modified nanoparticles enhanced prediction accuracy for water while the use of both octadecyl and aminopropyl groups was indicated to be a key for a better prediction accuracy for ethanol and methanol. As the prediction accuracy for ethanol and methanol was improved by introducing two additional nanoparticles with finely controlled octadecyl and aminopropyl amount, the feedback obtained by the present machine learning was effectively utilized to optimize material design for better performance. We demonstrate through this study that various information which was extracted from plenty of experimental data sets was successfully combined with our knowledge to produce wisdom for addressing a critical issue in gas phase sensing. KEYWORDS: nanoparticle, surface functionality, nanomechanical sensing, sensor array, machine learning, odor, quantification, MSS

M

indicates that plenty of information on the origin of odors can be derived through proper analytical approaches, leading to diverse applications in such fields as food, security, medicine, cosmetics, the environment, and so on. To meet the demand for odor analysis, various types of multichannel sensors, a socalled sensor array, have been developed for decades.1−4 In contrast to conventional techniques such as chromatography, a sensor array can directly measure odors without a separation process. In combination with advanced MEMS technology, it

ost substances have their own odors, which are caused by a complex mixture of gaseous molecules. There are more than 400,000 types of molecules that are potential constituents of various odors and are termed “odorants”. Some odors are composed of a few thousands of odorants, and such complicated composition is one of the major reasons why odor analysis, especially quantification of each component, is technically difficult. There are many kinds of odorants which have a similar molecular structure, and thus, it is another challenging issue to quantitatively evaluate them. Furthermore, such molecules coexist with water molecules in ambient conditions to form a more complex and variable gaseous mixture. On the other hand, the compositional complexity © XXXX American Chemical Society

Received: June 1, 2018 Accepted: July 31, 2018

A

DOI: 10.1021/acssensors.8b00450 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors

solutions (Solutions A−E) were prepared. Detailed composition of each solution is summarized in Table S1. Solutions A, B, C, and D were individually flowed in perfluoroalkoxyalkane (PFA: 1.0 mm inner diameter, 1/16 in. outer diameter, product of YMC Co., Ltd.) tubes with a syringe pump (CXN1070, product of ISIS, Co., Ltd.) at 10 mL/min. Solutions A and B and Solutions C and D were mixed respectively in a polytetrafluoroethylene (PTFE) fluidic channel with a Y-shaped junction (the channel cross section of approximately 1 mm2, KeyChem mixer, product of YMC Co., Ltd.). After that, the two resultant reaction solutions, i.e., Solution A + B and Solution C + D, were mixed in the second fluidic channel placed just after the first two fluidic channels. The first and second fluidic channels were connected with 10 cm PFA tubes. Then, the mixture of the four Solutions A-D was flowed through a PFA tube with 70 cm in length and was added into Solution E under magnetic stirring. After the addition, the final reaction solution was aged at room temperature for a certain period a few months or moredepending on the type of silane coupling reagents. Silica nanoparticles (SNPs) functionalized with aminopropyl or phenyl groups were synthesized by the reported procedure with minor modifications. Detailed composition of each solution is summarized in Table S2. Unlike STNPs, only Solutions A and B were mixed in a single step without further mixing with Solutions C, D, and E. The mixture of Solutions A and B was collected in a glass vial and then was aged at room temperature for a few months, although such a long reaction period is unnecessary to achieve sufficiently high yield. To prepare a powder sample for various characterizations, typically 20 mL of the suspension was centrifuged (9000 rpm, 10 or 20 min, depending on a sample). The precipitate was washed with IPA repeatedly and then dispersed in EtOH. The EtOH suspension was dried at 60 °C to obtain powder sample. Ph(1)−OH(3)−SNPs suspension was remarkably stable so that the NPs did not sediment. For this reason, the suspension was directly dried without any washing to obtain a powder sample. Characterization. Fourier Transform-infrared (FT-IR) spectra were measured using a Nicolet 4700 FT-IR spectrometer (Thermo Fisher Scientific Inc.) at the resolution of 2.0 cm−1 and in the range from 4000 to 400 cm−1. The sample powder was homogeneously mixed with KBr, and then the mixture was pressed to form a KBr disk for the transmission measurements. Scanning electron microscope (SEM) images were obtained using a Hitachi Ultrahigh Resolution Scanning Electron Microscope SU8000 at the accelerating voltage of 10 kV. Prior to each measurement, samples were coated with a few nanometers of platinum. Thermogravimetric-differential thermal analysis (TG-DTA) curves were recorded on a SII EXSTAR 6000 TG/DTA6300 at the heating rate of 10 K/min under air flowing. α-Alumina powder was used as a reference material to obtain DTA curves. Spray Coating of Various NPs onto MSS. NH2(1)−OH(3)− SNPs, NH 2 −STNPs, C18(1)−NH 2 (1)−STNPs, C18(0.25)− NH2(1)−STNPs, and C18(1)−NH2(0.25)−STNPs were spraycoated onto the surface of MSS by using a spray coater (rCoater, product of Asahi Sunac Co.) after preparing suspensions. For preparation of the NP suspensions, all the functional NPs except NH2(1)−OH(3)−SNPs were centrifuged at 9000 rpm for 20 min. The sediment was carefully washed with IPA several times and then the IPA/water mixture (vol/vol = 3/5) was added. The concentration of the suspensions was set at approximately 1 g/L. Before spraycoating, the suspensions were fully ultrasonicated to get the NPs dispersed as much as possible (some aggregates were still recognized). For NH2(1)−OH(3)−SNPs, the suspension was prepared by using MeOH instead of IPA. As-synthesized NH2(1)−OH(3)−SNPs were mixed with MeOH without any washing processes to have a basic MeOH/water suspension (vol/vol = 3/5) with the concentration of approximately 1 g/L. Then, the suspension was loaded in a syringe and was flowed through a PTFE tube at 3 mL/min by using a syringe pump (YSP201, product of YMC Co., Ltd.). The suspension was introduced into a spray nozzle and then was sprayed with the help of two types of carrier air (atomizing air: 0.030 MPa, patterning air: 0.030 MPa) to

is possible to fabricate a compact device with multiple sensing channels, being useful for analyzing complex samples. To derive quantitative information from odors, the analysis of a data set obtained with sensing tests is of critical importance. Although several approaches including principal component regression and neural networks have been applied to quantify the constituents of odors based on the data sets, there are still some challenges in a gaseous mixture consisting of a few components with ppm or ppb level concentration.5−12 In a practical situation, odors have to be measured in the presence of humidity, which usually reaches a few percent of volume fraction. Thus, desired information needs to be obtained in such a condition. For this reason, each channel of a sensor array has to have specific properties in terms of hydrophilicity/hydrophobicity to precisely quantify target species in the presence of large amounts of water molecules. Therefore, flexibly designed sensor channels with controlled properties are indispensable, while ready-made sensor arrays without any systematic optimization have been used so far. In this paper, functional nanoparticles (NPs) with systematically optimized surface properties are prepared to demonstrate their use as a sensing material for quantifying concentration of target species coexisting with model contamination gases including water. Aminopropyl, hydroxyl, phenyl, and octadecyl groups were immobilized on an NP surface with various ratios to control hydrophilicity/hydrophobicity as well as to realize different affinity to target species. The NPs were coated on each channel of a nanomechanical Membrane-type Surface stress Sensor (MSS) array.13,14 It was confirmed that the NPs composed of silica−titania hybrid15 worked as an effective receptor material for nanomechanical sensing. We have already reported that specific information such as alcohol content was estimated with high accuracy from the odor of a liquor that usually contains various flavors.16 For more detailed investigation toward practical applications, here we focus on a mixture of water, ethanol, and methanol as a model system and quantify concentration of all the species based on odor. Combining the controlled sensing properties of the NPs which have finely tuned surface properties with a datadriven knowledge obtained by a machine learning technique (e.g., a regression model based on Gaussian process), we demonstrate that the concentrations of not only a target but also coexisting contamination gases are precisely predicted.

■

EXPERIMENTAL SECTION

Chemicals. Tetraethoxysilane (TEOS: Tokyo Chemical Industry Co., Ltd., > 97.0%), 3-aminopropyltriethoxysilane (APTES: Sigma, Inc., > 98%), octadecyltriethoxysilane (ODTES: Tokyo Chemical Industry Co., Ltd., > 85.0%), trimethoxyphenylsilane (TMPS: Tokyo Chemical Industry Co., Ltd., > 98.0%), triethoxyphenylsilane (TEPS: Tokyo Chemical Industry Co., Ltd., > 99.0%), titanium tetraisopropoxide (TTIP: Tokyo Chemical Industry Co., Ltd., purity N/A), methanol (MeOH: Kanto Chemical Co., Inc., > 99.8%), ethanol (EtOH: Wako Pure Chemical Industries, Ltd., > 99.5%), isopropyl alcohol (IPA: Wako Pure Chemical Industries, Ltd., > 99.7%), 1,1,2,2tetrachloroethane (TCE: Wako Pure Chemical Industries, Ltd., > 97.0%), aqueous ammonia solution (NH3aq: Kanto Chemical Co., Inc., 28.0−30.0%), N,N-dimethylformamide (DMF: Kanto Chemical Co., Inc., > 99.5%) and octadecylamine (ODA: Aldrich, Inc., 97.0%) were utilized in the present study. All the chemicals were used as received. Material Preparation. Silica−titania hybrid nanoparticles (STNPs) with various surface functionalities were synthesized by a multistep nucleation-controlled growth method15 which we reported previously with some minor modifications. Briefly, five starting B

DOI: 10.1021/acssensors.8b00450 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors form homogeneous droplets. An MSS chip was mounted on a stage which was heated at approximately 100 °C to quickly evaporate the droplets. The stage was moved back and forth, while the spray nozzle was also moved from left to right at 15 mm/s with 0.3 mm pitch. The distance between the spray nozzle and stage was set at 75 mm. The coating process was repeated to obtain coating thickness of a few microns. To avoid any cross contamination, a mask was used to cover three channels when one channel was in a coating process. Coating with Inkjet. C18−STNPs, Ph−STNPs, and Ph(1)− OH(3)−SNPs were deposited onto the surface of MSS by inkjet spotting. An inkjet spotter (LaboJet-500SP) and a nozzle (IJHBS300) were purchased from the MICROJET Corporation. Three types of NPs were dispersed in different media; C18−STNPs was dispersed in TCE, Ph−STNPs was dispersed in DMF and Ph(1)−OH(3)− SNPs was dispersed in DMF containing a small amount of ammonia, MeOH, and water, respectively. C18−STNPs and Ph−STNPs were washed with IPA several times as described above and then dispersed in each medium with the concentration of approximately 1 g/L. For Ph(1)−OH(3)−SNPs, the as-synthesized NPs suspension was diluted with DMF ten times to be available for inkjet. A certain amount of droplets was dropped on an MSS channel by the inkjet to form a receptor layer coating. The numbers of shots were 400, 200, and 300 for C18−STNPs, Ph−STNPs, and Ph(1)−OH(3)−SNPs, respectively. A stage of the inkjet was heated at 60 °C for TCE and 80 °C for DMF to control evaporation. Detailed Procedure and Conditions for the Sensing Experiments. The vapor sensing tests were performed with the experimental setup described here. The MSS coated with various NPs was mounted in a chamber and the chamber was carefully sealed with O-rings. Two mass flow controllers (MFCs; FCST1005C-4F2-F100N2, purchased from Fujikin Inc.) were utilized to introduce nitrogen into the chamber at a flow rate of 100 mL/min. One MFC was for purging (i.e., accelerating desorption of adsorbents), and the other one was for introducing the sample vapor together with nitrogen as a carrier. In the present case, 1 mL of sample liquid (mixture of water, ethanol, and methanol) was added into a small vial capped with a rubber lid and two needles connected to PTFE tubes were stuck into the headspace of the vial through the rubber lid. One end of the PTFE was connected to MFC and the other end of the PTFE tube was connected to a vacant vial, so-called “mixing vial”, to make the mixed gas sample homogeneous. Another PTFE tube stuck into the mixing vial was connected to the chamber. Another MFC and vacant vial were set in the same way and connected to the mixing vial. The two MFCs were switched every 30 s to perform a sample introduction− purging cycle. This cycle was repeated four times and the data were recorded at the bridge voltage of −0.5 V and a sampling rate of 20 Hz. The data collection program was designed by LabVIEW (National Instruments Corporation). All the experiments were conducted under ambient conditions without any temperature/humidity control. To remove adsorbed species and get the receptor coatings back to a similar state, water vapor was flowed into the chamber at 100 mL/min for 30 s together with nitrogen as a carrier, followed by nitrogen purging at 100 mL/min for 30 s as described. This cycle was also repeated four times after every sample measurement. Features of Each Response in a Signal. As features of a signal in machine learning, the following parameters were extracted from a response in a signal measured by the MSS:

Parameter 1:(b − a)/(tb − ta)

(1)

Parameter 2:(c − b)/(tc − tb)

(2)

Parameter 3:(d − c)/(td − tc)

(3)

Parameter 4:(e − a)

(4)

Figure 1. Schematic of a feature extraction from a response in a signal. Four parameters are defined as features by using a, b, c, d, e, ta, tb, tc, and td. the information on the adsorption capacity. In the present analysis, in accordance with our previous work,16 the latter three responses where ta = 90, 150, and 210 in a signal were treated independently, and three sets of the parameters were extracted from a signal. As the latter three responses are not exactly independent, we considered another two types of features: (i) features generated from the mean of three responses and (ii) features generated from three responses such as the change of outputs between responses. The simple analyses using these features were performed; however, a critical improvement of prediction was not achieved. Thus, we decided to use the conventional features in this study. Definition of Cross Validation Error. In this paper, the performance of machine learning model is evaluated by a prediction accuracy to prevent overfitting. As a prediction accuracy, we adopted the cross validation error whose definition is as follows. First, let D = {Xk, yk}k=1,···,N be the training data set where the number of data points is N = 54 which was generated from three responses from 18 training samples. Here, the feature vectors {Xk}k=1,···,N are normalized by using z-score. The data set D was randomly divided into L data subsets. In this paper, we fixed L = 18. Each data subset is expressed by Dl which is labeled by l = 1,···,L, and the number of data points in each data subset is N/L. One of the L data subsets is regarded as validation data set, while the remaining L − 1 data subsets are used as training data set. The number of validation data points and that of the training data points are given by Nva = N/L and Ntr = N(L − 1)/L, respectively. Next, for each data subset Gl = D\Dl consisting of Ntr data, we performed Gaussian process and the predicted concentration y*(l)(X) was estimated. We calculated the mean-absolute error between the real concentration in Dl and predicted concentration: Δ(l) =

1 Nte

∑ k ∈ Dl

|yk − y*(l)(X k)|

(5)

Furthermore, the cross validation error (CVE) is obtained by averaging L different mean-absolute errors, which is defined as

Δ=

1 L

L

∑ Δ(l) l=1

(6)

Using this CVE, we discussed the prediction accuracy of the machine learning model depending on the combinations of channels. Note that when the concentrations of the three test samples are predicted, we used the machine learning model with Gaussian process regression for all training data, that is, D. Gaussian Process Regression in COMBO. By Gaussian process regression,17 a prediction value with uncertainty in any feature vector is estimated by the prepared training data sets. Since regularization is included in the Gaussian process, overfitting is prevented by properly choosing the values of hyperparameters. In our analysis, the Gaussian process was trained using the package called COMBO.18 In COMBO, by using a random feature map,19 the Gaussian process is approximated by the Bayesian linear model. The hyperparameters in the Gaussian process are automatically searched so that the type-II

where a, b, c, d, e, ta, tb, tc, and td are denoted in Figure 1. In this case, tb = ta + 2 [s], tc = ta + 30 [s], and td = ta + 32 [s] were used. Parameters 1 and 3 should include the information on the adsorption and desorption processes, respectively. Furthermore, parameter 2 should reflect a quasi-equilibrium state, and parameter 4 should have C

DOI: 10.1021/acssensors.8b00450 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors likelihood20 reaches a maximum. Thus, by using COMBO, the machine learning model with high prediction accuracy can be trained based on the Gaussian process. Note that it was confirmed that the overfitting is prevented in our analyses by focusing on the prediction accuracy depending on the number of channels (see Figure S1).

■

RESULTS AND DISCUSSION Synthesis and Characterization of Functional NPs. A variety of NPs with different functionalities were synthesized by sol−gel reaction of silicon and titanium alkoxides (experimental details are described in SI). Briefly, SNPs functionalized with aminopropyl groups (denoted as NH2(1)− OH(3)−SNPs) were obtained through cohydrolysis and condensation reaction of tetraethoxysilane (TEOS) and 3aminopropyltriethoxysilane (APTES) in aqueous ammonia solution. Phenyl-functionalized SNPs (denoted as Ph(1)− OH(3)−SNPs) were synthesized in the same manner using triethoxyphenylsilane instead of APTES. Silica−titania hybrid NPs with various surface properties were prepared by the microfluidic approach reported previously.15 APTES, octadecyltriethoxysilane (ODTES), and trimethoxyphenylsilane were hydrolyzed in the presence of TEOS, resulting in the formation of aminopropyl-functionalized STNPs (denoted as NH2− STNPs), octadecyl-functionalized STNPs (denoted as C18−STNPs), and phenyl-functionalized STNPs (denoted as Ph−STNPs), respectively. Furthermore, STNPs dual-functionalized with aminopropyl and octadecyl groups (denoted as C18−NH2−STNPs) were synthesized through cohydrolysis and condensation reaction of APTES, ODTES, and TEOS. In the case of nanomechanical sensing, physical properties of coating materials, especially Young’s modulus, significantly affect response intensity.21 The STNPs were reported to have core−shell structure with a few nanometers of titania core and tens of nanometers of silica shell.15 Thus, STNPs and SNPs should have similar Young’s moduli, which do not have a significant effect on the sensing properties. Accordingly, in the present study, sensing signals can be simply attributed to the surface chemical properties of the NPs. SEM observation revealed that almost all the NPs have similar sizes in the range from 10 to 30 nm while NH2(1)− OH(3)−SNPs showed the average size of approximately 100 nm (Figure 2). In all cases, no products where the size and/or shape were apparently different were seen, indicating that most of the silicon and titanium sources were consumed to form the hybrid NPs immobilized with various functional groups. Judging from FT-IR spectra shown in Figure 2, the presence of aminopropyl groups, octadecyl groups, and phenyl groups is evident because specific absorption bands that are characteristic of each functional group are seen in each spectrum. More specifically, the NH2 vibration mode of polyaminosiloxane is observed for NH2(1)−OH(3)−SNPs and NH2−STNPs at around 1575 cm −1 . 15 In addition, two characteristic absorptions at 1635 and 1489 cm−1 are attributed to the asymmetric and symmetric bending modes of an NH2 group. The FT-IR spectra of C18-STNPs and C18(1)−NH2(1)− STNPs showed intense absorption bands at around 2916 and 2848 cm−1 that can be ascribed to the C−H stretching vibrations of the octadecyl groups.22 Similar bands are also observed for the other four NPs since they have C−H-based structures in the form of an aminopropyl group or a phenyl group. The absorption bands appeared at around 1430 and 738 cm−1 for Ph-STNPs and Ph(1)−OH(3)−SNPs come from a phenyl group covalently attached to silicon, while the

Figure 2. SEM images and FT-IR spectra of the six types of NPs synthesized in the present study. [1] NH2(1)−OH(3)−SNPs, [2] NH2−STNPs, [3] C18(1)−NH2(1)−STNPs, [4] C18−STNPs, [5] Ph−STNPs, and [6] Ph(1)−OH(3)−SNPs.

absorption bands appeared at around 1594, 1571, 1490, 1067, 1027, and 694 cm−1 stem from various vibration modes of the phenyl group itself.23 The composition and yield of the NPs were then determined by taking account of TG-DTA results (Figure S2) and the amount of samples collected from the reaction solutions (Table 1). As the desorption of physically adsorbed water and Table 1. Details of the Six Types of NPs Synthesized in the Present Study

[1] [2] [3] [4] [5] [6]

sample name

size (nm)

molar Si/Ti

ideal molar Si/Ti

yield (%)

NH2(1)−OH(3)−SNPs NH2−STNPs C18(1)−NH2(1)−STNPs C18−STNPs Ph−STNPs Ph(1)−OH(3)−SNPs

100 10 20 30 20 10

N/A 0.72 3.5 2.3 7.5 N/A

N/A 4.1 4.1 2.7 4.0 N/A

94 44 94 98 59 95

the decomposition of the organic functional groups occur in the temperature range from 25 to 150 °C and 150 to 600 °C, respectively, the sample yields were estimated by considering the weight of residues calcined at 600 °C. It should be noted that a phenyl group covalently bonded with silicon is thermally stable even around 600 °C.24 Thus, the weight of residues calcined at 700 °C was utilized for Ph(1)−OH(3)−SNPs and Ph−STNPs. For SNP series, both NPs were collected with the yield of 95%, suggesting that most of the silicon sources were converted to NPs. In the case of STNPs, especially C18− STNPs and C18(1)−NH2(1)−STNPs, the yields were approximately 95% or higher, and the molar Si/Ti ratios D

DOI: 10.1021/acssensors.8b00450 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors

Figure 3. Response intensity of the six types of NPs to the vapor of 21 samples. The composition of each sample is listed in the table below. The three white bars in the right of each graph are those used as a test data set, and the other black bars are those used as a training data set in our machine learning model.

Figure 4. Plots of CVE in relation to the presence of a specific functional group such as hydroxyl, phenyl, octadecyl, and aminopropyl group (top) and the combination of two functional groups (bottom). The parity plots of real and predicted concentrations for each case are shown in Figures S10 and S11.

were similar to those of ideal values. In contrast, NH2−STNPs and Ph−STNPs showed moderate yields of around 50%, and these values did not dramatically increase even though the reaction time was extended, meaning that the reaction has reached an equilibrium state. In any case, since six types of well-defined NPs with different functionalities were obtained, these NPs were used as a receptor layer of the MSS for the vapor measurements. NPs Coating for Vapor Measurements. The coating of the NPs was performed by spraying and inkjet spotting

approaches (Figure S3). The NPs that can form a stable suspension such as C18−STNPs, Ph−STNPs, and Ph(1)− OH(3)−SNPs were coated by inkjet spotting, while those that easily aggregate and/or sediment in solution were coated by spraying (Figure S3). Despite the coating looking rough in the macroscopic viewpoint, the sensor surface was fully covered by spraying. By applying inkjet spotting, the coating was formed more locally on the surface. Being different from the case with a cantilever, MSS is more robust in terms of coating inhomogeneity on its surface25 so that the present inkjet E

DOI: 10.1021/acssensors.8b00450 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors

respectively. According to previous reports, 4-nonylphenol was selectively adsorbed onto the interlayer space of octosilicate dual-functionalized with octadecyl and phenyl groups by cooperative interaction.27,28 In the present case, a similar but different “cooperative” phenomenon was observed in the datadriven analysis. In other words, CVE largely decreased when a machine learning model was trained by features extracted from both C18(1)−NH2(1)−STNPs and NH2−STNPs. To clarify dominant factors for the lowest CVE, further analysis was conducted by calculating additional CVEs that were derived using the combination of two features; one from C18(1)−NH2(1)−STNPs and the other from NH2−STNPs (Table S3). Interestingly, the combination of parameter 1, which correlates with adsorption process, provides low CVEs while parameter 1 itself provides quite high CVEs in most cases. This result indicates that, again, two adsorption processes work cooperatively to decrease CVE. From an experimental point of view, the coexistence of octadecyl and aminopropyl groups is expected to induce hydrophobic interaction with an alkyl group of alcohol and hydrogen bonding with a hydroxyl group of alcohol, respectively, being regarded as a prerequisite to realize selective adsorption. Considering that the combination of C18−STNPs and NH2− STNPs provided quite large CVEs than the case of the lowest CVE (Table S4), fine-tuning of the NP surface significantly affects prediction accuracythis will be discussed later. In any case, the actual responses measured by each of the two NPs potentially contain specific information relating to the molecular structure of the two alcohols. Thus, the present decrease in CVE would be achieved by using experimental data collected from the NPs with proper functionalities. Addition of C18 and NH2 Dual-Functionalized NPs with Finely Controlled Composition for Optimized Prediction Accuracy. In principle, the prediction accuracy of a machine learning model is enhanced with the number of effective features, and the decrease in CVE was achieved by the combination of NPs in the present analyses (see Figure S1). For this reason, all the features obtained by six types of NPs were used to train a machine learning model. Consequently, lower CVEs (Figure 4, red dotted lines) were obtained than that of one or two combinations of NPs, as expected (parity plots are shown in Figure S12). Although the prediction accuracy already seems to be quite high, we explored possibility to decrease CVE further, especially for ethanol and methanol, by introducing properly designed additional NPs. As discussed, the results of the machine learning implied that the coexistence of octadecyl and aminopropyl groups dominantly affected CVE for ethanol and methanol. Accordingly, for the experimental verification of the implication obtained by machine learning, two more NPs dual-functionalized with octadecyl and aminopropyl groups were synthesized and utilized for the concentration quantification. These NPs were synthesized in the starting solution with different molar ratios of octadecyl and aminopropyl precursors, and resultant NPs are denoted as C18(4)−NH2(1)−STNPs and C18(1)−NH2(4)−STNPs, respectively. SEM images are shown in Figure 5. The size of the NPs was around 20 to 30 nm, which is similar to the other NPs. The yields estimated by TG-DTA results (Figure S13) and collected amount of samples were 76% and 24% for C18(4)−NH2(1)−STNPs and C18(1)−NH2(4)−STNPs, respectively. Despite the incomplete reaction, there are clear differences in their FT-

coatings are enough to be used for the following vapor measurements. After the coating, 21 liquid samples consisting of water, ethanol, and methanol with various concentrations (v/v%) were measured. The detailed composition of each sample is listed in Figure 3. All the measured signals are shown in Figures S4−S9. The measurement was performed by alternatively injecting the sample vapor and nitrogen. This injection cycle was repeated four times. Response intensity was extracted from the sensor response of the fourth injection cycle in each signal, and the intensity values are summarized in Figure 3. Apparently, all six types of NPs exhibited different response trends, indicating that these NPs with various functionalities could be effective at discussing the relationship between surface functionality and prediction accuracy in quantification of concentration. The NPs could also be useful to discover an optimized combination that results in high prediction accuracy. Effects of Surface Functionalities on Prediction Accuracy: Use of a Specific Functional Group and Its Combination. To begin with a simple discussion, we first investigated the effect of the presence of a specific functional group such as hydroxyl, phenyl, octadecyl, and aminopropyl groups on prediction accuracy (Figure 4). In the machine learning, four parameters were extracted from the second to fourth injection cycles in each signal as features, and each concentration is used as a label, meaning that three data points are generated from each sample. It should be noted that the sensing signals for the first injection cycle were not used for the analysis because the signals suffer from the memory effect; the sensing signals are affected by the previous measurement.26 A training data set consists of the samples indexed as 1−18 where the number of training data points is 54, while the samples indexed as 19, 20, and 21 were used as a test data set. For each component, the machine learning model based on Gaussian process regression17 was developed by using the COMBO,18 and prediction accuracy was calculated using 54 training data points. To evaluate a prediction accuracy, we adopted a cross validation error (CVE). In the cases of ethanol and methanol, the NPs functionalized with phenyl groups, especially Ph−STNPs, resulted in the lowest CVE. It was reported that Ph−STNPs were also useful for predicting alcohol content,16 i.e., ethanol concentration, supporting the validity of the present result. On the other hand, for the prediction of water concentration, octadecyl functionalized NPs such as C18− STNPs and C18(1)−NH2(1)−STNPs provided low CVEs. This trend was the same when the combination of two NPs was used for the prediction (Figure 4). The combination of C18−STNPs and C18(1)−NH2(1)−STNPs led to the lowest CVE as expected. This result suggests that the hydrophobic property coexisting with slight hydrophilicity is favored to derive useful information for predicting water concentration. For the prediction of ethanol and methanol concentrations, on the other hand, the combination of phenyl modified NPs (Ph−STNPs and Ph(1)−OH(3)−SNPs) showed the second lowest CVEs, while the combination of octadecyl modified NPs (C18−STNPs and C18(1)−NH2(1)−STNPs) provided the highest CVEs, probably due to the higher CVEs of octadecyl modified NPs itself compared to phenyl modified NPs itself. It should be emphasized that the combination of C18(1)−NH2(1)−STNPs and NH2−STNPs showed the lowest CVE, although NH2−STNPs and C18(1)−NH2(1)− STNPs resulted in the highest and second highest CVEs, F

DOI: 10.1021/acssensors.8b00450 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors

methanol decreased (Figure 7). The parity plots of real and predicted concentrations of the training/test data are shown in Figure 7, and the predicted concentration for the test data set is also summarized in Table 2. Although the variation between Table 2. Summary of the Real and Predicted Concentrations of Water, Ethanol, and Methanol for Test Dataa water 1 2 3

Figure 5. SEM images and FT-IR spectra of C18(1)−NH2(4)− STNPs and C18(4)−NH2(1)−STNPs.

ethanol

methanol

real

predicted

real

predicted

real

predicted

60 40 0

65.2 ± 5.1 36.6 ± 3.0 2.3 ± 1.3

0 60 40

0.8 ± 5.1 55.7 ± 3.5 38.0 ± 1.3

40 0 60

34.0 ± 4.2 7.7 ± 3.9 59.7 ± 1.4

a

The predicted values were obtained by the machine learning model with the data collected from the eight types of NPs. The error bars in the predicted values are evaluated as 95% confidence interval.

IR spectra shown in Figure 5; each spectrum shows characteristic absorption bands that can be assigned to octadecyl and aminopropyl groups as described, and relative absorption intensity varies, meaning that the NPs with different surface compositions were obtained. Since the responses with additional trends were experimentally confirmed by measuring the vapors of the 21 liquid samples (Figures 6, S14−16), all eight types of NPs were used

the cases with six and eight types of NPs seems rather small, the addition of C18(4)−NH2(1)−STNPs and C18(1)− NH2(4)−STNPs was effective, leading to the lowest CVE as expected. The higher CVE for water would imply that unnecessary information deteriorates prediction accuracy. While the CVEs decreased, it is interesting to notice the highly accurate quantitative discrimination of such similar compounds: ethanol and methanol. To identify the effective factors for successful discrimination, we tried to extract additional information from the data set. The effects of the combination of two features on CVE were examined by focusing on a series of C18 and NH2 modified NPs. The CVE values are summarized in Tables S3, S5−S11. It was found that the combination of parameter 2, which relates to interaction with an adsorbate at quasi-equilibrium state, results in lower CVE for methanol than ethanol in most cases. As for the receptor material, C18(1)−NH2(4)−STNPs were useful. These trends indicate the importance of receptor materials with multiple chemical functionalities for the modulation of the adsorption mechanism to bring out small differences between structurally similar molecules. As demonstrated in the present experiments, the knowledge for further improvement in the prediction accuracy of a specific target could be obtained through the chemical interpretation of the information provided by the data-driven approach.

Figure 6. Response intensity of C18(4)−NH2(1)−STNPs (left) and C18(1)−NH2(4)−STNPs (right) to the vapor of 21 samples. The three white bars in the right of each graph are those used as a test data set and the other black bars are those used as a training data set for the present machine learning.

to train another machine learning model, and its performance for concentration prediction was examined. As a result, CVE for water increased, while those values for ethanol and

Figure 7. Parity plots of real concentration versus predicted concentration of water, ethanol, and methanol. Thirty-two features obtained from the data taken with the combination of eight types of NPs are used. The blue points represent the training data set that was used to train the machine learning model. The red points represent the test data set. The error bars are evaluated as 95% confidence interval. Insets show a plot of CVE in relation to the presence of the six and eight types of NPs. G

DOI: 10.1021/acssensors.8b00450 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors Through this study, it was demonstrated that properly designed materials in terms of surface properties and their combination enabled one to quantify the concentration of each species in a multicomponent mixture even if a target is coexisting with a contamination gas. It is also possible to apply the present approach to complicated odors composed of a large number of constituents with low concentration after improving both receptor materials and data analysis; sensitivity can be improved by modifying the properties of NPs21 while prediction accuracy can be enhanced by selecting more effective features than those used in the present study.29 Further work is now in progress and will be reported elsewhere. The number of papers that discuss and interpret chemical aspects of a specific phenomenon with the help of data scientific approaches is still small;30,31 however, as demonstrated in this study, new insight that is never reached by experiments only would be discovered through data-driven interdisciplinary research.

■

AUTHOR INFORMATION

Corresponding Authors

*E-mail: [email protected]. *E-mail: [email protected]. ORCID

Kota Shiba: 0000-0001-7775-0318 Kosuke Minami: 0000-0003-4145-1118 Koji Tsuda: 0000-0002-4288-1606 Author Contributions

K.S. and R.T. designed the research; K.S. and T.S. synthesized all the nanoparticles; Y.K., K.K., and E.S. conducted the coating and sensing data collection; K.M., H.T.N., G.I., and G.Y. constructed the sensor measurement setup; R.T. and K.T. carried out the data analysis. The manuscript was written through contributions of all authors. All authors have given approval to the final version of the manuscript.

■

CONCLUSION Various nanoparticles (NPs) functionalized with hydroxyl, aminopropyl, phenyl, and/or octadecyl groups were synthesized through hydrolysis and co-condensation reaction of a silane coupling reagent and titanium tetraisopropoxide. A microfluidic reactor was introduced into the reaction to separate nucleation and growth processes, leading to the formation of six different types of well-defined functional NPs. The NPs were coated on a multichannel nanomechanical membrane-type surface stress sensor (MSS), and the NPscoated MSS was used to measure vapor samples that were mixtures of water, ethanol, and methanol with diverse concentrations. After collecting plenty of sensing data sets, four features were extracted from each sensing signal, making a machine learning model based on Gaussian process regression for predicting concentration of each component in the mixture. As a general trend, prediction accuracy, which was evaluated by the cross validation error, was improved as the number of features used to train the machine learning model increased. The best prediction accuracy for water was achieved by using 24 features from all six types of NPs. The examination of the surface functionality effects of each NP on the prediction accuracy revealed that the octadecyl-modified NPs enhanced prediction accuracy for water, while the use of both octadecyl and aminopropyl groups led to a better prediction accuracy for ethanol and methanol. By introducing two additional NPs modified with different ratios of octadecyl and aminopropyl groups, i.e., 32 features from eight types of NPs, we experimentally confirmed that the best prediction accuracy for ethanol and methanol was obtained. The present ternary mixture is one of the model systems that simulate a practical sensing situationa target is coexisting with structurally similar species under humidified conditions. Through the successful quantification demonstrated here, we believe that this approach can be adopted for quantitative odor analysis in a practical situation.

■

Additional results for mean absolute value, TG-DTA curves, optical microscope images, responses to the 21 samples, parity plots, tables summarizing CVEs (PDF)

Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS This work was supported by CREST (JPMJCR1665 and JPMJCR17J2) and the Materials research by Information Integration Initiative, JST, Japan; a Grant-in-Aid for Research Activity Start-up, 17H07350, MEXT, Japan. The authors thank all the support from the MSS alliance and the World Premier International Research Center Initiative (WPI) on Materials Nanoarchitectonics (MANA).

■

REFERENCES

(1) Konvalina, G.; Haick, H. Sensors for Breath Testing: From Nanomaterials to Comprehensive Disease Detection. Acc. Chem. Res. 2014, 47, 66−76. (2) Gutiérrez, J.; Horrillo, M. C. Advances in artificial olfaction: Sensors and applications. Talanta 2014, 124, 95−105. (3) Potyrailo, R. A. Multivariable Sensors for Ubiquitous Monitoring of Gases in the Era of Internet of Things and Industrial Internet. Chem. Rev. 2016, 116, 11877−11923. (4) Yoon, J.-W.; Lee, J.-H. Toward breath analysis on a chip for disease diagnosis using semiconductor-based chemiresistors: recent progress and future perspectives. Lab Chip 2017, 17, 3537−3557. (5) Hsieh, M.-D.; Zellers, E. T. Adaptation and Evaluation of a Personal Electronic Nose for Selective Multivapor Analysis. J. Occup. Environ. Hyg. 2004, 1, 149−160. (6) Then, D.; Vidic, A.; Ziegler, C. A highly sensitive self-oscillating cantilever array for the quantitative and qualitative analysis of organic vapor mixtures. Sens. Actuators, B 2006, 117, 1−9. (7) Pinnaduwage, L. A.; Zhao, W.; Gehl, A. C.; Allman, S. L.; Shepp, A.; Mahmud, K. K.; Leis, J. W. Quantitative analysis of ternary vapor mixtures using a microcantilever-based electronic nose. Appl. Phys. Lett. 2007, 91, 044105. (8) Zhao, W. C.; Pinnaduwage, L. A.; Leis, J. W.; Gehl, A. C.; Allman, S. L.; Shepp, A.; Mahmud, K. K. Identification and quantification of components in ternary vapor mixtures using a microelectromechanical-system-based electronic nose. J. Appl. Phys. 2008, 103, 104902. (9) Sunny; Mishra, S. V. N.; Dwivedi, R.; Das, R. R. Quantification of Individual Gases/Odors Using Dynamic Responses of Gas Sensor Array With ASM Feature Technique. IEEE Sens. J. 2014, 14, 1006− 1011.

ASSOCIATED CONTENT

S Supporting Information *

The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acssensors.8b00450. H

DOI: 10.1021/acssensors.8b00450 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors (10) Sunny; Kumar, V.; Mishra, V. N.; Dwivedi, R.; Das, R. R. Classification and Quantification of Binary Mixtures of Gases/Odors Using Thick-Film Gas Sensor Array Responses. IEEE Sens. J. 2015, 15, 1252−1260. (11) Adak, M. F.; Akpinar, M.; Yumusak, N. Determination of the Gas Density in Binary Gas Mixtures Using Multivariate Data Analysis. IEEE Sens. J. 2017, 17, 3288−3297. (12) Szulczyński, B.; Namieśnik, J.; Gębicki, J. Determination of Odour Interactions of Three-Component Gas Mixtures Using an Electronic Nose. Sensors 2017, 17, 2380. (13) Yoshikawa, G.; Akiyama, T.; Gautsch, S.; Vettiger, P.; Rohrer, H. Nanomechanical Membrane-type Surface Stress Sensor. Nano Lett. 2011, 11, 1044−1048. (14) Yoshikawa, G.; Akiyama, T.; Loizeau, F.; Shiba, K.; Gautsch, S.; Nakayama, T.; Vettiger, P.; de Rooij, N. F.; Aono, M. Two Dimensional Array of Piezoresistive Nanomechanical MembraneType Surface Stress Sensor(MSS) with Improved Sensitivity. Sensors 2012, 12, 15873−15887. (15) Shiba, K.; Sugiyama, T.; Takei, T.; Yoshikawa, G. Controlled growth of silica-titania hybrid functional nanoparticles through a multistep microfluidic approach. Chem. Commun. 2015, 51, 15854− 15857. (16) Shiba, K.; Tamura, R.; Imamura, G.; Yoshikawa, G. Data-driven nanomechanical sensing: specific information extraction from a complex system. Sci. Rep. 2017, 7, 3661. (17) Bishop, C. M. Pattern Recognition and Machine Learning (Information Science and Statistics), Jordan, M.; Kleinberg, J.; Scholkopf, B., Eds.; Springer: 2006. (18) Ueno, T.; Rhone, T. D.; Hou, Z.; Mizoguchi, T.; Tsuda, K. COMBO: An efficient Bayesian optimization library for materials science. Mater. Discovery 2016, 4, 18−21. (19) Rahimi, A.; Recht, B. Random features for large-scale kernel machines. In Proceedings of the 20th International Conference on Neural Information Processing Systems; Curran Associates Inc.: Vancouver, British Columbia, Canada, 2007; pp 1177−1184. (20) Rasmussen, C. E.; Williams, C. K. I. Gaussian Processes for Machine Learning; The MIT Press: 2006. (21) Yoshikawa, G. Mechanical analysis and optimization of a microcantilever sensor coated with a solid receptor film. Appl. Phys. Lett. 2011, 98, 173502. (22) Sasidharan, M.; Zenibana, H.; Nandi, M.; Bhaumik, A.; Nakashima, K. Synthesis of mesoporous hollow silica nanospheres using polymeric micelles as template and their application as a drugdelivery carrier. Dalton Trans. 2013, 42, 13381−13389. (23) Ou, D. L.; Seddon, A. B. Near- and mid-infrared spectroscopy of sol−gel derived ormosils: vinyl and phenyl silicates. J. Non-Cryst. Solids 1997, 210, 187−203. (24) Hashimoto, R.; Ogawa, M. Facile syntheses of nanoporous organosilica spherical particles. J. Porous Mater. 2018, 25, 425−431. (25) Loizeau, F.; Akiyama, T.; Gautsch, S.; Vettiger, P.; Yoshikawa, G.; de Rooij, N. F. Comparing membrane- and cantilever-based surface stress sensors for reproducibility. Sens. Actuators, A 2015, 228, 9−15. (26) Hierlemann, A.; Gutierrez-Osuna, R. Higher-Order Chemical Sensing. Chem. Rev. 2008, 108, 563−613. (27) Ide, Y.; Ogawa, M. Interlayer modification of a layered titanate with two kinds of organic functional units for molecule-specific adsorption. Angew. Chem., Int. Ed. 2007, 46, 8449−8451. (28) Ide, Y.; Iwasaki, S.; Ogawa, M. Molecular Recognition of 4Nonylphenol on a Layered Silicate Modified with Organic Functionalities. Langmuir 2011, 27, 2522−2527. (29) Imamura, G.; Yoshikawa, G.; Shiba, K.; Washio, T. Signal analysis and sample gas discrimination method; Japanese Patent Application 2016-040625, March 3, 2016. (30) Wexler, R. B.; Martirez, J. M. P.; Rappe, A. M. Chemical Pressure-Driven Enhancement of the Hydrogen Evolving Activity of Ni2P from Nonmetal Surface Doping Interpreted via Machine Learning. J. Am. Chem. Soc. 2018, 140, 4678−4683.

(31) Nielsen, M. K.; Ahneman, D. T.; Riera, O.; Doyle, A. G. Deoxyfluorination with Sulfonyl Fluorides: Navigating Reaction Space with Machine Learning. J. Am. Chem. Soc. 2018, 140, 5004−5008.

I

DOI: 10.1021/acssensors.8b00450 ACS Sens. XXXX, XXX, XXX−XXX