Multiwalled Carbon Nanotube-N-Doped Graphene/Poly(3,4-ethylenedioxythiophene):Poly(styrenesulfonate) Nanohybrid for Electrochemical Application in Intelligent Sensors and Supercapacitors.

In this study, we reported the preparation of a conducting polymeric/inorganic nanohybrid consisting of multiwalled carbon nanotubes (MWCNT), N-doped graphene (NGr), and poly(3,4-ethylenedioxythiophene):poly(styrenesulfonate) (PEDOT:PSS), and its electrochemical application in intelligent sensors and supercapacitors. The multilayer thin film of the PEDOT:PSS-supported MWCNT-NGr nanohybrid was prepared by a facile layer-by-layer assembly strategy. The obtained conducting polymeric/inorganic nanohybrid modified electrode displayed superior electron transfer ability and a high specific surface area, which was used for electrochemical applications in intelligent sensors and supercapacitors. Remarkably, the fabricated amaranth sensor exhibited a broad linear range of 0.05-10 μM with a limit of detection of 0.015 μM under the optimal conditions. With the help of the response surface methodology, multivariate optimization was used as a substitute for the traditional single variable optimization to reflect the complete real effects of multivariate optimization in a sensing platform. Machine learning implemented by hybrid genetic algorithm-artificial neural network was used as an intelligent analysis model to replace the traditional regression analysis model for realizing intelligent analysis and output of sensing system. The MWCNT-NGr/PEDOT:PSS modified electrode exhibited a considerable specific capacitance of 6.5 mF cm-2 at a current density of 2.0 mA cm-2. The proposed results provided a new thought for a nanosensing platform equipped with a supercapacitor as a self-powered electrochemical energy storage system and machine learning as an intelligent analysis and output system in the near future.

Introduction

With the diversified development of science and technology and rapid integration of multiple disciplines, especially the booming development of the internet of things (IoTs), electrochemical sensors have also seen new developments. Lab-based research to field applications and the miniaturization of large equipment to portable devices are now made possible in the process of technological transformation.[1,2] The data recording and processing instruments have developed from traditional desktop computers/laptops to tablets/smartphones, and data transmission has also evolved from wired to wireless.[3] Therefore, traditional data processing also needs to be changed. Traditional electrochemical sensors usually achieve data acquisition via electroanalytical methods; data processing was commonly performed through linear models to establish regression equations, which cannot meet the demand of the booming growth of electroanalytical mini devices in the IoT era, as it lacks the capacity to achieve intelligent analysis and smart transmission.

Therefore, it is necessary to develop novel and effective data processing methods, in this case, machine learning (ML) was used as a key artifice and was employed to realize intelligent analysis and processing of data, and further applied in materials science and chemistry.[4] As an effective and forceful machine learning model, artificial neural networks (ANNs) have drawn attention to materials in the field of science and analytical chemistry due to their attractive merits including rapid learning, fast computation, and ease of implementation.[5] The ANN can learn various kinds of continuous functions and fit in both linear and nonlinear systems, we previously proposed novel strategies by integrating electrochemical sensing to ML to realize intelligent sensing of maleic hydrazide or carbendazim.[6,7] At the same time, the power supply patterns have evolved from common external power supply to self-powered mini device, which prompted the rapid development of nanogenerators or self-powered/energy devices like supercapacitors and diverse batteries.[8,9]

Supercapacitors, also called electrochemical capacitors, have been widely used in energy storage systems, self-energy sensing systems, intelligent electronic microdevices, and other fields.[10] Due to their outstanding properties including fast charging ability, high power density, and stable cycling life span, supercapacitors can fill the gap between conventional electrolytic capacitors and batteries.[11] As the demand for advanced electronic devices increases, it is necessary to integrate greater functions into energy storage devices and develop novel electrode materials that are suitable for electrochemical applications.[12,13] During the long development process, the working electrodes have evolved from traditional electrodes such as mercury electrodes, glassy carbon electrodes, gold electrodes, platinum disk electrodes, graphite electrodes, and carbon paste electrodes to new electrodes including screen printing electrodes,[14] inkjet printing electrodes,[15] 3D printing electrodes,[16] and laser direct writing electrodes,[17] which promoted the rapid and large-scale development of chemically modified electrodes to special and industrial electrodes.

PEDOT:PSS is the most successful commercial water-processable conducting polymer, which has been widely used as the core material in polymeric/inorganic electronics and is expected to play a major role in flexible electrodes in the future.[18,19] We have previously enhanced the performance of PEDOT:PSS by incorporating ionic liquids,[20] Nafion,[21,22] poly(vinyl alcohol),[23,24] carboxymethyl cellulose,[25−27] and poly(vinylpyrrolidone),[28] which have proven benefits in electrochemical sensors. Carbon nanomaterials are the most studied and promising novel inorganic nonmetallic nanomaterials used as electrode materials; the structures and properties of carbonaceous materials can be effectively tailored by heteroatoms doping, the doping of nitrogen atoms into graphene resulted in superior nitrogen-doped graphene (NGr),[29,30] which has been widely employed in electrochemical applications such as in bio/chem sensors and supercapacitors due to its favorable detectability, superior electrocatalytic ability, high adsorption capacity, and large specific surface area.[31−34]

Herein, we developed a conducting polymeric/inorganic nanohybrid based on PEDOT:PSS-supported MWCNT-NGr nanohybrid. We then investigated its electrochemical properties, supercapacitive performance, and sensing performance (Scheme ). Finally, the response surface methodology (RSM) was selected for multivariate optimization of AM sensing parameters by one variable at a time methodology, which was selected to reflect the complete effects of two or more variables in a measured response and time-consuming practical processing, ML based on a hybrid genetic algorithm-artificial neural network (GA-ANN) model was used for DPV intelligent analysis and output of the AM sensor in comparison with the univariate regression model. which could meet current needs of the burgeoning development of electronic products/devices where traditional data analysis cannot realize intelligent processing and output of data in digital display instruments/apparatus.[35,36]

Scheme 1

Schematic Illustration of the Fabrication of Multilayer Nanohybrid and its Potential Applications in an Intelligent Sensor and a Supercapacitor

Results and Discussion

Surface Morphology and Elemental Analysis

Figure A–D presents the SEM images of PEDOT:PSS, MWCNT, NGr, and MWCNT-NGr/PEDOT:PSS, respectively. The PEDOT:PSS film has a uniform, dense, and smooth surface (Figure A) due to its excellent film-forming capacity, which is in accordance with previous reports.[28,37,38] MWCNT showed cylindrical shapes tightly entangled with each other (Figure B) owing to the electrostatic interactions, while NGr displayed thin wrinkled-paper-like shapes (Figure C) and a wavy nanosheet structure. In the case of MWCNT-NGr/PEDOT:PSS, the typical nanohybrid contains wrinkled NGr sheets and randomly oriented MWCNT that covered the surface of the PEDOT:PSS film, some micropores formed by NGr and MWCNT could also be observed (Figure D), indicating the formation of a 3D MWCNT-NGr nanohybrid that is supported on the PEDOT:PSS film. This hierarchically porous structure may provide a large active surface area as an effective electrocatalyst in practical electrochemical processing.[39,40] EDS (Figure E–H) analysis showed the components in the MWCNT-NGr/PEDOT:PSS nanohybrid, the presence of the peaks corresponding to the elements C, N, O, and S confirms the presence of the PEDOT:PSS substrate, while N came from NGr (Figure G) and S originated from PEDOT:PSS (Figure E).

Figure 1

SEM images of (A) PEDOT:PSS, (B) MWCNT, (C) NGr, and (D) MWCNT-NGr/PEDOT:PSS (the white circles indicate micropores formed by the nanohybrid) and the corresponding EDS elemental analysis spectra (E–H).

The atomic force microscopy (AFM) images of PEDOT:PSS and MWCNT-NGr/PEDOT:PSS are displayed in Figure . The PEDOT:PSS film exhibited smooth and uniform morphology with a surface roughness value of 2.38 nm, while MWCNT-NGr was coated onto the PEDOT:PSS, the surface morphology changed as well, and rough tubular structure belonging to MWCNT were seen in Figure B.

Figure 2

Atomic force microscopy (AFM) images of the samples: (a) PEDOT:PSS and (b) MWCNT-NGr/PEDOT:PSS.

Here, Raman spectrum was used to investigate the vibrational modes, crystalline quality, and generated defects of MWCNT, NGr, and MWCNT-NGr. The Raman spectrum of the MWCNT (Figure A) shows two typical peaks at 1350 and 1591 cm–1 corresponding to D and G bands, whereas the Raman spectrum of NGr exhibits the relative peaks at 1359 and 1595 cm–1. Besides, 2D bands at 2701 cm–1 (MWCNT) and 2621 cm–1 (MWCNT-NGr) can be seen, which originated from the E2g mode and from second-order backscattering. It’s clearly seen that the Raman spectrum of MWCNT-NGr exhibits sharp and intense peaks at 2945 cm–1 corresponding to D′ + G′ band, while the D′ + G′ band of the pristine MWCNT are located at 2938 cm–1, which could be attributed to the asymmetric C–H stretching of surface CH3. Besides, FT-IR spectroscopy was used to characterize the MWCNT, NGr, and MWCNT-NGr. As shown in Figure B, the curve of MWCNT shows bands at 2917 cm–1 (O–H stretch), 1582 cm–1 (C=C stretch), and 1176 cm–1 (C=O stretch), the peaks at 1110 and 1570 cm–1 in NGr corresponds to the C–N and C–C stretch. Meanwhile, the bands at 1654 cm–1 (C–C stretch) and 1088 cm–1 (C–N stretch) can be seen in the curve of MWCNT-NGr.

Figure 3

(A) Raman spectrum for MWCNT, NGr, and the nanocomposites MWCNT-NGr. (B) FT-IR spectrum of MWCNT, NGr, and MWCNT-NGr.

Electrochemical Properties

Electrochemical impedance spectroscopy (EIS) of bare GCE (a), and PEDOT:PSS (b), MWCNT (c), NGr (d), MWCNT-NGr (e), and MWCNT-NGr/PEDOT:PSS (f) modified GCE were carried out to analyze the electron transfer kinetics in 10 mM [Fe(CN)6]3–/4– solution (ratio 1:1) containing 0.1 M KCl (Figure A). The value of the electron transfer resistance (Rct) could be estimated from the diameter of the semicircle in Nyquist plots of EIS and ordered as follows: bare GCE (180 Ω) > MWCNT (22.06 Ω) > NGr ≥ PEDOT:PSS ≈ MWCNT-NGr ≈ MWCNT-NGr/PEDOT:PSS, and the four values belonging PEDOT:PSS, NGr, MWCNT-NGr, and MWCNT-NGr/PEDOT:PSS were approximately 0 and the relative curves appeared as straight lines in both high and low frequencies due to favorable conductivity of PEDOT:PSS or NGr.

Figure 4

Nyquist plots (A) of different electrodes in 10 mM [Fe(CN6)]3–/4– containing 0.1 M KCl. (B) Anodic peak currents of [Fe(CN)6]3– (1 mM) vs v1/2 at different electrodes: (a) bare GCE, (b) PEDOT:PSS/GCE, (c) MWCNT/GCE, (d) NGr/GCE, (e) MWCNT-NGr/GCE, and (f) MWCNT-NGr/PEDOT:PSS/GCE.

To estimate the effective surface areas of different modified electrodes, CVs at different scan rates were obtained in 1 mM K3[Fe(CN)6] containing 0.1 M KCl. It is possible to observe a linear increase of anodic peak currents (Ipa) with the square root of scan rate (v1/2) for different electrodes (Figure B). The effective surface area can be calculated according to the Randles–Sevcik equation as followswhere Ipa is the anodic peak current (μA), n is the electron transfer number, A is the effective surface area (cm2), C0 is the concentration (mM) of K3[Fe(CN)6], DR is the diffusion coefficient (cm2 s–1), and v is the scan rate (V s–1). For 1 mM K3[Fe(CN)6] with the 0.1 M KCl electrolyte: n is 1 and DR is 7.6 × 10–6 cm2 s–1. From the slope of Ipa vs v1/2, effective surface areas of GCE, PEDOT:PSS, MWCNT, NGr, MWCNT-NGr, and MWCNT-NGr/PEDOT:PSS were calculated to be 0.0387, 0.0553, 0.0659, 0.0667, 0.0692, and 0.1394 cm2, respectively. Obviously, the introduction of the PEDOT:PSS film as the substrate of MWCNT-NGr hybrid will provide a higher electroactive surface area, which is 3.6-fold higher than bare GCE. In summary, the EIS and surface area results indicated that the MWCNT-NGr/PEDOT:PSS hybrid with efficient charge transfer and high electroactive surface area is well-suited for electrochemical applications.

Sensing Performance of MWCNT-NGr/PEDOT:PSS

Electrochemical Behaviors of AM

The electrochemical behavior of 98 μM AM at different modified electrodes was investigated by CV (Figure ) in 0.1 M PBS (pH = 6.0) with scan rate at 50 mV s–1. The result indicated that one oxidation peak was observed on the surface of all modified electrodes and no reduction peak was found, which showed that the electrochemical reaction of AM on the surface of all electrodes was an irreversible process. A slightly undetectable oxidation peak at 0.84 V was observed for bare GCE (a) but obvious improvements of both peak current and the shape of the whole oxidation peak could be seen for MWCNT/GCE (c), NGr/GCE (d), MWCNT-NGr/GCE (e), and MWCNT-NGr/PEDOT:PSS/GCE (f), especially the oxidation peak current reached a maximum for the last one. The most sensitive voltammetric response of AM at MWCNT-NGr/PEDOT:PSS/GCE might be due to the excellent intrinsic electrocatalytic ability and favorable conductivity of MWCNT and NGr, and a synergistic effect of MWCNT-NGr nanohybrid supported on the PEDOT:PSS substrate.[41]

Figure 5

CVs of 98 μM AM in 0.1 M PBS (pH = 6.0) at different modified electrodes: (a) bare GCE, (b) PEDOT:PSS/GCE, (c) MWCNT/GCE, (d) NGr/GCE, (e) MWCNT-NGr/GCE, and (f) MWCNT-NGr/PEDOT:PSS/GCE. The inset shows the CVs of (a) bare GCE and (b) PEDOT:PSS/GCE.

In addition, the oxidation peak potentials shifted with increasing scan rates toward a more positive potential (Figure S1A), curves of oxidation peak currents increased linearly with the square root of scan rates (v1/2) (Figure S1B), while the curves of oxidation peak currents versus scan rates (v) showed a downward bending characteristic (Figure S1B inset), indicating that the oxidation reaction of AM on MWCNT-NGr/PEDOT:PSS/GCE was a typical diffusion-control process. Figure S1C depicts the linear relationship between the anodic peak potential and the Napierian logarithm of v for AM, the value of αn (α is the transfer coefficient and n is the electron transfer number) could be calculated by the obtained formula (slope = RT/αnF) according to the Laviron equation, which could be determined from the slope of oxidation peak potentials versus the natural logarithm of v. The values of αn were 1.31 for CV responses of AM. Generally, the value of α is between 0.3 and 0.7, so, from this we obtained the value of α as 0.5. Furthermore, the number of n in the electroreduction of AM was calculated to be 2. Figure S1D showed DPV curves of 98 μM AM at MWCNT-NGr/PEDOT:PSS/GCE in 0.1 M PBS with different pH values. Oxidation peak potentials of AM shifted negatively with the increase of pH, indicating that protons participated in the electrochemical process of AM on the surface of the MWCNT-NGr/PEDOT:PSS/GCE electrode. Figure S1E presents the linear relationship between the oxidation peak potential and pH for DPV responses of AM, the number of protons (m) was calculated by the obtained formula (dEpa/dpH = −2.303 mRT/nF) according to Nernst equation, which could be determined from the slopes of Epa vs pH. the slope of dEpa/dpH plots was 37.26, the m/n ratio of AM was about 1:1, suggesting that the number of electrons and protons was equal, so the value of m was calculated to be 2.

Multivariate Optimization of the Experimental Conditions

The optimization of multiple factors (n < 3) for the detection of AM was carried out by RSM under the same experimental conditions as in Figure A,B. Five independent factors including pH (5.0–7.5), preconcentration time (50–500 s, with a gradient of 50 s), MWCNT amount (0.3–1.8 mg mL–1), and NGr amount (0.3–1.8 mg mL–1) were investigated at five levels; 3D surface plots for DPV responses of AM in (Figure A,B) indicated that the optimized values of pH, preconcentration time, MWCNT amount, and NGr amount were 6.48, 275 s, 1.16, and 0.76 mg mL–1, respectively.

Figure 6

3D surface plots for DPV responses of AM using RSM with (A) both pH and preconcentration time; (B) both MWCNT amount and NGr amount at MWCNT-NGr/PEDOT:PSS/GCE in 0.1 M PBS containing 98 μM AM.

To confirm the optimization of multiple factors for the detection of AM, the optimization of one factor at a time was carried out. The oxidation peak currents increased linearly and reached a maximum value at pH 6.0; however, the oxidation peak current decreased beyond pH 6.0 (Figure S2A). The relationship between oxidation peak currents and preconcentration time indicated that oxidation peak currents increased from 50 to 250 s, while the oxidation current decrease after 250 s up to 500 s (Figure S2B). The concentration of MWCNT showed the same tendency. Thus, the peak current was the maximum when the concentration of MWCNT increased to 1.1 mg mL–1 (Figure S2C). Similarly, the oxidation peak currents first increased and then decreased when the concentration of NGr increased (Figure S2D), reaching a maximum value at 0.75 mg mL–1.

The results for the optimization of one factor at a time were almost in accordance with those for the optimization of multiple factors, the optimization of multiple factors for detection of AM by RSM was feasible. However, the methods and plots for the optimization of multiple factors (n ≥ 3), especially in the same number of experiments remain unsolved.

Electrochemical Detection of AM

Figure shows the electrochemical detection of AM at various concentrations using MWCNT-NGr/PEDOT:PSS/GCE. The oxidation peak currents of AM gradually increased with increasing AM concentrations [AM], indicating that the as-prepared sensor exhibited an advanced performance for AM electrocatalytic oxidation (Figure A). Oxidation peak currents of AM versus [AM] (Figure B) showed a specific functional relationship utilizing common linear regression methods, which could be used to determine the unknown content of AM using the fabricated electrochemical sensor in combination with traditional linear regression methods. Moreover, the sensor also showed a wide linear range from 0.05 to 10 μM, the equation is Ipa = 1.2330C + 0.2894, the limit of detection (LOD, defined as 3σ/S) and limit of quantification (LOQ, defined as 10 σ/S) were 0.015 and 0.05 μM, respectively (where σ is the standard deviation of the blank response and S is the slope of the calibration curve with linear range of lower concentrations of the analyte). Additionally, the performance of the AM sensor electrode modified with different materials in previous research is listed in Table S1. The fabricated MWCNT-NGr/PEDOT:PSS/GCE displayed a wider detection range and a lower detection limit, indicating that the prepared sensor based on MWCNT-NGr/PEDOT:PSS/GCE had better sensing performance.

Figure 7

(A) DPV responses of AM on MWCNT-NG/PEDOT:PSS/GCE at different AM concentrations (0.05–10 μM) in 0.1 M PBS (pH = 6.0). (B) The relationship between the oxidation peak currents and [AM] using the univariate linear regression method. Optimization history by the (C) GA and (D) GA-ANN model with optimal parameters. (E) Training and testing of the GA-ANN for AM concentrations estimation.

Developed GA-ANN Model for Data Intelligent Analysis

To obtain the optimum performance of predicted modeling, the hybrid GA-ANN method was developed to realize intelligent sensing for AM. AM concentrations as output values were predicted by the GA-ANN and input data of peak current values, which were obtained from the experimental results of the fabricated electrochemical sensor.

During modeling, different parameters such as learning rate, activation function, and maximum epoch number were tested to achieve the best network by various neuron arrays. The achieved results of networks were surveyed and the ANN modeling with the 1 × 6 × 1 architecture, epoch number of 500, and a training rate of 0.1 were selected. Moreover, the genetic breeding pool size is set as 50, maximal genetic algebra is 100, and crossover and mutation probability is 0.1 and 0.3, respectively, in the optimum process of GA. Optimization history up to 100 generations is illustrated in Figure C. In doing this, a powerful ANN model was built using optimum weights and bias values selected by GA. Figure D shows the developed GA-ANN model of currents and different concentrations of AM, which was built using the optimum parameters. Interestingly, the proposed GA-ANN model could predict values of AM concentrations quite well, and even where the slope of the diagram changes. Hence the developed ANN complemented the fabricated electrochemical sensor based on MWCNT-NGr/PEDOT:PSS/GCE for data intelligent analysis of AM measurement infield. Besides, Figure E demonstrates the comparative diagram of the experimental and predicted values using GA-ANN where the satisfactory generalization ability of the model is shown.

Evaluated and Compared Models

To obtain the ideal linear regression model, researchers have considered to transform the original data (i.e., squaring, taking the exponent and logarithm, etc.). However, this strategy is always a subjective choice, which has great limitations and is difficult to adapt to the actual changing data. Of note, the true characteristics of the original data could not be fully reflected. A nonlinear method of multilayer ANN could overcome these problems for regression analysis. Based on the same training and testing samples, univariate linear regression and ANN models were selected to assess the capability of the developed GA-ANN, data corresponding to the statistical criteria for univariate linear regression, the ANN and GA-ANN models are shown in Table . It could be observed that both ANN methods were feasible and better than univariate linear regression in modeling and generalization capability. In addition, the performance of the ANN model optimized by GA is obviously improved. R2 values of GA-ANN for both training and testing were 0.9999, reflecting how well the performance of the optimized ANN (GA-ANN) model was. In addition, it was observed that the values of MRE, MAE, and RMSE for training and testing were smaller, it also illustrated that the GA-ANN is built in an effective way. MRE, MAE, and RMSE values in the testing set displayed a negligible increase (a very small range) in comparison with the training network, which was normal. Compared to the ANN, it shows better robustness that R2 values of GA-ANN is closer in the training and testing set.

Table 1

Comparison and Analysis of Univariate Linear Regression, ANN, and GA-ANN Models

	training set				testing set
model	R2	MRE	RMSE	MAE	R2	MRE	RMSE	MAE
univariate linear regression	0.9922	0.4250	0.2710	0.1852	0.9944	0.2984	0.2743	0.2470
ANN	0.9987	0.2589	0.1265	0.0613	0.9966	0.1483	0.2219	0.1594
GA-ANN	0.9999	0.1435	0.0385	0.0337	0.9999	0.0773	0.1034	0.0817

Repeatability, Reproducibility, and Interference

The repeatability of as-prepared MWCNT-NGr/PEDOT:PSS/GCE in terms of repetitive use for the voltammetric detection of 5 μM AM was estimated by 20 successive assays (Figure S3A), and a relative standard deviation (RSD) value of 2.48% was obtained, implying good repeatability of MWCNT-NGr/PEDOT:PSS/GCE for sensing AM. Under the same conditions, five MWCNT-NGr/PEDOT:PSS modified electrodes were prepared to assess the reproducibility of the fabricated sensing electrode (Figure S3B), and the RSD value is 3.32%, which illustrated good reproducibility of the fabricated sensing electrode. In addition, as is discussed in detail on the interference in the ESI, the changes in the oxidation peak currents in the presence of these substances were less than 5% in Figure S4, indicating that the modified electrode has good anti-interference.

Supercapacitive Performance of MWCNT-NGr/PEDOT:PSS

Electrochemical sensors combined with self-powered electrochemical energy storage systems play an important role in the development of future electronics, here capacitive properties of MWCNT-NGr/PEDOT:PSS were investigated by CV in different electrolytes (aqueous and organic), as shown in Figure A,B, in 0.1 M H2SO4 and 0.1 M ACN-LiClO4, the CV curves exhibited a nearly rectangular shape and a larger area than that in other electrolytic media, indicating a pseudocapacitive behavior of MWCNT-NGr/PEDOT:PSS. Moreover, CV curves of the nanohybrid electrode can be observed at various scan rates such that the nanohybrid maintains current increases without obvious distortion in the shape of curves even at a scan rate of 200 mV s–1 in the two selected electrolyte systems (Figure S5A,B), confirming the good rate capability and efficient ionic transport within electrode materials, which means that MWCNT-NGr/PEDOT:PSS exhibits a good supercapacitive performance and a high rate of charge/discharge ability. A similar phenomenon was reported elsewhere for the graphene-MWCNT nanohybrid, and these carbon nanomaterials contribute to the pseudo-capacitance behavior.[42] This supercapacitor exhibited a skewed triangular shape of the charge–discharge curves of the nanohybrid and the increased discharge time compared with the symmetrical triangular curves with limited discharge time for MWCNT-NGr/PEDOT:PSS. The special capacitances were calculated from GCD curves (Figure C,D), the specific capacitances were calculated to be 8.90, 8.13, 7.81, 7.16, and 6.50 mF cm–2 at current densities of 0.125, 0.25, 0.5, 1.0, and 2.0 mA cm–2, respectively, in H2SO4 and 4.46, 4.30, 3.81, 3.32, and 3.04 mF cm–2 at current densities of 0.125, 0.25, 0.5, 1.0, and 2.0 mA cm–2 in 0.1 M ACN-LiClO4. The specific capacitances decreased with increasing current density (Figure E), indicating a favorable electrochemical performance and excellent rate capability of MWCNT-NGr/PEDOT:PSS.[12,13,43] The specific capacitance decreased with the increase of scanning rates (Figure F), implying a favorable excellent rate capability of MWCNT-NGr/PEDOT:PSS. Figure G shows the Ragone plot of MWCNT-NGr/PEDOT:PSS, the power densities increase from 5.29 × 10–6 to 8.65 × 10–5 W cm–2 when energy densities decrease from 7.91 × 10–4 to 5.78 × 10–4 W h cm–2, respectively, revealing a comparable energy and power density of MWCNT-NGr/PEDOT:PSS. The specific capacitance of the MWCNT-NGr/PEDOT:PSS nanohybrid retained 98.04% in the H2SO4 system and 94.02% in the ACN-LiClO4 system of its initial capacitance after 1000 charge–discharge cycles (Figure H), which was because of the swelling behavior of PEDOT:PSS that resulted in the partial dissolution of PSS in organic solvent or water. After this, the capacitance retention dropped only up 3.93% in the H2SO4 system and 1.17% in the ACN-LiClO4 system from 1000 to 5000 cycles (Figure H), indicating that the nanohybrid displayed good cycle stability.

Figure 8

Cyclic voltammograms of MWCNT-NGr/PEDOT:PSS/GCE in (A) aqueous electrolytes and (B) organic electrolytes. Galvanostatic charge/discharge curves of an asymmetric supercapacitor based on MWCNT-NGr/PEDOT:PSS/GCE in (C) 0.1 M H2SO4 and (D) ACN with 0.1 M LiClO4. (E) Rate capability, (F) specific capacitance at different scan rates, (G) Ragone plots, and (H) cycling stability of an asymmetric supercapacitor based on MWCNT-NGr/PEDOT:PSS/GCE.

Conclusions

A sensing platform based on a multilayer MWCNT-NGr/PEDOT:PSS nanohybrid has been developed, and the intelligent analysis of AM was realized by ML based on the ANN algorithm. The nanohybrid displayed fast electron transfer, high effective surface area, good film electrode stability, excellent electrocatalytic capacity with a relatively low oxidation peak potential, and a wide working range of 0.05–10 μM with a low LOD of 0.015 μM. Both RSM and ANN displayed better feasibility and superiority. Notably, the nanohybrid also offers a remarkable supercapacitive performance with a specific capacitance of 6.5 mF cm–2 at a current density of 2.0 mA cm–2 for future application in self-energy sensors. These results indicate that this work provides a broad scope for self-powered nanosensing platforms equipped with a supercapacitor as the electrochemical energy storage system and machine learning as the intelligent analysis system for multivariate optimization of parameters and intelligent output of sensing information.

Experimental Section

Chemicals

AM (purity 94.7%) was purchased from J&K Scientific Ltd. PEDOT:PSS aqueous dispersion (Baytron P, 1.3 wt %) was obtained from Bayer AG. MWCNT aqueous dispersion (50–100 nm, 4.39 wt %) was used as received from Chengdu Organic Chemicals Co. Ltd. NGr (purity 99.9%) powder was provided by Nanjing XFNANO Materials Tech Co. Ltd. A series of phosphate-buffered solutions (PBS) with different pH values was obtained by mixing 0.1 M Na2HPO4 solution with 0.1 M NaH2PO4 solution. All other reagents were of analytical grade and used as received.

Instruments

Scanning electron microscopy (SEM) images and energy dispersive spectrometer (EDS) data were obtained using an S-4800 scanning electron microscope. All of the electrochemical measurements (CV, EIS, and DPV) were performed using a CHI660D electrochemical workstation (Chenhua Instrument Co. Ltd. Shanghai, China) in a three-electrode system composed of a saturated calomel electrode (SCE) as the reference electrode, a Pt wire as the counter electrode, and unmodified or modified glassy carbon electrode (GCE, d = 3 mm) as the working electrode.

Preparation of the Modified Electrode

NGr powder (10 mg) was ultrasonically dispersed in 10 mL of N,N-dimethylformamide (DMF) for 1 h to produce a homogeneous NGr-DMF (referred as NGr) suspension (1 mg mL–1), and NGr suspensions with different concentrations (0.25, 0.375, 0.5, 0.625, 0.75, 0.875, 1.0, and 1.125 mg mL–1) were simultaneously prepared as well. The mixture of MWCNT aqueous dispersion (0.3, 0.5, 0.7, 0.9, 1.1, 1.3, 1.5, and 1.8 mg mL–1) and NGr dispersion was stirred for 12 h at room temperature. Pristine PEDOT:PSS dispersion was stirred for 5 h to obtain a homogeneous PEDOT:PSS suspension, and different concentrations of PEDOT:PSS suspensions (0.0325, 0.065, 0.13, 0.26, 0.52, and 1.04 wt %) were obtained by diluting PEDOT:PSS (1.3 wt %) with double-distilled water.

The GCE was polished with a chamois cloth with 0.05 μm alumina slurry until a mirror-like surface was obtained and washed with nitric acid (1:1), ethanol, and DI water in turns. After this, 5 μL aliquot of PEDOT:PSS was drop-coated on the surface of the clean GCE, and placed under an infrared lamp until dry, 5 μL of MWCNT-NGr dispersion was dropped on the GCE surface and dried again.

GA-ANN Model

Artificial Neural Network

An ANN refers to mathematical models that were inspired by information processing in the human brain (Scheme A), and aims to establish a complicated and nonlinear relationship between input variables (e.g., anodic peak currents) and network output (e.g., AM concentrations).[44] A representative back-propagation ANN contains nodes structured into an input layer, several hidden layers, and an output layer (Scheme B), each between a pair of nodes in adjacent layers is connected by weight and bias. It is a prerequisite to train the ANN with some optimization algorithm to achieve the best fit to training data.[45] In addition, one of the most difficult tasks in researching ANN is to find a suitable architecture that has the optimal number of hidden layers and nodes. Here, the architecture was selected through the trial and error method.

Scheme 2

(A) The Structure of a Biological Neural Network (left) and the Common ANN Model (right). (B) The Information Transmission Between Synapses (left) and a Typical Back Propagation ANN Model (right)

Genetic Algorithm

Genetic algorithm (GA) is a parallel searching method for solving optimization problems depending on the Darwin evolution theory of “survival of the fittest” and biological species evolution.[46] In general, GA employs a chromosome encoding strategy to randomly build the gene pool of n individuals. The following GA is generated by combining selection, crossover, and mutation until a satisfactory solution for the problem is identified. As genetic algorithm is independent of gradient information in solving the optimal solution, it has strong robustness in finding global optimal solution. With its exceptional performance, the GA is useful in solving the optimization problems in various fields.[47−49]

GA-ANN Combination

An ANN usually uses the gradient descent method to update the weights to obtain the optimal result. However, it is an empirical risk minimization method based on gradient information, which may lead to premature convergence to fall into partial optimization. For enhancing generalization ability and robustness of the ANN, GA can be used to modify the weights and biases of the ANN. In the GA-ANN model, sum squared error is applied as a fitness function,[50] and the global minimum discovered in target space using GA is adopted in the ANN. In the present study, Table S1 shows the optimal results of experimental tests with 20 different concentrations of AM samples versus response peak currents of AM for training and testing. In the ANN and GA-ANN models developed, 15 (75%) samples data were used for training, and 5 samples (25%) were used as data sets for testing, which were not employed during training. In this work, the MATLAB environment was employed for GA optimization using the genetic algorithm toolbox and ANN validation by the artificial neural network toolbox.

Evaluation of GA-ANN Performance

The performance of the GA-ANN model was statistically evaluated using the coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), and mean relative error (MRE). The values of R2 were close to 1, MAE, MRE, and RMSE were smaller, indicating that the model was performed perfectly. Different parameters were computed using equations as followswhere, n is the number of samples for modeling, vEXP is experimental values, and vANN is the predicted values of GA-ANN.