Title Page

Contents

Abstract 18

Chapter 1. INTRODUCTION 20

1.1. Machine learning in electrochemical sensors 20

1.1.1. Concentration prediction with machine learning 21

1.1.2. Chemical identification with machine learning 23

1.1.3. Analysis of electrochemical data using convolutional neural network 23

1.1.4. Generative model for data exploration and optimization 24

1.1.5. Smart point-of-care detection 25

1.1.6. Materials for electrochemical sensing platforms 26

1.2. Research objective and outline 32

Chapter 2. BASIC PRINCIPLES AND TERMINOLOGIES, AND FUNDAMENTAL CALCULATIONS 34

2.1. Electroanalytical methods 34

2.1.1. Cyclic voltammetry (CV) 35

2.1.2. Electrochemical impedance spectroscopy (EIS) 39

2.2. Machine Learning for data analysis 44

2.2.1. Machine learning algorithm 44

2.2.2. Convolutional Neural Network (CNN) 50

2.2.3. Variational autoencoder (VAE) 52

Chapter 3. ELECTROCHEMICALLY EXFOLIATED WS₂ NANOSHEETS FOR THE ELECTROCHEMICAL IMPEDIMETRIC DETECTION OF NADH 55

3.1. Introduction 55

3.2. Experimental Section 59

3.2.1. Reagents 59

3.2.2. Electrochemical exfoliation of bulk WS₂ 59

3.2.3. Bulk WS₂ electrode was preparation 60

3.2.4. Electrochemical measurements 60

3.3. Results and Discussion 61

3.3.1. Morphological and structure analysis 61

3.3.2. Cyclic voltammetry characteristic of the ex-WS₂/SPCE 66

3.3.3. Electrochemical impedimetric detection of NADH on ex-WS₂/SPCE 70

3.4. Conclusion 83

Chapter 4. ANALYSIS OF ELECTROCHEMICAL IMPEDANCE DATA: USE OF MACHINE LEARNING AND DEEP NEURAL NETWORKS 84

4.1. Introduction 84

4.2. Experimental Section 88

4.2.1. EIS data simulation 88

4.2.2. Deep neural network building and training 89

4.3. Results and discussion 90

4.3.1. EIS dataset 90

4.3.2. Deep neural network construction for the EIS model classification 95

4.3.3. Deep neural network construction for the EIS parameters regression 101

4.3.4. Training and evaluation of the EIS model classification 102

4.3.5. Training and evaluation of the EIS parameters regression 115

4.4. Conclusions 121

Chapter 5. DEEP GENERATIVE LEARNING FOR EXPLORATION IN LARGE ELECTROCHEMICAL IMPEDANCE DATASET 122

5.1. Introduction 122

5.2. Experimental Section 129

5.2.1. EIS data acquisition 129

5.2.2. Model building and training process 130

5.3. Results and discussion 135

5.3.1. Exploration of synthetic EIS dataset with a VAE model 135

5.3.2. Visualization and prediction of EIS parameter with a VAE model 145

5.4. Conclusions 152

Chapter 6. CONCLUSION 153

6.1. Conclusion 153

References 156

논문요약 174

Table 1-1. Performance of sensor electrode based on MXene, g-C3N4, and h-BN materials. 28

Table 1-2. Performance of sensor electrode based on MOFs, and TMDs materials. 31

Table 3-1. Comparison of performance of the sensor developed in the present study with other reported electrochemical NADH sensor. 78

Table 4-1. Electrochemical and physical parameters and range. 94

Table 4-2. List of hyperparameters with the best fit values 104

Table 4-3. Comparison of R-squared of the predicted EIS parameters on the different circuit models. 118

Table 5-1. Comparison of R-squared of the predicted EIS parameters on the different circuit models. 149

Figure 1-1. Illustration of (a) conventional approach for electrochemical sensor and (b) ML approach for electrochemical sensor. 22

Figure 2-1. Four type of commonly used detection methods in electrochemical sensor. (a) cyclic voltammetry, (b) differential pulse... 38

Figure 2-2. Illustration of voltammogram of the redox reaction of Fc⁺ solution. 38

Figure 2-3. Illustration of (a) Nyquist plot and (b) Bode plot for EIS data. 40

Figure 2-4. EIS spectrum simulate form the respective equivalent circuit. 43

Figure 2-5. (a) The linear regression algorithm and (b) the k-nearest neighbor algorithm. 45

Figure 2-6. Performance comparison between modern ML and classical ML. 47

Figure 2-7. Performance comparison between modern ML and classical ML. 47

Figure 2-8. Illustration of convolution algorithm in the CNN. 49

Figure 2-9. Illustration of the pooling method. 49

Figure 3-1. FESEM images of (a) bulk WS₂ and (b) ex-WS₂. HRTEM images of (c, e) bulk WS₂ and (d, f) ex-WS₂. (g) XRD pattern of bulk-WS₂ and ex-WS₂ 64

Figure 3-2. (a) EDX spectrum of bulk-WS₂, (a) EDX spectrum of ex-WS₂, and (c) UV-Vis spectrum of bulk-WS₂ and ex-WS₂ 65

Figure 3-3. (a) CV response of ex-WS₂/ SPCE, bulk WS₂/SPCE and bare SPCE measured in 0.1M PBS solution containing 1mM NADH. (b) The dependance of... 67

Figure 3-4. Baseline corrected CV data showing the oxidation peak of bare SPCE and ex-WS₂/SPCE measured in 0.1 M PBS solution containing 1 mM NADH. 68

Figure 3-5. CV response data of ex-WS₂/SPCE in 0.1 M PBS solution containing different concentrations of NADH (0 mM to 5 mM). 69

Figure 3-6. CV data, effect of NADH concentration on Bulk WS₂/SPCE electrode. 69

Figure 3-7. (a) A plot between -Z" and log (frequency) measured for ex-WS₂/SPCE in 0.1 M PBS solution containing 1mM NADH at various applied DC... 72

Figure 3-8. (a) Nyquist plot measured for ex-WS₂/SPCE in different concentrations of NADH viz. 0 mM to 5 mM. (b) Their respective Bode plot representation. 73

Figure 3-9. (a) Bode plot measured for ex-WS₂/SPCE in 0.1 M PBS containing wide range of NADH concentrations (2 µM to 2048 µM). (b) Calibration plot... 75

Figure 3-10. (a) Schematic representation of the NADH sensing mechanism of ex-WS₂/SPCE. (b) Interference study showing the value of mod |Z| versus... 82

Figure 4-1. Five different equivalent circuit models. 93

Figure 4-2. The example of the EIS spectrum from five equivalent circuits. The orange and blue lines present possible EIS spectrum patterns for the respective circuits. 93

Figure 4-3. Input data for DNN models; (a) imaginary part of impedance (Z"), (b) phase angle (φ), and (c) magnitude of the impedance (|Z|). (d-f) show... 97

Figure 4-4. Illustration of the feature extraction of six input data using 1D convolution with a kernel size of 32. The solid red box shows the size of the... 98

Figure 4-5. The scheme of the neural networks used for (a) the EIS circuit classification and (b) parameters regression. 99

Figure 4-6. The effect of hyperparameters; (a) effect of the number of data on the accuracy, (b) effect of batch size and dropout rate on the accuracy, and (c)... 104

Figure 4-7. The training curve comparison of the DNN model with and without flipping features. The red color presents a DNN model without flipping features... 106

Figure 4-8. The training result of the validation set. (a) Confusion matrix for the DNN classification model using the optimal hyperparameter evaluated on the... 106

Figure 4-9. (a) The confusion matrix and (b) receiver operating characteristic (ROC) curve for the DNN classification were evaluated on the test set, yielding... 107

Figure 4-10. The example of correct classification results from the DNN model with the probability distribution on the right-hand side, in the case of (a) C3,... 110

Figure 4-11. The example of misclassification results from the DNN model with the probability distribution. (a) C3, (b) C4, and (c) C5. 111

Figure 4-12. The noise dataset with various signal-to-noise ratios (SNR) of (a) SNR 80 dB, (b) SNR 40 dB, and (c) SNR 20 dB, and their confusion matrix... 114

Figure 4-13. A demonstration using the DNN model for real EIS data prediction. 114

Figure 4-14. The prediction of C3 parameters for (a) Rs, (b) R1, (c) Q1, and (d) σ. Presented by first 100 spectra with R-squared and MAE value. The... 117

Figure 4-15. A demonstration using the DNN model for NADH concentration prediction; (a) EIS spectrum of NADH with different concentration, (b) The... 119

Figure 5-1. The schematic of (a) machine learning "black box" model and (b) conventional interpretation method for the classification EIS data. 125

Figure 5-2. The schematic of a deep generative model (VAE) for the EIS data classification. 128

Figure 5-3. Five different equivalent circuit models. 128

Figure 5-4. Schematic diagram of (a) VAE with classification model (b) VAE with regression model. 133

Figure 5-5. The schematic of a supervised VAE model for the EIS data classification. 134

Figure 5-6. The classification result with the confusion matrix and ROC curve on (a) circuit C1 to C4 and (b) circuit C1 to C5. 137

Figure 5-7. The classification result of the VAE model on the test set with the confusion matrix. 137

Figure 5-8. The classification result of (a) K-Nearest Neighbor model, (b) shallow neural network model, (c) deep neural network model, and (d) VAE model. 139

Figure 5-9. (a) the visualization of latent space in the EIS dataset and (b) the EIS reconstruction result using the respective latent variables (in the red... 142

Figure 5-10. Beeswarm plot of (a) feature impact on L1 and (c) feature impact on L2. Input data with colored the feature impact to the value of (b) L1 and (d) L2. 143

Figure 5-11. Illustration of the effect of the feature on the out value of L₁ and L₂. 146

Figure 5-12. The prediction of C3 parameters for (a) Rs, (b) R1, (c) Q1, and (d) σ. The first 50 spectra with an R-squared value are displayed. The orange... 148

Figure 5-13. Latent space with the mapping of (a) Rs, (b) R1, (c) Q, and (d) σ. 151

Figure 5-14. The reconstructed EIS spectrum from the latent variable. 151