AI-Driven Differentiation and Quantification of Metal Ions Using ITIES Electrochemical Sensors

Abstract

Electrochemical sensors, particularly those based on ion transfer at the interface between two immiscible electrolyte solutions (ITIES), offer significant advantages such as high selectivity, ease of fabrication, and cost effectiveness for toxic metal ion detection. However, distinguishing between cyclic voltammograms (CVs) of analytes with closely spaced half-wave potentials, such as Cd²⁺ and Cu²⁺, remains a challenge, especially for non-expert users. In this work, we present a novel methodology that integrates advanced artificial intelligence (AI) models with ITIES-based sensing to automate and enhance metal ion detection. Our approach first employed a convolutional neural network to classify CVs as either ideal or faulty with an accuracy exceeding 95 percent. Ideal CVs were then further analyzed for metal ion identification, achieving a classification accuracy of 99.15 percent between Cd²⁺ and Cu²⁺ responses. Following classification, an artificial neural network was used to quantitatively predict metal ion concentrations, yielding low mean absolute errors of 0.0158 for Cd²⁺ and 0.0127 for Cu²⁺. This integrated AI–ITIES system not only provides a scientific methodology for differentiating analyte responses based on electrochemical signatures but also substantially lowers the expertise barrier for sensor signal interpretation. To our knowledge, this is the first report of the AI-assisted differentiation and quantification of metal ions from ITIES-based CVs, establishing a robust framework for the future development of user-friendly, automated electrochemical sensing platforms for environmental and biological applications.

1. Introduction

Toxic metal contamination from cadmium, arsenic, and lead, etc., poses serious health risks due to bioaccumulation in the food chain. These metals are known to contribute to various physiological and neurological disorders, including cognitive deficits, kidney dysfunction, and developmental delays in children, making their detection and monitoring critically important [1,2]. While detecting toxic metals in environmental samples is less complicated, their detection in biological samples remains a significant challenge. Non-electrochemical techniques, such as atomic absorption spectroscopy [3], inductively coupled plasma mass spectrometry (ICP-MS) [4] chromatography [5], and colorimetric assays [6] are highly sensitive and widely used for metal toxicity analysis. However, these methods often require expensive instrumentation, trained personnel, and complex sample preparation steps that can alter metal speciation, a critical factor influencing metal toxicity [7], and thus limit their applicability in field-based or real-time monitoring [8,9].

In contrast, electrochemical methods offer a cost-effective alternative with minimal sample preparation and real-time analysis capabilities. Redox-based and non-redox-based electrochemical techniques have been explored for metal detection, with surface-modified electrodes enhancing sensitivity. Shan et al. used a graphene/graphene oxide-modified electrode for Cd²⁺ detection via differential pulse voltammetry (DPV) [10]. Sacara et al. employed square wave anodic stripping voltammetry with modified glassy carbon electrodes for Cd²⁺ detection [11]. More recently, fast-scan cyclic voltammetry (FSCV) has been optimized for detecting multiple metal ions, including Cu²⁺ [12,13], Cd²⁺ [14], and As³⁺ [15]. Despite their sensitivity, redox-based electrochemical methods rely on electron transfer reactions at the electrode surface, which can produce unwanted signals and limit detection to redox-active metal ions. These reactions often lead to overlapping voltammetric peaks and electrode fouling, particularly in complex matrices such as biological fluids, where multiple redox-active species may coexist. Ion transfer at the interface of two immiscible electrolyte solutions (ITIES) offers a promising alternative, enabling the detection of both redox-active and redox-inactive ions based on their transfer across aqueous and organic phases under an applied potential [16]. This technique measures ion transfer currents rather than faradaic redox processes, significantly reducing signal interference. Since detection is governed by intrinsic ion properties such as charge, size, and hydrophobicity, ITIES-based sensors often do not require additional surface modification strategies for selectivity. Moreover, the associated electrodes are generally simpler and more cost-effective to fabricate compared with conventional redox-based systems. These advantages make ITIES a compelling platform for metal ion sensing, particularly in complex or biologically relevant environments.

Originally developed for biological analytes, ITIES has been extended to metal ion detection. The Shen group demonstrated that tris(crown ether) ionophores facilitate ion transfer at nanopipette-based ITIES electrodes, enabling lower transfer potentials [17]. Ishimatsu et al. investigated ion transfer kinetics for Ag⁺, K⁺, Ca²⁺, Ba²⁺, and Pb²⁺ using ionophore-doped membranes [18]. Our group recently employed ITIES-based sensors using glass nanopipettes for the detection of Cd²⁺ in both simple matrices, such as electrolyte solutions, and complex environmental samples, including artificial seawater [19]. We performed a comprehensive evaluation of the sensor’s performance, assessing its sensitivity, limit of detection (LOD), selectivity, and stability. Furthermore, we demonstrated the sensor’s applicability in analyzing a water sample collected from a local lagoon, with results comparable to those obtained using previously reported ICP-MS data. Despite the advancements in ITIES-based sensing platforms, interpreting electrochemical data remains challenging, particularly when signals from multiple analytes overlap. Although these systems offer the advantage of redox-independent ion detection, the electrochemical signatures of different ions are often closely spaced, especially in complex or biologically relevant matrices. This makes it difficult to resolve individual ion transfer events with high confidence. Moreover, while the fabrication of ITIES electrodes is relatively straightforward, maintaining a stable and reproducible interface presents practical challenges. Glass capillaries are commonly used to construct these electrodes due to their availability, low cost, and ease of manipulation [20]. However, the hydrophilic nature of glass necessitates surface modification with hydrophobic materials, such as chloromethylsilane, to retain the less polar organic phase inside the capillary [20]. This silanization process is highly sensitive to factors such as humidity and curing time, often resulting in variability in coating uniformity. Inconsistent surface properties can lead to unstable liquid interfaces and altered ion transfer kinetics, all of which contribute to signal distortion, baseline drift, and reduced reproducibility. Taken together, these issues highlight the need for more robust data analysis strategies and tighter fabrication control to fully realize the analytical potential of ITIES in multi-analyte detection. Artificial intelligence (AI) -based approaches are emerging as powerful tools to address these challenges. These AI methods offer significant advantages over classical analytical techniques, particularly when dealing with complex, high-dimensional, or noisy datasets. Traditional methods often rely on predefined calibration curves and manual interpretation, which can be time-consuming, sensitive to operator bias, and prone to errors in complex matrices. In contrast, AI models, such as artificial neural networks (ANNs) and convolutional neural networks (CNNs) can learn patterns from large datasets, automatically correct for signal variability, and distinguish subtle differences between analytes. They also enable real-time data processing, eliminate the need for extensive calibration, and enhance selectivity and sensitivity by filtering out noisy or faulty signals. These capabilities make AI particularly valuable for rapid, on-site analysis and for improving the robustness and accuracy of sensing technologies.

Several studies have successfully integrated AI with electrochemical sensors to enhance data processing and interpretation. Wang et al. developed a dual-modal fluorescence and electrochemical sensor combined with a one-dimensional CNN (1D CNN) for the simultaneous detection of Cd²⁺ and Pb²⁺ in complex water samples [21]. Similarly, Zhang et al. integrated 1D CNN with square wave anodic stripping voltammetry (SWASV) for Pb²⁺, Cd²⁺, and Cu²⁺ detection, achieving a high accuracy (R² > 0.99) and low relative errors (<5%) [22]. Choi et al. compared deep learning models with principal component regression for FSCV data analysis, implementing a 1D CNN using TensorFlow and Keras to estimate neurotransmitter concentrations from voltammograms [23]. Kudr et al. developed an automated electrochemical system using differential pulse anodic stripping voltammetry combined with a multilayer perceptron ANN to detect Zn, Cd, Cu, and Pb ions in environmental samples [24].

Specifically for cyclic voltammogram (CV) analysis, Cetó et al. [25] developed a bio-electronic tongue (BioET) system integrating cyclic voltammetry with a four-sensor array and a Bayesian regularization ANN for resolving overlapping electrochemical signals from wine polyphenols. The ANN used 23 manually selected current values from cyclic voltammograms as inputs to predict polyphenol concentrations without chromatographic separation, achieving a high accuracy (slopes > 0.97) for synthetic samples. Despite the promising results, the study was limited by a relatively small dataset, consisting of only 37 total samples, 27 used for training and 10 for testing. Such a limited dataset constrains the model’s ability to generalize and reduces predictive reliability, particularly for real-world applications where sample variability is significantly higher. Additionally, the data were extracted manually, introducing the possibility of selection bias. Kennedy et al. [26] developed a CNN model for classifying CVs according to three fundamental electrode reaction mechanisms (E, EE, and EC) by treating CVs as grayscale images to leverage image recognition techniques. Despite automating electrochemical process identification, the model showed significant limitations: it was trained exclusively on synthetic data generated by the MECSim simulation package without accounting for experimental artifacts, addressed only basic reaction mechanisms while excluding more complex pathways (EC’, ECE, and catalytic cycles), and demonstrated a susceptibility to high noise levels, with noise amplitudes ≥20% causing misclassification. These constraints highlighted the need for incorporating experimental data and expanding mechanistic scenarios in future work to enhance model robustness for real-world applications.

Previous studies that focus on the classification of data or electrochemical analysis include Bonet-San-Emeterio et al. [27] who integrated a graphene-modified electrode with discrete wavelet transform (DWT)-processed ANNs to resolve overlapping cyclic voltammetry signals from quaternary mixtures of dopamine, serotonin, ascorbic acid, and uric acid. The approach compressed raw CV data (206 points) into 55 approximation coefficients via Daubechies-3 decomposition before feeding them into a 55-9-4 feedforward ANN. When trained on 36 synthetic mixtures and tested on 9 samples, the system achieved an excellent performance (correlation coefficients > 0.974, normalized RMSE of 0.0583), outperforming partial least squares regression (NRMSE = 0.0918). However, the limited dataset (45 total samples) restricted the model’s applicability to complex biological matrices where imperfections such as noise and other factors impact the quality of the signal, highlighting challenges for clinical implementation. The Trammell group [28] investigated the application of artificial intelligence and machine learning for classifying heavy metals (Cd²⁺, Cu²⁺, Hg²⁺, and Pb²⁺) and explosives in seawater using cyclic square wave voltammetry (CSWV). The study evaluated seven distinct models, including deep learning architectures (FCNs, LSTM, LSTM-FCN, and ALSTM-FCN) and traditional methods (PCA-SVM, LDA, and 1NN-DTW), with CSWV data formatted as time series inputs. While the deep learning models achieved a strong performance (ROC AUC greater than 0.99 for metal classification), the study was constrained by limited dataset sizes (36 to 80 samples per metal class and only 6 samples per explosive class) and high computational demands. For example, the LSTM model required 2000 training epochs and access to graphics processing unit (GPU) hardware, making it less feasible for deployment in resource-limited environments. Although this work demonstrated the potential of combining CSWV with advanced AI architectures, it also highlighted the need for larger, more diverse training datasets and computationally efficient models to ensure practical applicability. In another study, the same group developed a multilayer epitaxial graphene electrode on silicon carbide for the multiplex detection of heavy metals, herbicides, pesticides, and industrial contaminants in seawater [29] using CSWV. The graphene electrodes outperformed conventional materials by delivering lower background currents, higher seawater stability, and effective oxygen reduction interference suppression. For data analysis, the researchers implemented an Attention-LSTM-Fully Convolutional Network (ALSTM-FCN) to classify contaminants directly from CSWV time series data, achieving ROC-AUC values above 0.90 for most targets while using class activation maps to identify key voltammetric features. However, the system showed significant limitations: classification accuracy notably declined for analytes below 500 ppb, and training the computationally intensive LSTM models required 2000 epochs on GPU hardware, highlighting substantial challenges for implementing this approach in real-time or field-deployable applications.

In contrast to prior studies, we employed a CNN to assess the quality of electrochemical signals, classifying them as ideal or faulty, arising from variations or instabilities at the liquid interface of our ITIES-based Cd²⁺ sensor, and to enhance selectivity in the presence of a competing metal ion [19]. Additionally, an ANN was used to predict metal ion concentrations (Figure 1). Our models are relatively light weight, requiring 20 epochs or less, and were trained on a comprehensive dataset that includes 2175 individual raw CV signals (Table S1). The CNN model was trained to first identify imperfection in signals, which is used to filter out good signals from faulty ones. This significantly improves the model’s accuracy while addressing some of the practical world concerns when deploying these models for field applications. Then the CNN model was used to differentiate Cd²⁺ signals from Cu²⁺, a known interfering ion. The ANN model was subsequently trained to predict metal ion concentrations, in which smart feature extraction was enabled to reduce the data sampling for the ANN model, which addresses some of the matrix interference which is observed during conventional external calibration methods typically required for electrochemical sensors. Based on our literature search, this is the first study to integrate AI with an ITIES-based sensor for both qualitative (signal quality and selectivity improvement) and quantitative (concentration prediction) purposes, demonstrating the potential of AI to revolutionize electrochemical metal ion detection.

2. Materials and Methods

2.1. Chemicals

All chemicals were purchased from Sigma-Aldrich (St. Louis, MO, USA). CdCl₂ and CuSO₄ were used as sources of Cd²⁺ and Cu²⁺ ions, respectively. Aqueous solutions of the metal ions were prepared in 0.3 M KCl.

2.2. Electrode Fabrication and Electrochemical Experiments

Electrodes were fabricated following previously reported methods [19]. Briefly, borosilicate glass capillaries (O.D. = 1.0 mm, I.D. = 0.58 mm; Sutter Instruments, Novato, CA, USA) were pulled into nanopipets with tip diameters of approximately 600 nm using a P-2000 laser puller. Pipet tips were imaged using a JEOL JSM-6380LV (Jeol Ltd. Tokyo, Japan) scanning electron microscope to confirm their dimensions. The inner walls of the pulled pipets were silanized by exposing them to chlorotrimethylsilane vapor under vacuum for 30 min to 1 h, depending on ambient humidity and temperature. The silanized pipets were then filled with an organic phase composed of 10 mM 1,10-phenanthroline (phen) and 0.1 M tetradodecylammonium tetrakis(pentafluorophenyl)borate (TDDATFAB) in dichloroethane (DCE). TDDATFAB was synthesized according to established procedures [30].

All electrochemical measurements were performed using a CHI660E potentiostat (CH Instruments, Bee Cave, TX, USA) in a three-electrode setup, with a lab-built Ag/AgCl electrode as the reference and a platinum wire (Alfa Aesar, Haverhill, MA, USA) as the counter electrode.

2.3. AI Modeling

AI techniques were applied to automate and improve the classification and quantification of electrochemical signals, CVs generated from ITIES-based sensors. MATLAB R2023b was used to generate CV images for CNN training and to preprocess numerical voltammetric data by extracting key features from regions with near-zero slope in the forward scan, thereby reducing dimensionality for ANN training. PyCharm 2024.1 served as the development environment for training both the CNN and ANN models. The CNN, based on a modified LeNet-5 architecture, was used to first distinguish ideal from faulty CVs, and then to classify ideal CVs as either Cd²⁺ or Cu²⁺. The ANN, implemented as a feedforward network with two hidden layers, was trained to predict ion concentrations directly from the preprocessed CV data.

3. Results and Discussion

3.1. Classification of Good and Faulty Signals for Cd²⁺ and Cu²⁺

ITIES is a non-redox-based electrochemical approach that facilitates the movement of charged ions between two distinct phases under an applied electric field. This process can be enhanced by incorporating a suitable ionophore into the organic phase, which selectively interacts with the analyte, reducing the activation energy required for ion transfer. As charged ions move across the interface, the resulting current is recorded, generating a cyclic voltammogram (CV). Each analyte exhibits a unique CV profile in a given medium, characterized by a distinct half-wave potential (E_1/2), which corresponds to the potential at which half of the maximum current is reached. This unique signature enables analyte identification.

In our previous study, we optimized the electrochemical parameters necessary for facilitating the transfer of Cd²⁺ from the aqueous phase into DCE in the presence of phen [19]. Although phen is not exclusively selective for Cd²⁺, we selected it as the preferred ionophore due to its affordability compared with other Cd²⁺-specific ionophores, such as ETH1062 [31], which are prohibitively expensive for developing point-of-care sensors. Our optimized conditions demonstrated a strong affinity of Cd²⁺ for phen; however, one notable interfering species is Cu²⁺. While a clear separation between the E_1/2 values of Cd²⁺ (~−0.55 V) and Cu²⁺ (~−0.43 V) was observed, along with differences in maximum current at equal concentrations, this distinction becomes increasingly challenging when Cd²⁺ is present at significantly lower concentrations than Cu²⁺. Furthermore, maintaining a stable liquid–liquid interface at the nanoscale presents a significant challenge, leading to the occurrence of distorted CVs alongside well-formed ones. These faulty CVs can closely resemble good CVs but exhibit subtle distortions, making differentiation difficult, particularly for non-expert users. To address this issue, we integrated AI into our data analysis workflow.

We collected over 700 ITIES CVs for each metal ion transfer at varying concentrations (Cd²⁺: 0.4 mM, 0.2 mM, 0.1 mM, and 0.05 mM; Cu²⁺: 0.2 mM, 0.1 mM, and 0.05 mM) from 0.1 M KCl to DCE in the presence of phen. Here, we first constructed a full calibration curve for Cu²⁺ and chose these concentrations from the linear range (Figure S1). Since we had already established the linear range of Cd²⁺ from our previous study [19], we did not construct a full calibration curve for Cd²⁺ but only collected data within the known linear range (Figure S2). From these, we identified four primary types of faulty CV signals for both Cd²⁺ and Cu²⁺: adsorptive, cross, incomplete, and noisy CVs (Figure 1 and Figure 2). As depicted in Figure 2a and Figure 3a, ideal CVs for Cd²⁺ and Cu²⁺ exhibit a sigmoidal shape with minimal hysteresis, where the steady-state plateau current can be reliably used for quantification. However, deviations from this ideal behavior result in faulty CVs. Adsorptive CVs (Figure 2b and Figure 3b) exhibit an extra peak or elevated background current due to interfacial adsorption. This effect is particularly pronounced in the Cd²⁺ CVs at the beginning of the scan, whereas in the Cu²⁺ CVs, a distinct peak emerges during the reverse scan (Figure 3b). Cross CVs (Figure 2c and Figure 3c) are characterized by a crossing of the forward and backward scans, often occurring due to unpredictable fluctuations in interfacial stability or the discharge of impurities, and are more prominent in Cu²⁺ CVs. Incomplete CVs (Figure 2d and Figure 3d) result when the applied potential window is insufficient to establish a stable steady-state, leading to truncated voltammograms. Although the potential window is generally maintained consistently, slight shifts can occur due to variations in the reference electrodes, particularly when working with lab-fabricated nanoscale electrodes. This issue is more pronounced in long-duration experiments involving multiple measurements. Noisy CVs (Figure 2e and Figure 3e) contain extra peaks or spikes, which may arise from instrumental noise, electrostatic interference, or electrochemical fluctuations, with noise artifacts being more frequently observed in Cu²⁺ CVs compared with Cd²⁺ CVs. The impact of these imperfections on analytical accuracy and precision varies with the metal ion; however, for ultra-trace detection, any deviation from an ideal CV is undesirable. To address this, we used these data to train our CNN and ANN models.

3.2. CNN Modeling

CNNs are a class of deep learning models designed to efficiently extract and analyze complex patterns from data. Initially developed for image recognition, CNNs have achieved remarkable success in diverse fields such as medical imaging, speech processing, and signal analysis. Their ability to learn hierarchical features makes them especially suitable for analyzing electrochemical signals, where subtle variations in CVs can provide critical analytical insights.

A typical CNN architecture (Figure 4) comprises convolutional layers, pooling layers, and fully connected layers. Convolutional layers apply learnable filters to the input data to detect local patterns such as peak structure, background current shifts, and hysteresis. In the context of electrochemical analysis, this allows CNNs to recognize localized variations and distinguish between ideal and distorted CVs. Pooling layers follow convolutional layers to reduce the spatial dimensions of feature maps, preserving key information while enhancing computational efficiency and reducing the risk of overfitting. Fully connected layers integrate the extracted features to perform tasks such as classification or regression, enabling the accurate identification of analytes or the quantification of metal ion concentrations.

One of the major advantages of CNNs over traditional machine learning models is their ability to learn features directly from raw data without requiring manual feature engineering [33]. Their use of local connectivity and parameter sharing minimizes computational complexity while improving generalizability, especially important when working with limited labeled datasets [34]. Furthermore, CNNs capture both low-level signal features (e.g., peak edges and noise artifacts) and higher-level patterns (e.g., overall CV shape and adsorption behavior), facilitating the robust differentiation between true electrochemical responses and background noise. This makes CNNs particularly effective for the detection and quantification of ultra-trace analytes in complex electrochemical environments.

The CNN implementation began by importing a dataset of CVs stored as PNG images. Each image was preprocessed through pixel value normalization and label encoding. The dataset was then divided into training, validation, and testing subsets to ensure robust model evaluation. For this study, a modified LeNet-5 architecture was employed, consisting of convolutional layers, max pooling layers, and fully connected layers. These modifications were tailored to suit the number of classification categories, the dataset size, and the resolution of input data. The model was compiled using binary cross-entropy as the loss function and optimized with the Adam optimizer. To enhance model performance and prevent overfitting, several training callbacks were implemented, including early stopping, learning rate scheduling, and model checkpointing. Upon training completion, the model’s performance was assessed using both validation metrics and test datasets. Visualization tools were used to track accuracy and loss over training epochs, and additional functions were developed to display predicted versus true labels for a subset of test images.

To effectively differentiate between Cd²⁺ and Cu²⁺ signals while accounting for imperfections in the CVs, we adopted a two-step approach. In the first step, we used a CNN model to classify CVs as either ideal or faulty for both Cd²⁺ and Cu²⁺ (Figures S3–S5). In the second step, only the ideal CVs were used to train and test the model for distinguishing between Cd²⁺ and Cu²⁺ signals (Figure 5). To minimize computational load and speed up model training during early-stage experiments, we utilized low-resolution CV images as the input for the CNN. This sequential setup enabled us to filter out low-quality data before analyte classification, resulting in a remarkable accuracy of 99.15% when distinguishing between the ideal CVs of Cd²⁺ and Cu²⁺. The initial classification step served as a quality control mechanism, ensuring that the final analysis was based on high-integrity data. This significantly enhanced the model’s performance, especially compared with previous studies that applied machine learning without addressing signal quality and reported a much lower accuracy [26].

Overall, the CNN model demonstrated strong classification capabilities in distinguishing ideal from faulty CVs for both analytes, with validation accuracies exceeding 95% (

Table 1. Comparison of CNN model performance in training, validation, and testing accuracies, along with learning rates, for five different comparisons [32].

CNN Classification	Training Accuracy	Validation Accuracy	Testing Accuracy	Learning Rate
Cd2+ vs. Cu2+	99.20%	99.15%	99.14%	0.0005
Good vs. Faulty for Cd2+	99.96%	96.12%	95.37%	0.001
Good vs. Faulty for Cu2+	99.62%	98.68%	98.34%	0.001
Faulty—Cd2+	96.80%	87.71%	92.45%	0.001
Faulty—Cu2+	93.63%	85.14%	89.19%	0.001

). When applied to only ideal CVs, the model achieved 99.15% accuracy in differentiating Cd²⁺ from Cu²⁺, indicating its ability to learn and detect unique peak features and signal patterns specific to each ion. The classification accuracy was somewhat lower (85–88%) when dealing with specific types of faulty CVs. One contributing factor is the presence of overlapping imperfections; for example, a CV might simultaneously exhibit both noise and adsorption effects, or noise and an incomplete signal. During testing, the model is designed to assign a single fault label. If the CV being tested predominantly exhibits one type of imperfection but is labeled with another, the model’s prediction may appear incorrect even if it correctly identifies one of the issues. These complexities, along with other sources of variability in the data, contributed to the reduced accuracy of the faulty CV classification. Nonetheless, the model’s ability to effectively screen and exclude compromised data before performing concentration analysis marks a substantial step forward. This approach enhances the reliability and robustness of electrochemical sensing platforms and lays the foundation for more accurate real-time applications.

We further evaluated model performance by analyzing line plots of accuracy and loss (for both training and validation) over multiple epochs across different classification tasks. Figure 6 illustrates these trends for the Cd²⁺ vs. Cu²⁺ comparison. A sharp increase in both training and validation accuracy occurs between epochs 2 and 4, indicating that the model rapidly begins to distinguish the elemental differences. Notably, validation accuracy briefly exceeds training accuracy, a common occurrence with small batch statistics—before both stabilize between 0.97 and 0.99. Correspondingly, training and validation loss decline smoothly from approximately 0.65 to 0.14 by epoch 10, with only a slight uptick in validation loss around epoch 8. This reflects remarkably fast convergence, achieving near-perfect separation in fewer than seven epochs.

Figure S6 presents the accuracy and loss trends for distinguishing between good and faulty CVs for Cd²⁺, evaluated over 20 epochs. The training accuracy increases steadily from approximately 0.72 to 0.99, while the validation accuracy climbs from around 0.86 to 0.97, closely mirroring the training curve. This parallel trajectory suggests strong learning with minimal overfitting. Both loss curves show a steep initial decline, leveling off around epoch 10. A minor rise in validation loss near epoch 8 likely reflects natural variation within the holdout set. Overall, the model demonstrates a fast and effective grasp of this binary classification, and training could likely be halted around 10–12 epochs with a negligible impact on performance.

Finally, Figure S7 shows the training and validation curves for classifying faulty CVs of Cu²⁺. Here, the training accuracy rises from roughly 0.52 to 0.94 by epoch 17. However, the validation accuracy lags slightly, plateauing around 0.85 and exhibiting more fluctuation. While training loss drops dramatically from 1.1 to 0.18, validation loss settles around 0.44 and varies slightly. The noticeable gap between the training and validation accuracy and loss suggests moderate overfitting, indicating that this particular classification task is more challenging. Improvements may be achieved by increasing the dataset size or incorporating regularization techniques such as dropout or weight decay.

3.3. ANN Modeling

ANNs are machine learning models inspired by the structure and function of biological neural networks. They consist of interconnected nodes, artificial neurons, organized in layers, which process and transmit information through weighted connections. These weights are optimized during training, allowing the network to learn complex, non-linear relationships in data. ANNs have been widely applied across domains such as image and speech recognition [35], natural language processing [36], and scientific data analysis [37]. Each artificial neuron in an ANN receives an input, applies a weight, and produces an output through an activation function. This non-linear activation enables the network to model intricate patterns. A typical ANN consists of an input layer, one or more hidden layers, and an output layer. While the input layer receives raw data, the hidden layers perform feature transformation, and the output layer produces predictions or classifications. The training process employs algorithms like backpropagation to adjust the weights based on the error between the predicted and actual outputs, minimizing this error through iterative optimization. In electrochemical applications, ANNs offer a robust approach for analyzing complex voltammetric data, such as CVs. By training them on labeled datasets with known analyte concentrations, ANNs can learn to correlate specific features of CVs with corresponding concentration values. Once trained, the model can predict analyte concentrations in new, unseen CVs, offering a rapid and accurate tool for quantitative analysis. This approach eliminates the reliance on traditional calibration curves for concentration determination. While it is relatively straightforward to simulate simple matrices for separate calibration studies, replicating the complexity of real-world environments, such as rivers, lagoons, and other natural water bodies remains challenging due to rapid fluctuations in metal speciation and water composition. In such contexts, the ability to reliably identify toxic metal ions like Cd²⁺ and Cu²⁺ and accurately predict their concentrations is essential for maintaining water quality and protecting public health.

In this study, we implemented an ANN model using a dataset of CVs represented by their x and y coordinates, extracted from text files. Unlike CNNs, which operate on image data, ANNs require numeric vectors as input. A key challenge in this process was ensuring consistency in the number of data points across all CVs in the dataset. For example, a CV with 800 data points could not be processed alongside one with only 750, as the ANN requires fixed-dimensional input vectors for each category, Cd²⁺ or Cu²⁺. To address this, we experimented with several data preprocessing strategies aimed at standardizing the number of (x, y) coordinate pairs per CV (e.g., reducing all to 700 points). These included trimming data from the tail end of the scan and using only the forward scan. Ultimately, we found that it was most effective to focus on the points that truly reflected concentration-related features. Specifically, we developed a MATLAB code that automatically identifies key regions in the CV—areas where the slope of the curve is approximately zero. The code selects a single point from the initial flat baseline and another from the final plateau region of the forward scan. These two representative points (Figure S6) were sufficient to achieve a high predictive performance when used as the input for the ANN model, minimizing complexity without compromising accuracy. Interestingly, the ANN’s ability to utilize the absolute current reading from the plateau region of ITIES–CVs provided an effective solution, particularly in cases where the baseline current fluctuates or is non-zero due to various unpredictable experimental conditions.

The dataset was then divided into training and testing sets. Our ANN architecture consisted of a feedforward neural network with an input layer of 64 neurons, followed by two hidden layers of 32 neurons each, and a single output neuron for regression. In total, the model had 2433 trainable parameters (Figure S8). The hidden layers utilized the ReLU activation function, while the output layer employed a linear activation to predict concentration values. We used the Adam optimizer with an initial learning rate of 0.01 and mean squared error (MSE) as the loss function. To prevent overfitting, we incorporated early stopping with a patience of 5 epochs and used a learning rate scheduler to halve the learning rate every 5 epochs for better convergence. Throughout training, model performance was tracked using both mean absolute error (MAE) and MSE metrics. Post training, the predicted values were visualized and compared with the actual concentrations, confirming the model’s accuracy. Notably, the ANN also performed well on CVs not seen during training, demonstrating strong generalization and making it a promising tool for real-world applications in electrochemical sensing.

Our ANN model demonstrated strong predictive capabilities in estimating the concentrations of Cd²⁺ and Cu²⁺ analytes from CV data, as shown in

Table 2. Comparison of the ANN model’s performance in terms of MAE, MSE, and learning rate for Cd²⁺ and Cu²⁺ CVs [32].

CV	MAE	MSE	Learning Rate
Cd2+	0.0158	4.0896 × 10−4	0.01
Cu2+	0.0127	2.6129 × 10−4	0.01

and

Table 3. Comparison of ANN-predicted and actual concentrations (randomized) for Cd²⁺ and Cu²⁺, demonstrating close agreement between predicted and true values [32].

Cd2+ Concentration (mM)		Cu2+ Concentration (mM)
Actual	Predicted	Actual	Predicted
0.10	0.120909	0.05	0.032442
0.40	0.392451	0.05	0.039273
0.05	0.036781	0.05	0.056565
0.10	0.120909	0.10	0.109722
0.05	0.036781	0.05	0.017079
0.20	0.182019	0.10	0.099719
0.05	0.058828	0.20	0.209858
0.40	0.392429	0.05	0.046686

. On the test dataset, the model achieved low MAE and MSE values, indicating a high accuracy and robustness in its regression performance. Specifically, for Cd²⁺, the model reached a test MAE of 0.0158 and a test MSE of 0.00040896, while for Cu²⁺, the MAE was 0.0127 and the MSE was 0.00026129. These results highlight the ANN’s ability to accurately predict analyte concentrations based on preprocessed CV inputs (Table 2). The model demonstrated a consistent performance across various concentration ranges, as evidenced by the close agreement between the predicted and actual values shown in Table 3. This consistency highlights the ANN’s strong generalization capability beyond the training data. Furthermore, as illustrated in Figure 7, the variance in predicted concentrations for both Cd²⁺ and Cu²⁺ remained low across the tested ranges, further confirming the model’s ability to reliably capture the underlying electrochemical behavior. The minor scatter or deviations observed are likely due to experimental noise or slight model inaccuracies, but these remain within acceptable limits for practical applications.

To facilitate evaluation, predicted concentrations were compiled alongside true values in a DataFrame, allowing for the clear visual assessment of model accuracy. This comparison not only confirms the model’s effectiveness but also aids in identifying any potential deviations or systematic trends. Importantly, the ANN also showed excellent performance when applied to unrecognized CV data, successfully predicting concentrations outside of the training set. This demonstrates the model’s generalization ability and highlights its practical potential in real-world applications. Whether for environmental monitoring or analytical sensing, the ANN proves to be a promising tool for the rapid and accurate quantification of heavy metal ions in complex samples.

Overall, our approach demonstrates several key advantages over previously reported studies, as summarized in

Table 4. Comparison of related studies integrating AI modeling with electrochemical techniques.

Ref.	Type of Sensor	Electrochemical Technique	Target Analytes	AI Model	Number of Datasets	Pros/Cons
[25]	Graphite–epoxy biosensor array, modified with enzymes and Cu nanoparticles	Cyclic Voltammetry	Catechol Caffeic acid Catechin	ANN	37 signals (27 train, 10 test)	Redox-active analytes only Small dataset used for training Features manually selected—possible selection bias
[26]	Synthetic generated data	Cyclic Voltammetry	Ferrocene polyoxometalate complexes	CNN	Synthetic data only	Simple models—lacks complexity Signal generation via simulation package Lacks non-ideal behavior in signals
[27]	Graphene-modified carbon electrode	Cyclic Voltammetry	Dopamine Serotonin Ascorbic acid Uric acid	ANN + DWT	45 signals (36 train, 9 test)	Redox-active analytes only Small dataset used for training Complex preprocessing of data leads to loss of valuable information
[28]	Screen-printed electrodes (carbon-based)	CSWV	CuSO4, PbCl2, HgCl2, CdCl2, Paraquat, Diquat, TNT, TNB, Bisphenol-A, Nonyl phenol	FCN, LSTM, ALSTM-FCN	36–80 signals per class for metals, 6 signals per class (explosives)	Redox-active analytes only Model trained on small dataset Intensive computational work with 2000 epochs on GPU hardware
[29]	Multilayer epitaxial graphene on silicon carbide substrate	CSWV	CuSO4, PbCl2, HgCl2, CdCl2, Diquat dibromide (DQBr2), Paraquat dichloride, Methyl parathion, Bisphenol A	ALSTM-FCN	16 samples per class including heavy metals and herbicides, total of 128 signals	Redox-active analytes only Small dataset used for training Multiple analytes without focus on closely interfering signals Declined classification accuracy for lower concentrations of analytes Intensive computational work with 2000 epochs on GPU hardware
This Study	Borosilicate glass-based electrode	ITIES	CdCl2, CuSO4	ANN, CNN	Large dataset with randomized data split: training—1522, validation—435, test—217	Analytes need not be redox-active Accounts for both ideal and faulty signals CNN—utilized raw unprocessed data, avoiding loss of valuable information ANN—data extracted using MATLAB code which removes selection bias Concentration prediction using ANN model Light weight model that converges within 20 epochs No GPU is needed

. Prior work has largely focused on redox-based electrochemical methods, with no attention given to non-redox-based techniques such as ITIES. Additionally, earlier studies primarily concentrated on analyzing ideal analyte signals, often overlooking the evaluation of electrode conditions or the identification of erroneous signals. In contrast, the first phase of our study employed a CNN model specifically designed to filter out faulty signals for both metal ions, enhancing the reliability of the data used for analysis. Furthermore, the use of an ANN to accurately predict metal ion concentrations without relying on external calibration curves represents a significant advancement, particularly for studies involving complex matrices. This feature allows for broader applicability in real-world environments where traditional calibration is challenging. Notably, our models were trained on a large dataset but required only minimal computational resources, with no need for a GPU, making the approach both scalable and efficient compared with other AI-integrated electrochemical studies. Additionally, while the use of reagents does involve some environmental considerations, our method uses only small volumes of organic solvents, and no heavy metals are introduced beyond those being measured. As a result, both the sensor design and AI modeling framework are environmentally friendly and cost-effective.

4. Conclusions

In this study, we developed an AI-enhanced sensing platform that integrates CNNs and ANNs with ITIES-based sensors for the accurate and reliable detection and quantification of Cd²⁺ and Cu²⁺ ions. Using a CNN architecture, we first established an effective screening process capable of distinguishing between ideal and faulty CVs, achieving over 95% accuracy in identifying compromised signals. Subsequently, the CNN successfully discriminated between Cd²⁺ and Cu²⁺ responses in ideal CVs, even when the half-wave potentials were closely spaced. The ANN model further enabled the precise prediction of analyte concentrations, demonstrating low mean absolute errors for both metal ions and strong generalization to previously unseen data. This capability is particularly valuable for in vivo electrochemical experiments, where traditional methods rely on in vitro calibration curves that often suffer from translatability issues due to differences between the biological and calibration matrices. Our ANN-based approach eliminates the need for a calibration curve by directly predicting analyte concentrations from the CV signals, offering a promising strategy for future in vivo studies.

This dual-model framework not only addresses a long-standing challenge in ITIES-based electrochemical sensing, the qualitative and quantitative interpretation of voltammetric data, but also reduces the level of operational expertise required for accurate analysis. Our results demonstrate that AI-driven data analytics can significantly enhance both the selectivity and accuracy in electrochemical detection, paving the way for more robust and user-friendly sensor technologies for environmental and potentially biological applications. To the best of our knowledge, this is the first report demonstrating the use of AI to both classify and quantify metal ion signals from ITIES-based sensors based on CV shape and key electrochemical features. The workflow established here can be readily expanded to incorporate additional analytes or modified sensor designs. Future work will focus on enlarging the training datasets, integrating additional sources of signal variation, and exploring transfer learning strategies to further enhance model adaptability and real-world applicability. Overall, the convergence of ITIES electrochemistry and AI paves the way for the next generation of point-of-care and on-site monitoring solutions for toxic metal contaminants.