1. Introduction

The field of machine learning (ML) is rapidly evolving, fueled by advances in computational power, algorithmic complexity and exponential growth in data availability [1]. This evolution of ML has enabled it to become one of the beneficial means on scientific exploration and technological innovation. ML, an integral branch of artificial intelligence, encompasses a variety of algorithms capable of learning from data and making predictions or decisions [2]. Among these, linear regression, support vector machines (SVM) random forests (RF), artificial neural networks (ANN), and extreme gradient boosting trees (XGBoost) stand out for their adaptability and robustness [3-6]. Each method has unique advantages, ranging from the interpretability of linear regression to the predictive power of deep neural networks. ML has demonstrated extraordinary capabilities in deciphering complex patterns, improving prediction accuracy, and promoting innovation in different scientific fields [7].

Electrochemical oxidation processes are of particular importance with the growing need for the treatment of environmental pollutants, the optimization of energy storage systems and the development of materials science [8,9]. Electrochemical oxidation processes are pivotal in addressing sustainability issues through the removal of contaminants, extending the life and efficiency of batteries, and synthesizing new materials with tailored properties [10,11]. However, the intricate dynamics of electrochemical reactions, influenced by numerous variables such as electrode material, electrolyte composition, and operating conditions, pose significant analytical and optimization challenges [12-14]. Traditional experimental approaches, while invaluable, are often time-consuming and may not capture the full spectrum of possible reaction pathways and outcomes.

ML, a powerful tool for modelling complex non-linear relationships, has achieved unparalleled success in image recognition, natural language processing and predictive modelling, demonstrating their potential to reveal the complexity of electrochemical systems [15,16]. Zhang et al. developed an accurate battery prediction system by combining electrochemical impedance spectroscopy with gaussian process regression (GPR) algorithm. The entire spectrum was used as input to train a GPR model, which automatically determined spectral features that predict degradation and accurately predicted the remaining battery lifetime without full information about the battery's past operating conditions [16]. The advantages of applying ML to electrochemical oxidation and related fields are manifold. Large datasets are analyzed by ML algorithms to identify patterns and relationships that are not immediately apparent to researchers, thus optimizing reaction conditions and reducing the need for complex experiments [17]. Sun et al. developed five ML models to predict the maximum value of the reaction rate using multiple electrochemical reaction conditions and quantum parameters of the target pollutant as input values. Subsequently, a particle swarm optimization (PSO) algorithm was used to predict the optimal reaction conditions in reverse, and the relative standard deviations of the predictions were verified to be < 5% through multiple experiments [18]. This predictive capability is especially advantageous in the design, development, and optimization of novel materials for electrochemical applications [19]. Furthermore, the flexibility of ML in dealing with different types of data makes it an ideal tool for integrating information from a variety of sources (e.g., experimental data, theoretical models, and computational simulations). This approach can lead to a more comprehensive understanding of electrochemical processes, from the macroscopic behavior of electrochemical oxidation to the microscopic mechanisms at the electrode-electrolyte interface [20,21].

In conclusion, the convergence of ML and electrochemical oxidation is expected to open up new frontiers in environmental management, energy storage, and material science. In this work, we firstly analyzed the hotspots at the intersection of ML and electrochemical oxidation through bibliometrics, highlighting on the importance and research trends of ML in the four dimensions of the field of electrochemical oxidation (pollutant removal, battery remediation, substance synthesis, and prediction of material characterization). We then outlined the mathematical principles and concepts of common ML algorithms including multivariate linear regression (MLR), SVM, XGBoost, ANN and RF. Subsequently, the applications of ML models in data analysis prediction capabilities in the four areas mentioned above were discussed, including predictive modelling, process control and material surface reaction mechanism prediction. This underscores the promising outlook of ML in these key domains. Finally, based on the results of the review, we presented future perspectives on data monitoring, model improvement, and autonomous detection with a special focus on the strengths and weaknesses of different ML methods applied to different areas in electrochemical oxidation. As such, this review aims to provide new paths for environmental management, energy innovation, and sustainable development.

2. Bibliometric analysis

Although much attention has been given to ML combined electrochemical oxidation, no studies have systematically summarized the mechanism and application of ML in electrochemical oxidation. Moreover, no previous review used the scientometric approach to map out the linkage or relationships among published papers in this field. In order to bridge these gaps in extant literature, mine high-frequency research hotspots in the field of ML in electrochemical oxidation, and explore possible mechanisms and applications in this field, this study used CiteSpace software to conduct a visual bibliometric analysis. Research trends and emerging hotspots in the application of ML to electrochemical oxidation can be identified through CiteSpace. The co-occurrences of citations and keywords are analyzed, revealing the primary research themes within the field and their evolution over time [22-24]. This not only facilitates a deeper understanding of the dynamics of ML applied in electrochemical oxidation but also effectively guides future research directions, enhancing the systematic and prospective nature of studies.

Based on CiteSpace software platform, the methods of bibliometrics, statistical analysis of data, keyword co-occurrence visualization, clustering and citation burst detection were adopted. The version of CiteSpace in this study was CiteSpace 6.1.R6. The specific operation settings were as follows: Keywords were set as nodes, respectively, for keyword co-occurrence and co-citation network mapping. The modified g-index in each slice was used for selection criteria. Clustering modes was implemented by the log-likelihood ratio (LLR) algorithm. In addition, the citation report function of web of science platform was used to have access to annual number of published papers. Literature search strategy was described in Text S1 (Supporting information). The flowchart of the bibliometric analysis using CiteSpace was shown in Fig. S1 (Supporting information).

302 documents were found in the database of web of science core collection from 2013 to 2022. As shown in Fig. S1, it can be found that the number of literatures published in this research field increased slowly and the number of articles per year did not exceed 10 before 2019 (7 articles in 2018). From 2019 to 2021, the number of published papers raised by > 20 per year. The most significant amplification of research in this field appeared from 2021 to 2022. From 2019 to 2022, 276 research studies were published, accounting for 91% of all the included studies published in the last 10 years. In general, the number of publications in the field of ML applied to electrochemical oxidation has showed a growth trend since 2013. It can be seen that this field has become a hot research field and has received widespread attention.

Keywords are precise generalizations of an article. The network (Fig. S2 in Supporting information) consists of nodes and links. The size of the node reflects the number of documents, namely the frequency of occurrence, containing the labelled keywords [25,26]. Table S1 (Supporting information) summarized top 20 frequently used keywords in the field, among which the most frequently used is ML, which is mentioned up to 113 times. Followed by degradation, model, performance, oxidation, lithium-ion battery, etc., which have been used > 25 times in the literature since 2014. These words macroscopically reflect the research framework for electrochemical oxidation related applications through ML. It also can be seen keywords, such as electrocatalyst, electrode, oxygen reduction, mechanism, prognostics (Table S1), reflecting that the research on electrode materials and surface reaction mechanisms in this field has been a research hotspot in recent years.

The articles with the same research content and the same citation are clustered. The greater the number of co-cited articles, the greater the correlation between the documents, and the higher the probability of clustering into one category. Fig. S3 and Table S2 (Supporting information) show that the clustering results include lithium-ion batteries, CO₂ reduction, ML, predict, etc. The largest cluster (#0) has 68 keywords within and a S value of 0.636, labeled as lithium-ion batteries, which includes keywords, such as degradation (2014), prediction (2015), diagnosis (2018), health estimation (2021) (Fig. S4 in Supporting information). It can be found that CO₂ reduction in the results is clustered into a single group and as the second largest category. Moreover, as two separate clusters, ANN and predict contain keywords such as extraction removal mechanism, cathode material, graphite anode that are highly related to electrochemical mechanisms and ML algorithmics.

Fig. S5 (Supporting information) presents the top 25 keywords with the strongest citation bursts. Among them, the first strongest burst keyword is carbon with beginning year 2019, which reflects that most studies focus on carbon related ML. More importantly, several keywords, for example, ion battery, electroreduction, electrooxidation, CO₂ reduction, water oxidation, related to electrochemical oxidation have indicated a relatively strong burst since 2021. And these keywords are hotly researched until 2023. Thus, ML research predicting applications related to electrochemical oxidation may be the focus of future research.

In general, the interdisciplinary research at the intersection of electrochemistry and ML exhibits a noticeable growth trend, signifying widespread attention to this field. Keywords such as degradation, model, performance, oxidation, lithium-ion battery, and CO₂ reduction reveal the thematic framework of the cross-disciplinary exploration between electrochemistry and ML. Therefore, ML has proven to be an effective technology applied in electrochemistry. As shown in Fig. 1, based on the results of bibliometric analysis, this review article provided a systematic overview of recent advances in environmental applications of ML for electrochemical pollutant treatment, battery repair and degradation, electrochemical generation and synthesis, and surface mechanism analysis of materials.

3. Mainly machine learning algorithms in electrochemical oxidation

This section analyzed and summarized the basic theory of common ML algorithms, including MLR and typical nonlinear regression (NR) algorithms such as SVM, XGBoost, ANN and RF.

The MLR model determines the optimal parameter estimates by minimizing the discrepancy between the actual observed values and the predicted values of model (usually the sum of squared differences). This model is capable of assessing and interpreting the weight of each independent variable with respect to the dependent variable [27]. Within the context of electrochemical oxidation, MLR can be utilized to understand how various process parameters such as current density, pH value, electrode material, and treatment time influence the efficiency of electrochemical oxidation. By analyzing data collected from experimental processes, MLR aids in predicting the outcomes of electrochemical treatment under varying conditions and in optimizing process parameters to achieve maximum efficiency [28]. The multiple linear regression equation is a mathematical model used to describe the linear relationship between a dependent variable (Y) and multiple independent variables (X₁, X₂, …, X_n). The equation is typically represented as follows (Eq. 1):

$ Y=\beta_1 X_1+\beta_2 X_2+\beta_3 X_3+\cdots+\beta_{\mathrm{n}} X_{\mathrm{n}}+\varepsilon $

(1)

where β_j (j = 1, 2, 3, …, k) represents the regression coefficient, while ε denotes the error component.

NR is a form of regression analysis where observed data is modelled by a function that is a nonlinear combination of the model parameters and depends on one or more independent variables. NR is particularly useful in environmental applications, as the relationships between variables are often not linear. This type of regression allows for more accurate modeling and prediction of complex processes [29]. Nonlinear models can be generally expressed in the form (Eq. 2):

$ Y=f(x, \theta)+\varepsilon $

(2)

where Y represents the response variable, while f denotes the function employed. The x indicates inputs, θ symbolizes the parameters that require estimation, and ε refers to the error component. If the second derivative of the function with respect to any parameter is non-zero, then that parameter is categorized as a nonlinear parameter [30]. Based on this principle, numerous models with superior performance for addressing nonlinear problems have been developed, such as SVM, ANN, RF, and XGBoost, as shown in Fig. 2.

As a general ML algorithm, the SVM idea was originally proposed by Vapnik and Chervonenkis in 1963 [31]. With constant improvement, SVM have emerged as a standard algorithm for ML, pattern recognition, and data fitting, and they are now applied to a wide range of real-world issues. SVM is based on finding a dividing hyperplane in the sample space based on the training set, thus differentiating different samples. The hyperplane in the sample space and the distance from any point to the hyperplane are denoted as Eqs. 3 and 4, respectively [4].

$ \omega^{\mathrm{T}} x+\mathrm{b}=0 $

(3)

$ r=\frac{\left|\omega^{\mathrm{T}} x+\mathrm{b}\right|}{\|\omega\|} $

(4)

where ω is the normal vector, which determines the direction of the hyperplane and b is the displacement term, which determines the distance between the hyperplane and the origin. If a hyperplane can correctly classify the training set, then the margin of the support vectors can be represented as Eq. 5. To find the hyperplane with the maximum margin, it is necessary to maximize Eq. 5. Consequently, this forms the fundamental expression of SVM, as represented by Eq. 6, which is the minimization of 1/γ. In practical scenarios, the original dataset may not always be linear. Therefore, when there is no hyperplane that can correctly classify the dataset in the original space, kernel functions are employed to map the samples from the original space to a higher-dimensional feature space. Thus, the equation is further derived as Eq. 7 [4].

$ Y=\frac{1}{\|\omega\|} $

(5)

$ \min\limits _\text{w, b} \frac{1}{2}\|\omega\|^2 $

(6)

$ \min\limits _{\mathrm{w}, \mathrm{b}} \frac{1}{2}\|\omega\|^2+\mathrm{C} \sum\limits_{\mathrm{i}=1}^{\mathrm{m}} l_{\varepsilon}\left(f\left(x_{\mathrm{i}}-y_{\mathrm{i}}\right)\right) $

(7)

where ω represents the regularization function, and l_ε is the epsilon-insensitive loss function.

In summary, the SVM model employs a learning strategy based on maximizing the margin to map original samples into a higher-dimensional space, thereby finding the optimal hyperplane to solve nonlinear scientific problems.

XGBoost represents an advanced implementation of gradient boosting algorithms, popular for its excellent performance in various predictive modelling tasks [32,33]. Designed for efficiency, flexibility, and portability, XGBoost operates by sequentially constructing a series of decision trees, each correcting the errors of the previous one, thereby enhancing the predictive accuracy of the model. The algorithm includes regularization parameters, which help prevent overfitting and make it robust to noise in the data [34]. The XGBoost model is defined by a regularized learning objective that combines a loss function with a term that penalizes complexity, as follows (Eq. 8):

$ \mathrm{Z}\left(x_{\mathrm{t}}\right)=\sum\limits_{i=1}^n \mathrm{l}\left(y_{\mathrm{i}}, \hat{y}_{\mathrm{i}}^{\mathrm{t}}-1+f_{\mathrm{t}}\left(x_{\mathrm{i}}\right)\right)+\Omega\left(f_{\mathrm{t}}\right) $

(8)

where, $\widehat{y_i}$ is the prediction for the i-th instance, $y_i$ is the corresponding target, l represents a differentiable convex loss function that measures the discrepancy between the prediction $\widehat{y_i}$ and target $y_i$, $\hat{y}_i^{t-1}$ is the prediction from t - 1 tree, and Ω presents the regression tree function used for regularization of the model. The inclusion of additional regularization terms helps to smooth the learned weights, thereby avoiding overfitting. The regularization term Ω(f_t) is defined as follows (Eq. 9):

$ \mathit{\Omega}\left(f_t\right)=\gamma T+\frac{1}{2} \lambda\|\omega\|^2 $

(9)

where, γ is a parameter that controls the complexity of the tree, T is the number of leaves in the tree, λ is a parameter for regularization on the leaf weights, and ω are the leaf weights.

In XGBoost, the model is trained additively, where each tree f_t is added to minimize the loss function l(t). The process involves using first and second-order gradient statistics on the loss function to find the best tree structure at each iteration. The optimization of the objective function is performed using a second-order approximation, which allows for a quick and effective optimization. When an instance lacks the feature required for splitting in XGBoost, the model classifies it in a default direction, which is determined during the learning process. This approach enables the model to effectively handle missing values and enhances its robustness.

ANNs have emerged as a powerful tool in the field of electrochemical oxidation. Their ability to learn complex patterns and predict outcomes makes them ideal for optimizing and understanding various electrochemical processes. ANNs are computational models inspired by the human brain, consisting of interconnected nodes or neurons that process information [35,36]. ANNs typically consist of an input layer, one or more hidden layers, and an output layer. Each layer contains a number of neurons, and the neurons are interconnected with weights that are adjusted during the training phase. The basic formula for the output of a neuron is given by (Eq. 10):

$ y=f\left(\sum\left(w_{\mathrm{i}} \cdot x_{\mathrm{i}}\right)+\mathrm{b}\right) $

(10)

where x_i and y are input and output, w_i is weight, b is bias term and f is activation function. ANNs are trained using a dataset that contains input-output pairs. During training, the network adjusts its weights and biases to minimize the error between the predicted and actual outputs. The backpropagation algorithm is commonly used for training, which involves propagating the error back through the network to update the weights [37].

RF is a bagging ensemble learning method developed from decision tree algorithms. In this technology, each tree in the forest is trained on a random subset of data and replaced, increasing the diversity between trees and improving the accuracy and robustness of the model. This ensemble learning method operates by constructing multiple decision trees during the training period and outputs the mode of the classes (for classification) or the mean prediction (for regression) of the individual trees. RF demonstrates convergence properties to achieve a lower generalization error, as expressed in Eqs. 11–13 [38].

$ m g=(\boldsymbol{X}, \boldsymbol{Y})=\alpha v_k I\left(h_k(\boldsymbol{X})=\boldsymbol{Y}\right)-\frac{\max }{j \neq Y} \mathrm{a} v_k\left(h_k(\boldsymbol{X})=j\right) $

(11)

$ P E^*=P_{X, Y}(m g(\boldsymbol{X}, \boldsymbol{Y})<0) $

(12)

$ \left.P_{X, Y}\left(P_{\Theta}(h(\boldsymbol{X}, \varTheta)=\boldsymbol{Y})-\frac{\max }{j \neq Y} \mathrm{~h}(\boldsymbol{X}, \varTheta)=j\right)<0\right) $

(13)

Eq. 11 defines the margin function for RF. It considers a set of classifiers h₁(x), h₂(x), …, h_k(x), trained on a dataset randomly drawn from the distribution of random vectors X and Y. I(·) represents an indicator function. The margin "mg" measures the extent to which the average vote for the correct class in X, Y surpasses the average vote for any other class. A larger margin indicates a higher confidence in classification. The generalization error is given by Eq. 12, which represents the probability of a random input X being misclassified in a RF (PE* < 0). Ideally, this probability should be as low as possible. In RF, $h_k(X)=h\left(\boldsymbol{X}, \varTheta_k\right)$, where for a large number of trees, as the quantity of trees increases, nearly all sequences $\varTheta_1, \ldots, P E^*$ converge to Eq. 13. This implies that with an increasing number of trees, the generalization error rate PE* of the RF will converge to a fixed value, leading to more accurate model predictions. The flowchart of the application of these five common ML algorithms is shown in Fig. S6 (Supporting information).

4. Application of electrochemical oxidation combined with machine learning

The integration of ML algorithms in the field of electrochemical oxidation is increasingly receiving attention. In line with the previously mentioned bibliometric analysis, a comprehensive review of the application mechanisms of common ML algorithms in areas such as pollutant treatment, battery restoration, electrochemical synthesis, and material characterization analysis has been conducted. These algorithms play a crucial role in enhancing the understanding of electrochemical processes and improving the overall efficiency and effectiveness of electrochemical applications.

4.1. Electrochemical treatment of pollutants

ML has been widely used in the field of electrochemical pollutant treatment. These algorithms have shown promising results in a variety of applications, including predicting pollutant concentrations, optimizing treatment processes and analyzing the complex interactions between various factors in electrochemical systems [18].

Sun et al. evaluated the performance of models trained with pollutant characteristics and reaction conditions as input features, using five common ML algorithms and 10-fold cross-validation [18]. The results showed that the XGBoost algorithm provided the best fit, evidenced by its superior R² values, while ANN model exhibited overfitting, as illustrated in Fig. 3a. Comparative analysis of four models' predictive performance further confirmed XGBoost's superiority (Fig. 3b). Subsequently, the PSO algorithm was employed to iteratively adjust the positions and velocities of ions to find optimal reaction conditions, represented as a swarm of particles each indicating a combination of reaction conditions, with initial positions and velocities denoting the values of reaction conditions and their movement speed in the search space, respectively (Figs. 3c and d) [18]. The trained XGBoost model was used to predict reaction rates for each particle, with these predictions serving as the return values of the fitness function. The fitness of each particle was calculated, recording the best positions found by each (pBest) and the global best position among all particles (gBest). Velocities and positions were adjusted based on pBest and gBest to find the optimal solution, with the process continuing until reaching a maximum number of iterations or identifying the optimal conditions, thereby enhancing the efficiency and accuracy of identifying reaction conditions. As shown in Fig. 3e, the relative error between the simulated value and the experimental value is 4.03%, and the model fitting performance is good [18].

Yu et al. used BP-ANN-P, with four inputs (initial NOR concentration, initial pH value, current density, and experiment time) and total organic carbon removal rate as the sole output, to demonstrate overall prediction accuracy of the model. Compared to first-order kinetic fitting results of conventional models, the BP-ANN-P achieved higher R² values in most cases, improving fitting accuracy by 1.619 to 127.137 times, as shown in Figs. 3f–h. Despite a decrease in the ANN model's R² with an initial NOR concentration of 200 mg/L (Fig. 3f), indicating that the effective sample size (86 samples) affects model accuracy, the R² value of the BP-ANN-P (0.969) remained higher than that of first-order kinetics (0.922) through the simulation of other five random dates. Meanwhile, the same conclusion (higher R² value for BP-ANN-P) was found on the comparison results of current density and initial pH factor, respectively (Figs. 3g and h) [39]. It shows the superiority of BP-ANN-P in modeling degradation performance.

Similarly, Rumky et al. assessed the relationship between different anode characteristics such as anode material and surface area, and the removal of chemical oxygen demand (COD), dissolved organic carbon (DOC), and color in various wastewater treatment plants. Based on a range of process characteristics considered, including electrode spacing, system pH, reactor volume, current density, and voltage, and combined with MLR modeling, it was determined that in electro-oxidation, the removal of both COD and color is dependent on the system's reaction time, while the removal of DOC is closely related to reactor volume. This MLR algorithm aids in identifying the most significant factors and optimizing the treatment process for improved performance [40]. Foroughi et al. used a three-dimension electrochemical system for the treatment of tetracycline-containing wastewater. The model of least squares SVM predicted that about 90.42% ± 2.3% of tetracycline was removed under optimal conditions (tetracycline concentration of 84 mg/L, pH 4.8, and current density of 15.72 mA/cm²), which was close to the predicted value [41]. This means that many ML algorithms have been fully applied in the field of environmental protection, especially in the treatment of electrochemical pollutants.

The integration of ML in electrochemical processes introduces innovative solutions for pollutant detection, removal optimization, and predictive performance in water treatment technologies. Through the application of ML algorithms, researchers are better equipped to predict the rates of pollutants removal in electrochemical oxidation, facilitating the development of efficient electrodes for pollutant detection and removal.

4.2. Rehabilitation of lithium-ion battery

Lithium-ion batteries have become an indispensable component in laptops and smartphones, playing a critical role in stabilizing grids powered by renewable sources such as solar and wind energy [42]. As market demand surges, so do the requirements for lithium-ion batteries. However, one challenge hindering the rapid advancement of battery technology is the time-consuming process of testing and monitoring battery health, which impacts battery lifespan [43]. Consequently, there is an imperative need to devise improved methods for predicting battery life. The application of ML in predicting the lifespan of lithium-ion batteries has gradually entered researchers' purview, including state of health prediction, state of charge (SOC) estimation, and heat generation rate (HGR) prediction. These algorithms offer a pathway to optimize their usage, extend their longevity, and contribute to a more sustainable energy future.

Thelen and colleagues accurately estimated the likelihood of lithium-ion battery capacity and the state of three main degradation modes by training ML models using limited early-life experimental data obtained through cyclic testing and simulated data from a half-cell model [20]. As shown in Fig. 4a, the hierarchical prediction of battery degradation allows for the assessment of capacity and power fade (Level 1) and the prediction of three degradation modes (Level 2) caused by various adverse chemical and physical processes (Level 3) such as graphite exfoliation, electrolyte decomposition, electrode particle cracking, as well as stress factors accelerating degradation modes such as operational time, temperature, current load, mechanical stress (Level 4). Specifically, the loss of lithium inventory (LLI) degradation mode leads to battery capacity fade, including lithium plating and solid electrolyte interface (SEI) growth consuming lithium ions, reducing the number of particles available for charge transfer. Loss of active positive material (LAM_PE), such as graphite exfoliation, resulting in the loss of lithium-ion intercalation active sites, thus reducing battery capacity. Loss of active negative material (LAM_NE) such as electrode particle cracking and loss of electric contact, creates small regions of dead active material on the electrodes, which are no longer available for lithium insertion. The loss of active material on both the anode and cathode leads to battery capacity and power fade [20].

The ML model is also capable of identifying subtle variations in VQ and dV/dQ curves and associating them with specific degradation behaviors, further enhanced by fitting simulated experimental data into ML models to better understand degradation mechanisms. As depicted in Fig. 4b and Fig. S7 (Supporting information), the experimental and simulated curves of batteries groups G1, G2, G3, and G4 for C1 (C/24, 37 ℃) V/Q and dV/dQ evolve over time. The results from the experiments and simulations for Group G1 batteries are very similar, suggesting minimal degradation and an excellent fit with the half-cell model, which is predominantly impacted by LLI. In contrast, Group G2 batteries show deviations at higher degradation levels, likely due to loss of active material (LAM), especially from the cathode. For Groups G3 and G4, significant discrepancies between dV/dQ curves and simulations suggest that LLI and LAM lead to faster degradation, with loss of active material being more pronounced at higher discharge rates (C/3), indicating that mechanical stress and higher operational loads exacerbate material loss.

Similar to this, the prediction of the HGR of batteries plays a crucial role in battery life and degradation rate. Efficient thermal management can prolong battery life by mitigating adverse thermal effects and slowing down the degradation process. Cao et al. utilized three algorithms (ANN, SVM, and GPR) to predict HGR, with results indicating that ANN exhibited the best performance in predicting HGR, with an R² value ranging from 0.89 to 1.00. This highlights its effectiveness in capturing the complex nonlinear relationships in battery HGR data [44]. Moreover, Jafari et al. leveraged the nonlinear relationships between voltage, current, and SOC, employing a data-driven approach that does not require initial SOC information, to accurately and in real-time predict the SOC of lithium-ion batteries using the XGBoost algorithm. The model predicted a root mean square error (RMSE) of 2.56, demonstrating that the XGBoost algorithm is a powerful and effective tool for estimating battery SOC, with significant advantages in prediction accuracy, operational efficiency, and application versatility [45].

In summary, current research highlights the effectiveness of ML in enhancing battery diagnostics, extending service life, and optimizing performance through predictive maintenance. By leveraging the predictive capabilities of ML, it is possible to break through the boundaries of battery technology, developing more durable, efficient, and sustainable energy storage solutions.

4.3. Generation and synthesis of substances

In the emerging field of materials science, especially in material synthesis, the process of empirical synthesis optimization is a daunting task, and the integration of artificial intelligence is an inevitable trend. ML algorithms offer unprecedented capabilities in predicting and optimizing the synthesis processes of various materials, thus facilitating the development of new materials with enhanced performance and functionalities. For example, the integration of ML algorithms helps optimize the synthesis of H₂O₂ and α-ketoiminophosphonates, enabling substance synthesis with previously unattainable properties and minimal data requirements. This marks a key transition in the field towards data-driven discovery [46,47].

Leem et al. employed the SVM algorithm to train on experimental data, including Faraday efficiency and the current density of H₂O₂ production, considering variables such as applied potential and electrolyte composition. Through a systematic approach involving 5-fold cross-validation and strict model evaluation metrics, it was determined that an optimized bicarbonate ion mole fraction of 0.225 at an applied potential of 3.25 V vs. RHE could achieve a maximum H₂O₂ current density of 2.16 mA/cm². Specifically, as shown in Figs. 5a, d and g, the model predicts total current density (J_total), faradaic efficiencies (FE_H2O2), and current densities toward H₂O₂ (J_H2O2) across a range of applied potentials and HCO₃⁻ mole fractions. To verify the model's accuracy, eight additional experimental conditions were strategically selected (marked as red dots and labeled 1 to 8) to validate the model, including untested conditions scattered throughout the feature space (points #1–6), conditions predicted to yield the highest FE_H2O2 (point #7), and the highest J_H2O2 (point #8). The performance of the SVM prediction model was evaluated using R² values for different prediction parameters, as shown in Figs. 5b, e and h, comparing measured values to predicted values for J_total, FE_H2O2, and J_H2O2. The model's predicted R² scores were high, with R² values of 0.86, 0.89, and 0.95 for J_total, FE_H2O2, and J_H2O2, respectively, indicating that the SVM model's predictions were very consistent with experimental results, emphasizing the model's accuracy. Combined with Figs. 5c, f, and i, the successful prediction of untested conditions as well as conditions with the highest predicted FE_H2O2 and J_H2O2 demonstrated the model's capability to effectively explore and optimized the parameter space for H₂O₂ production [46]. This indicates the effective application of the SVM model in electrochemical reverse engineering tasks, where SVM is capable of using discrete experimental data to construct a continuous image of electrochemical H₂O₂ production under various conditions and visualizing the effect of electrolyte composition on the two-electron water oxidation reaction, thereby enhancing H₂O₂ production.

Additionally, Kondo et al. utilize ML-assisted multiparameter screening, the optimal electrochemical oxidation conditions for α-amino phosphonates were effectively predicted, establishing an efficient synthesis method for α-ketoiminophosphonates. This ML-assisted approach employed GPR to predict across a multiparameter space, including current, reactant concentration, temperature, and reaction time. Continuous iterations of Bayesian optimization improved the accuracy of predictions, ultimately establishing a set of efficient conditions for synthesizing the desired compounds. The prediction of model consistency with experimental results demonstrated its robustness, achieving high yield and current efficiency. This process saved energy, time, and labor in reaction optimization, expanded the substrate range, and produced various α-ketoiminophosphonates with high yield and current efficiency [47].

The utilization of ML algorithms in conjunction with experimental data to provide rapid and efficient parameter screening highlights the benefits of applying ML in the domain of electrochemical oxidation synthesis. These methods have proven the ability to save energy, time, and resources by accurately predicting optimal conditions, thereby minimizing experimental trials. They have scalable and various abilities that improve the yield and comprehension of reaction pathways, enabling electrochemical oxidation synthesis in environmentally friendly conditions.

4.4. Prediction of material characterization

Common ML algorithms include NR such as decision trees, neural networks, GPR, etc. These algorithms help predict catalytic activity based on a dataset of known materials [48]. ML algorithms can significantly reduce the computational cost and time associated with density functional theory (DFT) calculations by learning from existing DFT datasets. Zhang et al. utilized ML to accelerate DFT studies on the catalytic performance of platinum-modified amorphous alloy surface catalysts. The study engineered ML potentials through the construction of distance contribution descriptors (DCD) and used ML-accelerated DFT calculations to determine the Gibbs free energies of 46,000 *H adsorption sites on the Pt@PdNiCuP surface. In the study, the optimal efficacy of DCD was optimized through SVM learning evaluation, as illustrated in Fig. 6a, where m = 3 minimized the mean squared error to 0.0695. The learning curve of the SVM model (Fig. 6b) indicated that increasing the sample size could improve the model's accuracy, achieving significant precision with approximately 500 samples. Further refinement of the model involved hyperparameter tuning through grid search for 800 sample sets, leading to a RMSE of 0.130 eV for the training set and 0.131 eV for the test set (Fig. 6c). Moreover, the relative error distribution confirmed the model's accuracy, with most prediction errors within ± 20%, showcasing the model's reliability (Fig. 6d) [49]. This approach, by effectively identifying optimal descriptors, focuses DFT computational resources on the most promising candidate materials. Ultimately, by narrowing down the search space and identifying key material descriptors that control catalytic activity. SVM has demonstrated significant advantages in enhancing the computational efficiency of DFT, especially in accelerating the computation of Gibbs free energies of adsorption sites on material surfaces. By significantly reducing computational cost and time, SVM not only accurately predicts material properties, but also facilitates the acceleration of new material discovery and performance optimization. This advancement highlights SVM's outstanding ability to predict key properties and their associated reaction mechanisms in materials science.

Similarly, Sha et al. employed a neural network algorithm to dissect the characterization of LiNi_0.5Co_0.2Mn_0.3O₂ single crystal cathode materials. This required the construction of a special convolutional neural network model capable of processing scanning transmission electron microscopy (STEM) images, identifying primary crystal structures, pinpointing defects, and categorizing them. The entire process of image analysis based on a neural network model is depicted in Fig. 6e. Specifically, the input of network is the simulated STEM images of various atomic models, and the output of the network is the location of point defects in atomic clusters. The different crystal structures are randomly mixed to form a valid dataset, which is then subjected to preprocessing steps, such as normalization of the image pixel values and conversion of labels into Boolean values. Subsequently, network parameters are iteratively adjusted using the stochastic gradient descent algorithm. Finally, the model parameters retained at the end of training are utilized for defect detection in real STEM images. With the support of ML algorithms, the analysis of LiNi_0.5Co_0.2Mn_0.3O₂ single crystal cathode materials encompassed the types, quantity, and distribution of defects and their correlation with electrochemical performance, thereby offering profound insights into the material's degradation process [50].

In summary, the incorporation of ML in electrochemical oxidation processes presents many benefits, notably predicting electrochemical oxidation performance, improving process optimization, and facilitating real-time monitoring. Nonetheless, this integration faces full of challenges. Key among these are the various computational requirements, complex parameter tuning, insufficient data sets and model interpretability. To fully exploit ML's capabilities in this domain, these challenges need to be fully considered and solved. Future research may focus on enhancing the predictive accuracy, generalizability, and interpretability of ML models through algorithmic adjustments, data distribution and structure optimization, and the development of novel attribution analysis methodologies. It is anticipated that these improved ML models will be able to analyze the complex interconnections found in electrochemical oxidation processes.

5. Current challenges and prospects

In summary, ML methods have been used in a variety of fields of electrochemical oxidation, and based on the advantages and disadvantages of different ML algorithms summarized in Table S3 (Supporting information), it can be found that there are different applicable algorithms for different electrochemical oxidation fields.

For the treatment of electrochemical pollutants, algorithms such as RF, SVM, and ANN are particularly well-suited. The low risk of overfitting, robust to outliers, and high accuracy of RF enable it to capture complex data patterns while resisting the impacts of impurities in data, such as noise and outliers, which is ideal for dealing with the variable and unpredictable data often encountered in pollutant treatment processes. SVM excels at handling high-dimensional data and boasts excellent generalization capabilities, effectively managing the nonlinear relationships present in electrochemical pollutant treatment within high-dimensional spaces. By seeking to maximize the decision boundary, SVM aids in the accurate classification of unknown pollutant data. The high flexibility and powerful learning capabilities of ANN make it adept at capturing and learning complex patterns within large datasets, suitable for modeling the intricate chemical processes involved in electrochemical pollutant treatment. Given the potential for vast amounts of experimental data in the field of electrochemical pollutant treatment, ANN's proficiency in processing large-scale datasets proves invaluable for understanding the dynamic processes of pollutant degradation.

The high flexibility of ANN allows for the adaptation of network architecture to suit problems of varying complexity. In the prediction of battery performance, this flexibility enables ANN to capture the complex nonlinear patterns occurring during the battery discharge and charging processes. Additionally, given the extensive data generated from battery testing, ANN, with its strong adaptability and powerful learning capabilities, can self-adjust to learn subtle patterns and long-term dependencies within large datasets, which is crucial for predicting battery life under changing operational conditions. XGBoost is efficient and flexible, exhibiting a significant advantage in handling missing data. In the context of battery restoration, it can rapidly analyze and model vast quantities of historical data, automatically learning from the missing values that occur for various reasons within large datasets to efficiently predict future battery performance. Furthermore, the high accuracy of XGBoost enables effective implementation to enhance prediction precision, thus facilitating timely repair and maintenance decisions. In addition, both algorithms are capable of handling large datasets, which is important in battery testing and monitoring.

The precision and predictive capabilities of algorithms such as SVM and ANN are greatly beneficial for the generation and synthesis of substances. The prediction of various chemical, physical, and structural properties of substances requires the processing of high-dimensional data. SVM is well-suited for such high-dimensional spaces, and its high accuracy and good generalization ability assist in accurately predicting the performance of unknown compounds [51,52]. The flexibility and strong adaptability of ANN make it capable of handling various types and sizes of data, which is particularly applicable in the field of synthesis science with its complex properties. Moreover, ANN with powerful learning capability can learn from a large amount of data, such as the numerous variables required in the synthesis process, and identify interaction modes, aiding in the understanding and prediction of the synthesis and performance of new substances.

The prediction of material characterization often involves extracting information about molecular structures from high-dimensional data. SVM shows its unique advantages in dealing with multiple influences and potential interactions, mainly due to its effectiveness in high-dimensional space and excellent classification ability. The kernel trick of SVM allows researchers to find the optimal boundaries, which is crucial for revealing the deep chemical behavior of materials. In addition, SVM's high accuracy and excellent generalization capability allow it to show excellent performance in the prediction of material properties where high accuracy is required. ANN, with its highly flexible network structure and excellent learning capability, simulates and learns nonlinear relationships for complex reactions and adapts to variable data patterns, which is very useful for understanding and predicting how materials react under different conditions. At the same time, ANN is able to process images, recognize and classify the internal structure of materials, and simulate the relationship between electrochemical parameters and surface reactions. RF not only maintains robustness in the presence of noise in the data, but also reduces the risk of overfitting by integrating the learning approach, thus enhancing the predictive accuracy of the model on new data. These features make RF a reliable choice for predicting complex chemical reactions. It is the ability of RF to provide accurate predictions while remaining robust to random variability in the data, which is critical for studying the uncertainty and dynamics of oxidation reaction processes.

Please note that the suitability of a particular ML algorithm can vary depending on the specific characteristics of the dataset and the problem at hand. The insights provided are based on the general capabilities of the algorithms and their common applications in related fields. Additionally, the actual implementation of these algorithms in the specified domains would require domain-specific knowledge and data to train and validate the models effectively. In conclusion, the ML algorithms need to be carefully adapted to the specific requirements and complexity of the electrochemical oxidation field. To effectively deploy ML algorithms in all areas of electrochemical oxidation, thorough data evaluation, careful algorithm tuning, rigorous validation, and continuous learning systems can be employed. Firstly, it is crucial to assess the nature, quantity and quality of the data, as algorithms that perform well on large datasets are algorithms such as ANN and XGBoost, whereas less high-dimensional data is more suitable for SVM. Secondly, it is necessary to fine-tune the parameters of each algorithm to match the specific data characteristics of each domain, such as tuning the kernel of the SVM for non-linear data in pollutant treatment, or conducting battery performance studies by modifying the network layer of the ANN. Thirdly, the model is validated using cross-validation techniques to ensure that it can accurately predict unknown data, thus validating its generalization ability. Finally, the establishment of a continuous learning mechanism enables the model to be adapted and improved as new data are encountered, which is particularly important in the field of dynamic electrochemical processes.

6. Conclusion and prospect

Selecting the appropriate ML algorithm that suits the specific requirements and complexities of the target domain is crucial. The effectiveness of these algorithms in their respective applications is expected to make significant advancements in understanding and optimizing processes within these domains. Review analysis of ML algorithms like ANN, NR, SVM, XGBoost, and RF highlights their unique advantages and the prospect of their wide applications across multiple domains, including electrochemical pollutant treatment, battery repair and forecasting, as well as material synthesis and surface mechanism prediction. In the domain of electrochemical pollutant treatment, RF, SVM, and ANN stand out for their robustness to noise, high-dimensional data handling capabilities, and modeling of nonlinear relationships, respectively. These characteristics are crucial for dealing with the typically complex data and unpredictable dynamics in pollutant degradation processes. Similarly, for battery life prediction and degradation rate forecasting, ANN and XGBoost algorithms are preferred for their adaptability in modeling complex battery behaviors and efficiency in handling missing data, critical for timely repair and maintenance decisions. Substance synthesis benefits from the precision of SVM and ANN algorithms, as well as their pattern recognition capabilities and flexibility in learning complex synthesis pathways. SVM, ANN and RF have better fitting effects on the characterization of materials. These algorithms are able to recognize and classify the internal structure of materials, learn complex data patterns, and predict unknown data more accurately. This facilitates a correct understanding of complex response mechanisms. Recent high-quality research also indicates that the integration of these algorithms, especially the RF-XGBoost-ANN combination, can provide superior predictive accuracy and minimal error, demonstrating the potential for wide application [53].

The future of ML in electrochemical oxidation hinges on developing more specialized algorithms that balance computational efficiency with high interpretability. Hybrid models, combining the strengths of different algorithms, and advancements in deep learning, particularly in enhancing interpretability and reducing training requirements, are likely to dominate future research in the following ways:

Model interpretability: Developing new tools and methods to make the decision-making process of complex models more transparent and easier to understand. Developing and promoting standards and assessment metrics for interpretability to help researchers better assess the transparency of models to address "black box" issues.

Autonomous real-time data processing and feedback: There is a need to develop ML frameworks that can autonomously receive and process experimental data in real time and feed this information back to the model for adjustment. The close integration of this model with experimental feedback enables the system to adjust and improve predictions in real time based on experimental results, improving the adaptability and response speed of the electrochemical oxidation process.

Data standardization and model integration: Optimizing data collection and developing new data pre-processing methods. Anticipated data distributions are predicted and recorded to forecast potential biases in the data. Diversification of data sources and quality checks are conducted to ensure the integrity of the data. Rigorous data preprocessing methods are implemented to eliminate bias, thereby enhancing the validity of the models. Finally, with a comprehensive understanding of the dataset and an exploration of model suitability, a more powerful AutoML framework is developed to automate model selection and hyperparameter optimization, which will effectively determine the best ML model for a given requirement.

Energy-efficient learning models: Developing sample less learning and transfer learning algorithms to reduce the dependence of model on large-scale data. Also developing energy-efficient algorithms and hardware to speed up model training and inference and reduce energy consumption.

Declaration of competing interest

The authors declare no competing financial interests that could have appeared to influence the work reported in this paper.

CRediT authorship contribution statement

Zonglin Li: Writing – review & editing, Writing – original draft, Resources, Methodology, Investigation, Conceptualization. Shihua Zou: Visualization, Software, Formal analysis. Zining Wang: Methodology, Investigation. Georgeta Postole: Validation, Supervision. Liang Hu: Validation, Supervision. Hongying Zhao: Supervision, Project administration, Funding acquisition, Conceptualization.

Acknowledgments

The authors acknowledge funding from the National Natural Science Foundation of China (Nos. 22122606, 22076142, 62276190), National Key Basic Research Program of China (No. 2017YFA0403402), National Natural Science Foundation of China (No. U1932119), the Science & Technology Commission of Shanghai Municipality (No. 14DZ2261100), the Fundamental Research Funds for the Central Universities.

Supplementary materials

Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.cclet.2024.110526.