Intelligent Pulsed Electrochemical Activation of NaClO₂ for Sulfamethoxazole Removal from Wastewater Driven by Machine Learning

Abstract

Sulfamethoxazole (SMX), a widely used antibiotic, poses potential threats to ecosystems and human health due to its persistence and residues in aquatic environments. This study developed a novel intelligent water treatment system, namely Intelligent Pulsed Electrochemical Activation of NaClO₂ (IPEA_NaClO2), which integrates a FeCuC-Ti₄O₇ composite electrode with machine learning (ML) to achieve efficient SMX removal and energy consumption optimization. Six key operational parameters—initial SMX concentration, NaClO₂ dosage, reaction temperature, reaction time, pulsed potential, and pulsed frequency—were systematically investigated to evaluate their effects on removal efficiency and electrical specific energy consumption (E-SEC). Under optimized conditions (SMX 10 mg L⁻¹, NaClO₂ 60~90 mM, pulsed frequency 10 Hz, temperature 313 K) for 60 min, the IPEA_NaClO2 system achieved an SMX removal efficiency of 89.9% with a low E-SEC of 0.66 kWh m⁻³. Among the ML models compared (back-propagation neural network, BPNN; gradient boosting decision tree, GBDT; random forest, RF), BPNN exhibited the best predictive performance for both SMX removal efficiency and E-SEC, with a coefficient of determination (R²) approaching 1 on the test set. Practical application tests demonstrated that the system maintained excellent stability across different water matrices, achieved a bacterial inactivation rate of 98.99%, and significantly reduced SMX residues in a simulated agricultural irrigation system. This study provides a novel strategy for the intelligent control and efficient removal of refractory organic pollutants in complex water bodies.

1. Introduction

SMX, a broad-spectrum antibiotic widely used for anti-infective therapy, has become an emerging environmental contaminant due to its persistence and bioaccumulation [1,2,3,4]. Its stable molecular structure and low biodegradability enable long-term retention in aquatic environments. Globally, about 11% of farmland is irrigated with municipal wastewater, exposing agroecosystems to antibiotics such as SMX [5]. SMX has been reported in irrigation water, reclaimed water, and contaminated surface waters at concentrations of tens to hundreds of ng L⁻¹, reaching several thousand ng L⁻¹ in heavily polluted sites [6]. Soil concentrations ranged from 2.39 to 13.4 µg kg⁻¹, while crop concentrations were reported in the ng g⁻¹ range [7]. Persistent SMX promotes the dissemination of antibiotic-resistant bacteria and genes and alters microbial community structures [8,9,10]. Its bioaccumulation through food chains also poses potential endocrine-disrupting risks to humans. These findings highlight the urgent necessity of developing effective treatment technologies for SMX removal, making it a major focus in current environmental remediation research [11,12,13].

Conventional treatments—e.g., adsorption and biological methods—often perform poorly against persistent contaminants such as SMX [14,15]. Adsorption is attractive because of its operational simplicity and rapid uptake, but it does not destroy the contaminant; moreover, adsorbent regeneration often incurs high energy costs, incomplete desorption, and potential secondary pollution from spent sorbents [16,17]. Biological treatments rely on microbial degradation, yet the introduction of antibiotics into biologically active systems can destabilize microbial consortia, reduce treatment efficiency, and create selective pressure that favors resistant strains. Even if low concentrations of antibiotic stress can be tolerated, the risk of horizontal transfer of resistance genes remains a serious problem and may exacerbate the spread of resistance [18,19]. These limitations have driven interest in advanced oxidation processes (AOPs) [20] that are capable of mineralizing or transforming recalcitrant contaminants while minimizing secondary pollution and energy consumption.

Among AOPs, electrocatalytic technology exhibits distinct advantages by enabling both direct electron-transfer oxidation and precursor oxidant activation through the regulation of electrode potential and interfacial charge distribution, thereby facilitating the removal of refractory organic pollutants [21,22]. Electrocatalytic systems are typically classified into direct and indirect oxidation modes. Direct electro-oxidation removes pollutants through electron transfer at the electrode surface and usually has a high removal efficiency for electroactive substances [23,24,25], but often less so for non-electroactive compounds [26]. Indirect oxidation, by contrast, relies on electrochemically activated oxidants (such as persulfate or peracetic acid) or the electrolysis of water to produce reactive intermediates in the vicinity of the electrode; these intermediates then attack target contaminants [27,28]. NaClO₂ has attracted attention as a precursor for selective oxidants, including chlorine dioxide (ClO₂), during electroactivation [29]. Its relatively low standard reduction potential (0.954 V vs. SHE) [30] allows it to participate in oxidative attack on aromatic moieties (such as the SMX benzene ring) at comparatively low overpotentials, thereby enabling efficient degradation under milder electrochemical conditions [31]. Nevertheless, conventional electrochemical oxidation frequently suffers from competition with the oxygen evolution reaction (OER), which consumes current, lowers energy efficiency, and can destabilize the system [32]; balancing active species generation against OER suppression is therefore central to the design of effective electrochemical treatment systems.

Pulsed electrochemical operation has emerged as a promising strategy to address this tradeoff [33,34,35]. By periodically modulating the electrode potential, pulsed operation rapidly toggles the electrode surface between oxidizing and reducing states, altering interfacial charge distributions and transient adsorption/desorption kinetics so as to favor active species formation while mitigating parasitic OER. Existing studies have demonstrated that pulsed potential operation, in comparison with constant potential mode, can effectively reduce energy consumption—with reductions reaching up to 50%—and concurrently enhance electrode stability [36,37]. However, pulsed electrochemical systems are intrinsically complex: pulse waveform, frequency, voltage, solution pH, electrolyte composition, ionic strength, and the presence of coexisting ions all interact in nonlinear ways to control radical generation and contaminant degradation kinetics [38].

Previous studies have shown that the combination of operating parameters strongly influences both radical generation efficiency and pollutant degradation rates [39,40]. In pulsed electrocatalytic degradation, parameters such as frequency and duty cycle regulate the production of interfacial reactive species (·OH, ·O₂⁻) and alter electrode microstructures (e.g., Pt grain size and dispersion), thereby determining whether direct electron transfer or radical attack dominates the degradation process [39]. In pulsed electrocoagulation, significant interactions occur among electrode type, polarity-reversal frequency, and current density: polarity reversal mitigates Al electrode passivation and enhances Faradaic efficiency, whereas excessive reversal frequency in Fe electrodes hampers ion migration and promotes side reactions (e.g., oxidation of adsorbed hydrogen), reducing efficiency to below 10% [40]. These results indicate that traditional linear or single-factor regression models cannot accurately describe the multi-parameter coupling behavior inherent to pulsed electrochemical degradation systems [41].

Given these challenges, machine learning (ML) methods capable of identifying and exploiting nonlinear relationships represent an attractive toolset for modeling and optimizing pulsed electrochemical degradation processes [42]. Among ML models, artificial neural networks (ANNs) are particularly well suited for multivariate, nonlinear systems and have been widely applied to water-quality prediction and process control in wastewater treatment [43,44]. BPNN, a class of multilayer feedforward ANNs trained by gradient-based error propagation, can learn complex input–output mappings without prespecifying a mechanistic mathematical model; this flexibility renders BPNN advantageous over traditional regression when dealing with noisy, high-dimensional experimental data [45]. Successful applications of BPNN and other ML algorithms to electrochemical water treatment include predictive modeling of COD removal in Fenton processes, optimization of electrochemical sludge oxidation, and adsorption kinetics modeling, where these approaches have achieved high accuracy and robustness [46,47,48].

Building on this background, the present study hypothesizes that systematic optimization of key operational parameters in Intelligent pulsed electrochemical activation via ML approaches can achieve highly efficient removal of SMX coupled with low-energy consumption. To test this hypothesis (Figure 1), a FeCuC composite catalyst was synthesized via a hydrothermal method in this study. After briefly comparing the removal efficiencies of conventional electrodes (Ti and Pt) with that of a Ti₄O₇ electrode, the catalyst was immobilized on a Ti₄O₇ electrode substrate (Figure S8) to fabricate a highly active composite anode [49]. A series of pulsed electrochemical experiments—centered on NaClO₂ electroactivation—were conducted to evaluate SMX removal as a function of key process variables, including initial SMX concentration, initial NaClO₂ concentration, reaction time, temperature, pulse frequency, and pulse potential. Single-factor tests were complemented by trials using real water matrices to assess applicability, including comparative measurements of SMX residues in treated water, irrigated soils, and crops, as well as disinfection performance. Experimental data were used to train and compare multiple ML models (BPNN; GBDT; RF), and BPNN was selected as the optimal predictor. Finally, a software prediction platform was developed to enable rapid estimation of SMX removal performance under varied water-quality and operational scenarios. If successful, this integrated approach offers a practical pathway to improve the removal of SMX and related recalcitrant organics and to advance intelligent, energy-efficient strategies for water treatment.

2. Materials and Methods

2.1. Chemicals

All chemicals were of analytical grade and used as received without further purification. Copper(II) nitrate trihydrate (Cu(NO₃)₂·3H₂O, purity ≥ 98.0%) was purchased from Macklin Biochemical Technology Co., Ltd. (Shanghai, China). Iron(III) nitrate nonahydrate (Fe(NO₃)₃·9H₂O, purity ≥ 98.5%) and anhydrous sodium sulfate (Na₂SO₄) were obtained from Kelong Chemical Reagent Co., Ltd. (Chengdu, China). Sodium chlorite (NaClO₂, purity ≥ 80.0%) was supplied by Yien Chemical Technology Co., Ltd. (Shanghai, China). Sulfamethoxazole (C₁₀H₁₁N₃O₃S, purity ≥ 98.0%) and sodium thiosulfate (Na₂S₂O₃, 99%) were provided by InnoChem Technology Co., Ltd. (Beijing, China) Tetratitanium heptoxide (Ti₄O₇) was purchased from a commercial supplier. Escherichia coli. (E. coli.) ATCC 25922 (non-receptor strain) was obtained from Sangon Biotech Co., Ltd. (Shanghai, China) for biological assays. Ultrapure water (Milli-Q, 18.2 MΩ cm) was used for all chemical experiments, while triply distilled water was used for biological experiments.

2.2. Synthesis and Characterization of Electrode Materials

2.2.1. Synthesis of FeCuC Composite

Stoichiometric amounts of Fe(NO₃)₃·9H₂O and Cu(NO₃)₂·3H₂O were dissolved in deionized water under magnetic stirring to form a homogeneous solution. Bamboo powder was selected as the carbon precursor owing to its abundant availability and low cost. As a fast-growing lignocellulosic material, bamboo is rich in cellulose, hemicellulose, and lignin, which, upon pyrolysis, yield a carbon-rich char with high carbon yield and readily tunable pore structure. The powder was then introduced in 3~5 batches under continuous stirring for 2 h to achieve uniform dispersion. The resulting suspension was transferred into a 100 mL PTFE-lined stainless-steel autoclave, preheated at 393 K for 2 h, then heated to 473 K at 5 K min⁻¹ and maintained for 12 h under hydrothermal conditions. After cooling to room temperature, the product was collected, washed repeatedly with deionized water and ethanol until the filtrate reached neutral pH, and dried at 343 K overnight.

2.2.2. Synthesis of FeCuC-Ti₄O₇ Composite Electrode

The previously prepared FeCuC powder and Ti₄O₇ were placed together in a new 100 mL PTFE-lined autoclave, followed by the addition of deionized water as the reaction medium. The mixture was sealed and subjected to a second hydrothermal process following the same temperature profile as above (393 K for 2 h, ramp 5 K min⁻¹ to 473 K, then held 12 h). The autoclave was cooled naturally to room temperature, and the product was washed alternately with deionized water and ethanol until neutral, then dried at 343 K overnight to obtain the FeCuC-Ti₄O₇ composite electrode.

2.2.3. Characterization

The crystalline structure of the composite was examined by X-ray diffraction (XRD, D8 Advance, Bruker, Karlsruhe, Germany). Surface morphology was observed using scanning electron microscopy (SEM, Sigma 360, ZEISS, Oberkochen, Germany), and elemental chemical states of Fe, Cu, C and O were analyzed via X-ray photoelectron spectroscopy (XPS, K-Alpha, Thermo Fisher Scientific, Waltham, MA, USA).

2.3. Electrochemical Removal Experiments

Unless otherwise specified, all experiments were performed at 298 ± 1 K. Sodium chlorite was accurately weighed according to the target concentration and dissolved in 100 mL of SMX solution (10 mg L⁻¹) prepared in a 250 mL glass beaker. A 90 mL portion of the solution was transferred into a 100 mL electrochemical cell equipped with a three-electrode configuration: FeCuC-Ti₄O₇ as the working electrode, a glassy carbon counter electrode, and a saturated calomel electrode (SCE) as the reference. The pulsed electrochemical process was initiated using an electrochemical workstation by applying a square wave potential. The waveform was generated by a programmable pulsed DC power supply, with the high level set at the pulse potential E, a duty cycle of 1:1, and the low level held at 0 V (i.e., no applied potential). At predefined time intervals, 1.0 mL samples were collected through 0.45 μm filters and immediately quenched with 0.4 mL of 200 mM Na₂S₂O₃ solution to neutralize residual oxidants. SMX concentrations were determined via HPLC (Agilent 1260, Santa Clara, CA, USA) on a Waters BEH C18 column (2.1 × 100 mm², 1.7 μm). Mobile phase A was 0.1% formic acid aqueous solution and mobile phase B was acetonitrile, with a flow rate of 0.3 mL/min. Mass spectra were recorded on an Agilent QTOF-6550 (USA) for exact mass confirmation. The chromatographic peak at ~7.05 min was integrated to calculate SMX removal efficiency.

The effects of six operational parameters were systematically investigated: initial SMX concentration (5~45 mg L⁻¹), NaClO₂ concentration (0~90 mM), reaction temperature (287~313 K), reaction time (30~90 min), pulsed potential (1~5 V), and pulsed frequency (0.01~1000 Hz). Each parameter was varied independently while others were kept constant. All removal and energy-consumption experiments were performed in triplicate (n = 3), yielding a total of 255 experimental runs for comprehensive evaluation of process efficiency and energy consumption.

2.4. Electrochemical Test

Electrochemical performance of the electrodes was evaluated using a CHI 760E electrochemical workstation (CH Instruments, Shanghai, China) in 0.1 M Na₂SO₄ electrolyte. Cyclic voltammetry (CV), linear sweep voltammetry (LSV), and electrochemical impedance spectroscopy (EIS) were performed under controlled conditions.

CV scans were recorded at scan rates of 0.01~0.1 mV s⁻¹ (increments of 0.01 mV s⁻¹) to estimate the electrochemically active surface area (ECSA). LSV experiments examined (1) the effect of NaClO₂ concentration (0, 1, 5, 10, 25, and 50 mM) on redox behavior at 0.1 mV s⁻¹; (2) various scan rates 0.7~0.9 mV s⁻¹ in steps of 0.05 mV s⁻¹ to estimate the number of electron transfers involved in the electrochemical process; and (3) temperature-dependent kinetics between 298 and 318 K with 5 K intervals. EIS measurements were acquired over 0.01~1000 Hz for electrodes immersed in 0.1 M Na₂SO₄ and in a mixed electrolyte of 50 mM NaClO₂ + 0.1 M Na₂SO₄ to investigate interfacial charge-transfer resistance.

2.5. Neural Network Model Development

A total of 255 experimental datasets were used to construct machine-learning models. The data were split into a training set (205 samples) and a test set (50 samples). Pulsed-frequency data were log₁₀-transformed, and all variables were normalized using min–max scaling (see Equation (1)).

$$\stackrel{\mathrm{-}}{\mathrm{z}}={\displaystyle \frac{\mathrm{z}-{\mathrm{z}}_{\min }}{{\mathrm{z}}_{\max }-{\mathrm{z}}_{\min }}}$$

(1)

Six operational parameters served as input variables, while the final SMX concentration and E-SEC were used as output targets.

Two BPNN models with a 6-n-1 architecture were constructed for predicting SMX concentration and E-SEC, respectively. The number of hidden neurons was determined empirically (Equation (2)) and optimized to 7 and 12 for the two models.

$$\mathrm{m}=\sqrt{\mathrm{p}+\mathrm{q}}+\mathrm{a}$$

(2)

Networks were trained using the Levenberg–Marquardt algorithm in MATLAB (R2023a) Neural Network Toolbox.

A GBDT model was also developed using the fitrensemble function, employing least-squares boosting with a minimum leaf size of 5 to mitigate overfitting. In addition, a RF model was constructed. The RF model was implemented via MATLAB’s TreeBagger function. Model hyperparameters were optimized through grid search. Model performance was assessed by mean squared error (MSE), root mean squared error (RMSE), and coefficient of determination (R²).

2.6. Practical Application

Four application tests were carried out to evaluate the practical feasibility of the developed technology (All experiments were performed in triplicate). The stability of the FeCuC-Ti₄O₇ electrode was examined in various water matrices containing common coexisting species (CO₃²⁻, Cl⁻, and humic acid), and SMX removal efficiency was quantified under each condition. Meanwhile, disinfection performance for reclaimed irrigation water was tested using E. coli. as the target microorganism under four conditions—raw water, electric field alone, NaClO₂ alone, and the synergistic electrocatalytic system. Bacterial concentrations were determined via the plate counting method. Furthermore, lettuce was cultivated using irrigation water treated before and after the IPEA_NaClO2 system process to assess SMX accumulation in plant tissues and soil. Finally, the predictive accuracy of the BPNN, DT, and GBDT models was validated by comparing predicted and experimental results under fixed operating parameters, thereby comprehensively verifying the system’s performance and reliability for practical wastewater treatment applications.

3. Results and Discussion

3.1. Construction of the Pulsed Electrochemical System

To establish the pulsed electrochemical system, the synthesized FeCuC-Ti₄O₇ working electrode was first characterized. As shown in Figure 2a, the electrode exhibits a geometric diameter of approximately 3 cm and a thickness of about 2 mm, with a metallic luster and black surface. The SEM image (Figure 2b) reveals a cellulose-like porous skeleton structure that provides a high specific surface area, facilitating electrolyte permeation and mass transfer. An enlarged image (Figure 2c) reveals that the porous framework is composed of well-dispersed nanoparticles with diameters ranging from 20 to 30 nm. The electrochemical surface area (ECSA) characterization (Figure S1) demonstrated a value as high as 5.82 mF cm⁻², providing a favorable interface for SMX removal.

XPS survey spectra (Figure S2) confirmed that the electrode surface contains C, O, Cu, and Fe elements, aligning with the designed composition. The Fe 2p spectrum (Figure 2d) exhibits a main Fe 2p_3/2 peak at 710.8 eV, characteristic of Fe(III), along with the Fe 2p_1/2 peak at 724.5 eV. A broad satellite peak near 716.5 eV, assigned to Fe(II), is also observed. These features confirm that iron is predominantly present as Fe(III) oxides, with possible trace presence of Fe(II) species. In Figure 2e, the 2p_3/2 peak (at ~932.5–933.6 eV) and 2p_1/2 peak (at ~952.3–954.1 eV) reflect the spin-orbit splitting of Cu 2p. The obvious satellite peak (at ~944 eV) confirms the presence of Cu(II), while the 2p_3/2 peak at ~932.5 eV implies the existence of Cu(I) (no satellite peak). Thus, the sample contains mixed Cu(I)/Cu(II) species [50,51]. In the C 1s spectrum (Figure 2f), the major peak at 284.8 eV corresponds to sp²-hybridized C–C bonds, while additional peaks at 286.5 eV and 288.6 eV are assigned to C–OH and O–C=O, respectively. The O 1s spectrum (Figure S3) shows a broad band at 531~534 eV, confirming the existence of oxygen-containing groups. These surface hydroxyl and carboxyl functionalities enhance the hydrophilicity of the electrode, promoting reactant transport through the porous structure and thus accelerating the electrochemical reactions.

X-ray diffraction (XRD) analysis (Figure 2g) was further conducted to determine the crystalline composition. The diffraction peaks match well with those of graphite carbon, iron carbide, and metallic copper, indicating that the composite consists of graphitic carbon, Fe–C phases, and Cu crystallites. The broad diffuse peak within 20~30° corresponds to amorphous carbon. The overlapping and slightly broadened peaks of copper and iron carbide suggest nanoscale particles highly dispersed within the carbon matrix. These results confirm that FeCuC consists of a carbon framework doped with a small amount of copper and iron, forming the CFe₁₅.₁ phase, signifying successful composite formation of Fe and Cu on the carbon substrate.

Moreover, the electrochemical analysis in Figure 2h,i reveals that the number of transferred electrons during NaClO₂ activation is approximately one, consistent with the expected redox way. This suggests a single-electron transfer activation of NaClO₂ to ClO₂ via Equation (1):

NaClO₂ − e⁻ → Na⁺ + ClO₂(aq.)

(3)

3.2. Effects of Experimental Parameters

To systematically evaluate the influence of operational parameters on SMX removal (Equation (3)) and E-SEC (Equation (4)), several key factors were analyzed. As illustrated in Figure 3a, when the initial SMX concentration increased from 5 to 45 ppm, the 60 min removal efficiency decreased from (84.53 ± 0.64)% to (64.48 ± 0.07)%. This can be attributed to the same generation of reactive species (such as ClO₂) within a fixed reaction time and initial NaClO₂. As concentration rises, fewer reactive species are available per SMX molecule, and mass transfer limitations become more significant, causing diffusion control to dominate. Nevertheless, the E-SEC decreased significantly from (1.46 ± 0.06 kWh m⁻³) to (0.25 ± 0.06 kWh m⁻³) (Figure 4a) because the total mass of SMX removed increased while total energy consumption remained nearly constant. As shown in Figure S4, extending the reaction time from 60 to 90 min slightly improved the removal efficiency (from 84.53% to 89.93%), yet the overall trend remained unchanged. Beyond 60 min, the marginal benefit diminishes, and the E-SEC correspondingly increases (0.8 ± 0.04 kWh m⁻³) to (1.03 ± 0.04 kWh m⁻³), reflecting reduced energy efficiency during the late reaction stage (Figure 4d).

$$\mathsf{\eta }\left(\mathrm{\%}\right)=(1-{\displaystyle \frac{\mathsf{\omega }1}{\mathsf{\omega }2}})\times 100\mathrm{\%}$$

(4)

$$\mathrm{E}-\mathrm{S}\mathrm{E}\mathrm{C}\left({\mathrm{k}\mathrm{W}\mathrm{h} \mathrm{m}}^{-3}\right)={\displaystyle \frac{\mathrm{V}\times \mathrm{I}\times \mathrm{t}}{{\mathrm{m}}_{\mathrm{S}\mathrm{M}\mathrm{X},\mathrm{r}\mathrm{e}\mathrm{m}\mathrm{o}\mathrm{v}\mathrm{a}\mathrm{l}}}}$$

(5)

where $\mathsf{\eta }$ represents the removal efficiency of SMX, $\mathsf{\omega }1$ the final concentration, $\mathsf{\omega }2$ the initial concentration, V the pulse potential, I the current, t the removal time, and m_SMX,removal the amount of SMX removed.

When the initial NaClO₂ concentration increased from 0 to 90 mM (Figure 3b), SMX removal efficiency first rose sharply and then reached a plateau at (82.4 ± 0.1)% around 90 mM. Below 60 mM, ClO₂ generation correlates positively with NaClO₂ concentration, whereas beyond 60 mM, excessive ClO₂ escape, side reactions (e.g., ClO₂ → ClO₃⁻), and saturation of active sites slow further enhancement. Meanwhile, E-SEC decreased continuously from (1.46 ± 0.06 kWh m⁻³) to (0.74 ± 0.05 kWh m⁻³) (Figure 4b) due to a higher pollutant removal rate per unit time.

Temperature also significantly affected removal performance (Figure 3c). As the temperature increased from 287 to 313 K, the removal efficiency improved from (72.17 ± 0.38)% to (90.03 ± 0.32)%, primarily due to enhanced mass transfer and accelerated intrinsic kinetics. However, above 313 K, thermal decomposition of NaClO₂ and electrode passivation hindered further improvement. The lowest E-SEC (0.71 ± 0.05 kWh m⁻³) was observed at 313 K (Figure 4c), as higher temperature enhanced conductivity and kinetic rate, but further heating raised energy demand, causing a rebound in E-SEC.

The influence of pulsed potential is presented in Figure 3e. As the potential increased from 1 to 1.5 V, the removal efficiency rose sharply from (32.1 ± 0.1)% to (78.3 ± 0.2)%. It is noteworthy that further raising the potential led to a pronounced decline in the removal rate (at 2 V pulse potential, the removal efficiency decreased by approximately 12% compared to that at 1.5 V), likely due to the synergistic effect between direct electron transfer and indirect oxidation by reactive chlorine species. Moderate increases in potential enhance ClO₂⁻ → ClO₂ activation and pollutant oxidation, while excessive potentials trigger parasitic reactions such as oxygen evolution (OER) and ClO₂ → ClO₃⁻, decreasing current efficiency and possibly causing electrode corrosion. As the voltage increased, the E-SEC reached its minimum value of (0.42 ± 0.05 kWh m⁻³) at 1.5 V and subsequently rose gradually to (1.12 ± 0.01 kWh m⁻³), since energy consumption scales quadratically with voltage and competing side reactions reduce energy utilization.

ClO₂ → ClO₃⁻,

(6)

H₂O → O₂,

(7)

As shown in Figure 3f, the optimal pulsed frequency range was 5~50 Hz, consistent with Figure 5d. At frequencies below 10 Hz, increasing frequency reduces ROS self-decay during intervals, enhancing removal. Above 100 Hz, double-layer discharge, mass transfer limitations, and side reactions dominate, reducing performance. E-SEC reached its minimum (0.66 ± 0.03 kWh m⁻³) near 10 Hz (Figure 4f); lower frequencies improved efficiency, while higher ones increased Joule heating and switching losses.

Overall, the effects of frequency and potential on contaminant removal efficiency exhibit distinct nonlinear characteristics: a series of complex potential-dependent electrochemical processes occur under different potential conditions (as shown in Equations (3), (6) and (7)), while pulse frequency acts on mass transfer and charge transfer processes through intricate mechanisms. Additionally, the limitations of single-factor experiments make it difficult to eliminate the interference of interaction effects between variables; thus, it is often challenging to screen out energy-efficient optimal process parameters that meet specific removal efficiency requirements. Beyond energy consumption, considerations of practical applicability led us to further evaluate the economic impact of sodium chlorite reagent costs. The associated consumption cost per unit of SMX removed was calculated using Equation (8) (see Figure S10). This further highlights the necessity of optimizing the parameters of this electrochemical process using machine learning techniques in the subsequent sections.

$$\mathrm{W}(\mathrm{y}\mathrm{u}\mathrm{a}\mathrm{n}/\mathrm{g} \mathrm{S}\mathrm{M}\mathrm{X})={\displaystyle \frac{\mathrm{D}\times 90.44}{\mathsf{\omega }2-\mathsf{\omega }1}}\times 0.0118$$

(8)

where W represents the sodium chlorite consumption cost, D represents the initial sodium chlorite concentration, $\mathsf{\omega }2$ is the initial SMX concentration, and $\mathsf{\omega }1$ is the final SMX concentration.

3.3. Electrochemical Analysis

To gain an electrochemical perspective on the reaction mechanism underlying SMX removal, LSV, CV, and EIS measurements were conducted. Figure 5a demonstrates that, at the same potential, current density at 0.738 V increases markedly with rising NaClO₂ concentration, indicating that NaClO₂ undergoes electrochemical activation at the FeCuC-Ti₄O₇ electrode and that the reaction rate escalates with reactant concentration; these observations confirm the significant catalytic activity of the electrode toward ClO₂⁻ activation. However, the OER and ClO₃⁻ generation may occur beyond about 1.6 V, which may decrease the efficiency of the target activation reaction. These LSV curves explained the peak removal of SMX at 1.5 V in Figure 3e. For temperature, Figure 5b further shows that current density increases overall as temperature rises, and the onset potential for the reaction shifts slightly, signifying that elevated temperature favorably impacts the electrochemical reaction kinetics, which matches the removal results in Figure 3c. EIS spectra of magnitude and phase angle versus frequency are presented in Figure 5c and Figure 5d, respectively; upon addition of NaClO₂, both the low-frequency impedance magnitude and the phase angle decrease significantly compared with the blank condition (magnitude reduced from 2.1 × 10⁵ to 4.7 × 10⁴ Ω; phase angle shifted from −44.3° to −7.5°), indicating that introduction of NaClO₂ substantially ameliorates the electrode surface state and reduces both mass-transfer and charge-transfer resistances associated with the electrode reaction. Collectively, these electrochemical test results demonstrate that the FeCuC-Ti₄O₇ electrode exhibits excellent electrocatalytic reduction activity toward NaClO₂ in the tested system; thermal activation further enhances the reaction kinetics, and the electrode interfacial properties are significantly optimized upon addition of the reactant—characteristics that jointly promote high-efficiency electrochemical removal of SMX in this pulsed system. While electrochemical analysis has inherent limitations, these findings still partially elucidate how various factors within this system impact both the contaminant removal efficiency and energy consumption outcomes.

3.4. Artificial Intelligence Optimization

As described in Section 3.2, both SMX removal efficiency and E-SEC are nonlinearly influenced by multiple operational parameters (initial concentration, NaClO₂ dosage, temperature, potential, and frequency), and a trade-off exists between pollutant removal and energy cost. The complex behavior—featuring nonlinear kinetics (e.g., polarization and pathway switching), strong parameter coupling (e.g., temperature–side reaction interactions), and multi-objective optimization—renders linear models (such as LASSO regression) inadequate. Therefore, a BPNN was introduced to capture high-dimensional nonlinear correlations based on the universal approximation theorem. This approach enables precise prediction of removal performance and energy efficiency, providing an intelligent optimization framework that surpasses empirical parameter tuning. The BPNN’s architecture is presented in Figure 6a,b. The model was initially developed using 255 historical datasets and has been iteratively refined to enhance its performance.

The accuracy and reliability of the established BPNN model were assessed using the testing dataset. The RMSE and R² were employed as evaluation metrics, calculated as follows:

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}})}^2}{\mathrm{n}}}}$$

(9)

$${\mathrm{R}}^2={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{(\overline{{\mathrm{y}}_{\mathrm{i}}}-\widehat{{\mathrm{y}}_{\mathrm{i}}})}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{(\overline{{\mathrm{y}}_{\mathrm{i}}}-{\mathrm{y}}_{\mathrm{i}})}^2}}$$

(10)

where ${\mathrm{y}}_{\mathrm{i}}$ represents the input value, $\widehat{{\mathrm{y}}_{\mathrm{i}}}$ the output value, $\overline{{\mathrm{y}}_{\mathrm{i}}}$ the mean of inputs, and $\mathrm{n}$ the number of test samples. The results confirmed the high fitting accuracy of the BPNN model, further validating its predictive robustness and suitability for multi-parameter electrochemical systems.

Model evaluation results consistently confirmed that the BPNN model delivered the most accurate predictions for both SMX removal efficiency and E-SEC across all metrics. As illustrated in Figure 6i, the BPNN’s prediction of E-SEC achieved an ${\mathrm{R}}^2$ value of 0.9996—approaching unity—signaling an almost perfect alignment between the model’s outputs and the experimental measurements, a level of precision rarely achieved by conventional regression methods. To contextualize this performance, Figure 6c–e compare the predictive capabilities of the BPNN, GBDT, and RF models for SMX removal: the BPNN (with an ${\mathrm{R}}^2$ of 0.9917) outperformed the GBDT (0.9084) and RF (0.6473) by a substantial margin, reflecting its ability to capture the subtle nonlinear relationships between operational parameters and pollutant degradation. A parallel trend emerged for E-SEC predictions (Figure 6f–h): the BPNN’s ${\mathrm{R}}^2$ of 0.9996 far exceeded the GBDT’s 0.9172 and the RF’s 0.7390, further validating its superiority in modeling energy consumption dynamics. Figure 6i synthesizes these metrics, highlighting that the BPNN not only secured the highest ${\mathrm{R}}^2$ values for both SMX removal (0.9917) and E-SEC (0.9996) but also minimized prediction errors: its RMSE for SMX removal (1.4813) and E-SEC (0.0671) were the lowest among the three models, while the GBDT performed moderately and the RF lagged notably. Finally, Figure 6j benchmarks model predictions against actual operational data: the BPNN’s predicted SMX removal (71.795) and E-SEC (0.3562) closely mirrored the real-world values (71.225 and 0.37), whereas the RF overestimated both metrics and the GBDT underestimated them. Collectively, these results underscore the BPNN’s unparalleled predictive reliability for optimizing both treatment efficacy and energy efficiency in this system.

3.5. Practical Application Assessment

For practical wastewater treatment applications, the stability of the removal process, compliance of the effluent with reuse standards, and potential ecological risks are key evaluation indicators. As shown in Figure 7a, when the water matrix composition varies, the removal efficiency for high initial SMX concentrations (100 ppm) declines to some extent; in particular, the presence of chloride ions (Cl⁻) and humic substances in the water reduces the SMX removal efficiency by approximately 14.22% and 19.66%, respectively. Nevertheless, because the concentrations of these matrix constituents in real wastewater are typically lower than the levels used in these controlled experiments, the IPEA_NaClO2 system is predicted to exhibit relatively stable removal performance under field conditions, with SMX removal efficiencies reaching 92.53% in the tested practical scenario. Figure 7b demonstrates that the total E. coli. counts (ATCC 25922) in the wastewater were significantly reduced after treatment with the IPEA_NaClO2 system, and plate colony counting (Figure 7c) further verifies that the microbial inhibition (inactivation) rate achieved by this treatment can reach 98.99%, meeting the microbial indices required for reclaimed irrigation water. Moreover, as shown in Figure 7d, compared to untreated wastewater, the use of IPEA_NaClO2-treated water for agricultural irrigation substantially decreases SMX residues in lettuce root tissues (reduced by ~72.89 μg g⁻¹), in aboveground parts (~18.31 μg g⁻¹ reduction), and in soil (~42.21 mg g⁻¹ reduction), indicating the technology’s potential to markedly lower ecological accumulation and associated environmental risk. Collectively, these findings demonstrate that the IPEA_NaClO2 system exhibits favorable stability for real wastewater treatment, delivers effective disinfection performance, and substantially reduces SMX residues in agricultural contexts—supporting its feasibility for large-scale implementation and application in reclaimed-water irrigation systems. Simultaneously, to evaluate its practical feasibility, long-term stability tests were further conducted (Figure S9). The results indicate that the composite electrode maintains stable performance over extended periods, which not only prevents secondary environmental pollution but also effectively reduces the operational cost of removal.

However, this study only took the removal of SMX and energy consumption as optimization objectives, without considering the formation of intermediates and TOC removal. This limitation can be addressed in future research. In more practical application scenarios, advanced AI should be capable of predicting multi-dimensional water treatment outcomes, such as TOC removal and the generation of toxic byproducts. Furthermore, it should adapt to variations in influent water quality, such as different water matrix conditions.

4. Conclusions

This study successfully developed an efficient removal system based on pulsed electrochemical activation of NaClO₂, integrating a FeCuC-Ti₄O₇ composite electrode with ML to achieve high-efficiency SMX removal and energy consumption optimization. Under optimized operational conditions (NaClO₂ concentration 60~90 mM, pulsed frequency 10 Hz, temperature 313 K), the system achieved an SMX removal efficiency of nearly 90% while maintaining a low E-SEC of 0.66 kWh m⁻³. Comparative analysis of multiple ML models confirmed that the BPNN outperformed GBDT and RF in predicting SMX removal efficiency and E-SEC, with an R² value approaching 1, demonstrating its high-precision modeling capability for complex electrochemical processes. Practical water matrix tests further verified that the IPEA_NaClO2 system exhibited robust removal stability across varying water qualities, efficient disinfection performance (bacterial inactivation rate > 98%), and significant reduction in SMX residues in simulated agricultural irrigation scenarios. These results highlight the technology’s potential to mitigate the ecological risks associated with SMX accumulation. Overall, this study provides a feasible technical pathway and methodological support for the intelligent control and efficient removal of refractory organic pollutants. Although additional data, as well as the incorporation of potential environmental variables and water matrix parameters, are still required to train the system for enhanced practical applicability, the integrated IPEA_NaClO2 system has initially established its framework and shows promising application prospects in reclaimed water reuse and environmental pollution control—offering a potential solution for addressing persistent antibiotic contamination in aquatic environments.