Interfacial engineering of GQD/TiO₂-embedded photocatalytic concrete for visible-light-driven wastewater treatment: from charge separation to machine-learning optimization

Abstract

This study presents the design and evaluation of graphene quantum dots (GT) titania embedded in cementitious concrete (GQD/TiO₂-loaded concrete, GTC) for visible-light-driven photocatalytic decolorization of Reactive Black 5 (RB5), integrating experimental characterization with machine learning (ML) modeling. GT photocatalysts produced using the sol–gel approach outperformed bare TiO₂ in terms of visible-light activity. The 1 mol% GT calcined at 300 °C achieved approximately 100% RB5 elimination within 80 minutes. Embedding GT in concrete efficiently immobilized the photocatalyst while retaining high efficiency (89.7%), proving structural stability and reusability over numerous cycles. X-ray photoelectron spectroscopy (XPS) verified Ti–O–C interfacial bonding and surface –OH enrichment, whereas photoluminescence quenching revealed reduced electron–hole recombination. Radical scavenging experiments identified ˙O₂⁻ and ˙OH as dominant reactive species. High-performance liquid chromatography (HPLC) and chemical oxygen demand (COD) analyses confirmed the cleavage of chromophoric and aromatic structures into low-molecular-weight intermediates, achieving substantial mineralization and outperforming conventional TiO₂ systems. Ten supervised and deep learning machine learning models were developed to capture complicated, nonlinear behavior, with CatBoost and convolutional neural networks (CNN) outperforming others in terms of predicting accuracy. The key regulating elements, according to feature importance analysis, are reaction time and GTC dosage, followed by initial dye concentration. The integrated experimental-ML framework offers a scalable and interpretable strategy to develop photocatalytic concrete systems, allowing for predictive design and effective wastewater treatment. This study develops an approach for long-lasting, high-performance photocatalytic concrete, integrating mechanistic knowledge with data-driven optimization to provide long-term environmental cleanup.

Graphene quantum dot-doped TiO₂ photocatalytic concrete enables efficient visible-light degradation of Reactive Black 5. Machine learning accurately predicts photocatalytic performance for sustainable wastewater treatment.

1. Introduction

Rapid industrialization, combined with an increase in energy demand, has increased environmental contamination, making water quality preservation a crucial global concern. Textile dye effluents, in particular, are among the most refractory and visually disruptive types of industrial wastewater, posing serious threats to aquatic ecosystems and human health. Notably, azo dyes such as Reactive Black 5 (RB5) exhibit high color persistence and complex aromatic structures that resist conventional biological degradation, necessitating the development of innovative remediation strategies.^1–3 Advanced oxidation processes (AOPs) that utilize heterogeneous photocatalytic methods with semiconductor materials, such as titanium dioxide (TiO₂), have garnered significant attention for their effectiveness in removing organic pollutants like RB5. This interest is largely due to TiO₂'s low cost, chemically inert nature, and strong oxidative capabilities when exposed to UV light. However, the wide bandgap (∼3.2 eV) and rapid electron-hole recombination of TiO₂ limit its efficiency under visible or indoor lighting, hindering its practical application in real-world wastewater treatment scenarios.^4,5 Recent research has focused on developing multifunctional heterojunction designs that facilitate simultaneous pollutant degradation and the synthesis of value-added products. For instance, Cu₂-xS@ZnIn₂S₄ hollow heterojunction nanoboxes have been shown to enhance photocatalytic H₂O₂ production while effectively degrading tetracycline.⁶

To make it visible light active by reducing the bandgap, different impurities like metal, non-metal metal or a combination of both can be doped for environmental applications. Carbon-based impurities like graphene quantum dots (GQDs), due to their unique features like electron transfer properties, optical, and flexible surface functionalization properties, are an excellent choice for bandgap tuning.^7–9 Multiple studies have reported GQDs/TiO₂ with improved photocatalytic performance (improved discoloration and mineralization kinetics of various industrial dyes) with extended visible light absorption, enhanced charge separation, and photosensitization effects.^10,11 Despite this fact, GQD/TiO₂ photocatalysts have limited recyclability due to partial GQD detachment, surface fouling from reaction intermediates, and oxidative degradation under continuous irradiation. These factors reduce visible-light activity during repetitive cycles. Catalyst recovery through filtration or centrifugation adds to material loss and operational complexity in slurry systems. These difficulties emphasize the requirement for effective GQD anchoring and immobilization techniques, such as integration into cementitious matrices, to improve catalytic stability and long-term reusability.^12–14

Concurrently, immobilizing photocatalysts inside cementitious matrices to generate photocatalytic concretes is an emerging technique for incorporating pollutant abatement into constructed infrastructure.^15–17 Photocatalytic concrete embedded with semiconductors doped composites like TiO₂ are being extensively studied for self-cleaning surfaces, gaseous pollutant removal (e.g., NO_x, VOCs), and aqueous pollutant degradation. This method not only allows for the large-scale deployment of heterogeneous photocatalysts, but it also reduces the catalyst recovery issues associated with slurry systems. The VOS viewer literature bibliographic map for photocatalytic concrete from the Science Direct database (Fig. 1) indicates TiO₂-based photocatalyst as the central research hub, strongly linked with self-cleaning concrete, air pollution mitigation, and NO removal since the last 10 years. Recent trends (2020–2025) show a shift toward heterogeneous photocatalysis, reflecting growing environmental and public health applications.

Fig. 1. VOS viewer bibliographic literature for photocatalytic concrete application.

However, one of the primary issues in materials design is ensuring certainty that nanoscale catalysts are successfully incorporated into concrete while retaining long-term durability and active site accessibility. For practical applications to advance, precise assessment and prediction of photocatalytic performance are crucial, in addition to material production and structural optimization. Response surface methodology (RSM), Langmuir–Hinshelwood kinetics, and classical isotherms are examples of traditional modeling and optimization techniques that frequently rely on oversimplified assumptions (e.g., homogeneous surfaces, constant equilibrium) that overlook the intrinsic nonlinear interactions between operational variables like pH, photocatalyst quantity, contaminants concentration, intensity of light, and concrete surface properties.¹⁸ These constraints reduce predictive dependability, especially outside of the narrow design space of the experimental settings. Recently, machine learning (ML) models have become an effective means for modeling nonlinear, complicated systems without the need for precise mechanistic formulations.^9,19 In heterogeneous photocatalytic systems, where multiple interdependent variables influence performance, machine learning algorithms (including ensemble and neural network models) outperform traditional statistical models in terms of learning intricate patterns from experimental data and predictive accuracy. To date, however, the complete applications of a combination of supervised, unsupervised, and deep learning frameworks for the prediction of photocatalytic operational efficiency have remained unexplored.

In this regard, the current study strengthens photocatalytic concrete research by combining materials design, mechanistic validation, and data-driven modeling. Controlled graphene quantum dot (GQD) doping of TiO₂ creates a stable Ti–O–C heterointerface, extending visible-light absorption, suppressing charge-carrier recombination, and enabling sustained ˙OH and ˙O₂⁻ radical generation. Immobilization within a cementitious matrix preserves photocatalytic efficiency and overcomes long-standing challenges of catalyst recovery, durability, and scalability. To test the idea, GQD/TiO₂ (GT) photocatalysts were produced with varied GQD loadings and calcination temperatures, followed by integration into concrete at different weight fractions to develop GT-loaded concrete (GTC). The degradation of Reactive Black 5 (RB5) by visible light was comprehensively explored under environmentally relevant settings, with performance assessed through discoloration, mineralization (COD removal), kinetic analysis, radical scavenging, and HPLC-resolved degradation pathways. Batch studies were performed to thoroughly evaluate the influence of major operational variables such as GTC dose, solution pH, initial dye concentration, and reaction duration on RB5 visible-light-driven photodegradation. Recognizing that immobilized photocatalytic systems have strong nonlinear dependencies that cannot be reliably captured by traditional kinetic or polynomial models, this work incorporates advanced machine learning models with explainable analytics to enable accurate performance prediction and rational optimization. This work integrates machine learning (ML) with mechanistic experiments to advance photocatalytic concrete research. Conventional kinetic models and one-factor-at-a-time techniques fail to represent the nonlinear dynamics that regulate immobilized photocatalytic systems. Advanced machine learning models were combined with explainability tools (SHAP and PDP) to obtain near-perfect predicted accuracy while maintaining physical interpretability. The hypothesis also predicts that ML algorithms can effectively evaluate and forecast complex, nonlinear relationships between various reaction parameters that conventional manual optimization or classical statistical approaches cannot under the same reaction conditions. Multiple machine learning (ML) models were developed in parallel to capture complicated, nonlinear connections among process variables. Predictive performance was evaluated using comprehensive error measures, including coefficient of determination (R²), Root Mean square error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Median Absolute Error (MedAE), Scatter Index (SI), and Log-Cosh Error (LogCosh). A comparative analysis revealed that ML models had higher prediction accuracy and generalization potential. This work is innovative in that it integrates ML modeling framework with sophisticated carbon-modified photocatalytic concrete to enable scalable optimization, mechanistic validation, and precise efficiency prediction. This integrated materials-process-modeling method holds promise for data-driven improvement of sustainable wastewater treatment technologies and offers a solid route for the systematic design of long-lasting, high-performance photocatalytic concrete systems.

2. Research methodology

2.1. Materials

All the chemicals used in this study, along with the estimated cost and lab-scale cost estimation of the synthesized photocatalysts, are listed in Table S1. The model pollutant was the textile azo dye, RB5, (C₂₆ H2₁ N₅ Na₄ O₁₉ S₆), provided by Koh e Noor Textile Mills. All chemicals and materials were used as received without further purification. For concrete fabrication, ordinary Portland cement (OPC) CEM I 42.5R compliant with TS EN 197-12012, fine aggregate, sand, and aggregate dust (≤0.5 mm) were procured from recycled aggregates from a local construction site in Abbottabad, Pakistan. The aggregates are customized to relevant American Society for Testing Soil Materials (ASTM) specifications (Unified Soil Classification System per ASTM D487-23).

2.2. Synthesis of GT photocatalyst and GTC photocatalytic concrete

2.2.1. Synthesis of GQDs, GQD/TiO₂ (GT) photocatalyst, and photocatalytic concrete (GTC)

GQDs (varying masses i.e., weights in g of citric acid (0.0404, 0.2053, 0.4213, 0.6671, and 1.0006 g) corresponding to GQD loadings of 1, 5, 10, 15, and 20 mol% were synthesized via pyrolysis of citric acid, adapted from a reported method.²⁰ The initial white crystals melted into a colorless liquid, which subsequently transitioned to an orange hue, indicating the formation of GQDs. The resultant orange liquid was cooled and then neutralized by dropwise addition of 1.0 M NaOH under continuous stirring to ensure a homogeneous dispersion of GQDs.

The GQD/TiO₂ photocatalysts were synthesized through the sol–gel method.^13,21 Two solutions were prepared; solution A contained 37 ml of TTIP dissolved in 60 ml of absolute ethanol. Solution B (a mixture of acetic acid (15 ml), ethanol (20 ml), DI water (10 ml), and a specified amount of the GQD solution) was added to solution A, under vigorous stirring. The resultant gel was then aged at room temperature (22 °C ± 1) for 24 hours, which was subsequently dried at 110 °C overnight. The dried gel was then ground into fine powder and annealed in a muffle furnace for 1 hour at different calcination temperatures (300, 400, and 500 °C) to obtain the calcined GT photocatalyst. This same procedure was repeated to synthesize different GQD loadings of 1, 5, 10, 15, and 20 mol%, each calcined at specified calcination temperatures.

Photocatalytic concrete (GTC) palettes were produced by incorporating the synthesized GT powder into a standard cementitious matrix, employing the same method as reported in our recent study.²² The concrete mix proportion is detailed in Table S2. All components were mixed manually in deionized water to a uniform paste, cast into molds (1 cm × 1 cm × 1 cm, per ISO 10678:2010), demolded after 24 hours, and cured under ambient laboratory conditions for 7 days. The cured palettes were air-dried for 24 hours prior to photocatalytic performance assessment.

2.3. Photocatalytic degradation studies

The photocatalytic activity of both the synthesized GT photocatalysts and the GTC palettes was evaluated through the discoloration of RB5 under visible light irradiation. For photocatalyst discoloration studies, the GT photocatalyst (1 mg mL⁻¹) was ultrasonically dispersed in the RB5 solution of known concentration. For GTC discoloration studies, a palette was immersed in the dye solution to be subjected to performance assessment. In both cases, the RB5 solution was first stirred in the dark (for 30 minutes) to establish adsorption–desorption equilibrium, followed by irradiation using a 500 W halogen lamp photoreactor at a certain distance from the top. Aliquot samples were withdrawn at regular intervals, centrifuged to remove particles, and the residual dye concentration was analyzed by UV-visible spectroscopy (PG Instruments T80+). Control experiments were performed without light and/or without a catalyst.

The decolorization/mineralization efficiency (%) was calculated using eqn (1).

1

where: C₀ is the initial dye concentration (mg L⁻¹), and C_t is the dye concentration (mg L⁻¹) at different intervals of time for the period of light irradiation.

In optimization studies on RB5 dye discoloration, various reaction conditions were evaluated to assess the effect of four operational variables: GTC dose, reaction pH, initial dye concentration (IDC), and light exposure time.

2.4. Kinetic modeling, mechanistic analysis, and recycling studies

Kinetic modeling, surface charge analysis, reactive species trapping, LCMS/HPLC analysis, and recyclability tests were used to assess the photocatalytic activity of graphene quantum dot-doped TiO₂-loaded concrete (GTC) against RB5 degradation. To investigate the link between initial dye concentration (IDC) and apparent reaction rate constants, the Langmuir–Hinshelwood (L–H) model was applied to the degradation kinetics. Scavenging experiments were performed to identify the radical scavengers or dominant reactive oxygen species (ROS) responsible for RB5 degradation. Four selective quenching agents were used: disodium ethylenediaminetetraacetate (EDTA-2Na) as a hole (h⁺) scavenger, isopropanol (IPA) as a hydroxyl radical (˙OH) scavenger, calcium chloride (CaCl₂) as an electron (e⁻) scavenger, and ascorbic acid (AA) as a superoxide radical (O₂˙⁻) scavenger. To monitor the transformation of RB5, HPLC analysis was performed before treatment (0 min) and after 80 minutes of visible-light exposure. Chromatograms were taken at 254 nm (sensitive to aromatic structures) and 365 nm (selective for conjugated chromophores). Peak area, peak intensity, and retention time were utilized to assess intermediate formation and molecular degradation. Over six sequential cycles, recyclability tests were performed to examine catalyst stability and reusability under visible-light irradiation. To identify the photocatalytic degradation intermediates of RB5, liquid chromatography-mass spectrometry (LC-MS) was performed using a Shimadzu system equipped with LabSolutions software. Aliquots collected at specific irradiation intervals were filtered through a 0.22 µm membrane to remove catalyst particles, followed by the injection of 3 µL samples. Mass spectrometric analysis employed electrospray ionization (EI) in full-scan mode to monitor the transformation of RB5 and its fragments. By processing the chromatograms and mass spectra, the decay of the parent dye and the sequential appearance of intermediates were tracked, allowing for the systematic determination of the degradation pathway.

2.5. Integrated experimental design and predictive data-driven modeling framework

This section describes an integrated experimental and data-driven framework that incorporates batch screening studies for systematic experimental design and optimization, as well as machine learning (ML) models to capture complex, nonlinear relationships, thereby enhancing the accuracy of photocatalytic dye degradation predictions. ML model performance assessment enables the selection of the most precise and accurate prediction for RB5/ML degradation effectiveness, thus highlighting the potential of ML modeling as a complementary optimization tool to experimental approaches in photocatalytic concrete systems. Machine learning models were developed for capturing complicated, nonlinear interactions between process variables and comparing their predictive accuracy to traditional approaches.

2.5.1. Machine learning (ML) modelling

This study provides a systematic assessment of the predictive performance of ten machine learning (ML) models, including: seven ensemble learning algorithms, and three neural network/deep learning architectures, including CatBoost (CB), Random Forest (RF), Bagging, Histogram-based Gradient Boosting (HGB), Light Gradient Boosting Machine (LGBM), Gradient Boosting Decision Tree (GBDT), Gaussian Process Regression (GPR), Artificial Neural Network (ANN), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN). The machine learning model description is explained in Table S3. The use of ensemble tree-based models was due to their strong ability to capture nonlinear interactions by aggregating or sequentially optimizing decision trees, and CB offers additional benefits for complex feature relationships. GPR was incorporated to make probabilistic predictions with uncertainty, and ANN, LSTM, and CNN architectures were employed to simulate complex nonlinear and temporal relationships in photocatalytic degradation datasets. The rate of model performance was numerically measured in terms of various loss and accuracy measures: coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), median absolute error (MedAE), scatter index (SI), and Log-Cosh Error (LogCosh). The ML Loss function statistical metric indicators are presented in Table S4. This materials-process-modeling system facilitates a powerful assessment of the GQD/TiO₂-based photocatalyst concrete operation and allows predictive evaluations on how to optimize scalable photocatalytic wastewater treatment systems.

2.5.1.1. Pearson correlation coefficient

The Pearson Correlation Coefficient (PCC), as a widely applied metric used to quantify the strength and direction of the linear relationship between two continuous variables, was used to assess the different features of this study. This helps evaluate how the deviation of variables from their mean corresponds proportionally to another, thereby providing a normalized measure of association. The coefficient values range from −1 to +1, where values close to +1 indicate a strong linear relationship, near −1 indicate a strong negative linear relationship, and values near 0 reflect an absence of linear dependence. The PCC assumes interval or ratio-scale data, approximate normality, and sensitivity to outliers, making it particularly suitable for analyzing linear associations in experimental data. The PCC is statistically expressed as follows. x_i

2

where x and y denote individual observations, x̄ and ȳ represent variable means, and n is the number of paired samples.²³

2.5.1.2. ML data preparation

The 1 mol% GT photocatalyst calcined at 300 °C (denoted as 1GT-3) was identified as having the highest RB5 photocatalytic efficiency, 98% within 80 min under visible-light irradiation, which was then added to cementitious matrices at various weight fractions (0, 2, 4, 6, 8, 10 and 20 wt%) to synthesize photocatalytic concrete palettes (GTC). A series of controlled batch photocatalytic tests was carried out to assess the impact of key operational factors on RB5 degradation, and these results served as the primary dataset for the machine learning (ML) study (Table 1). The target output was established as RB5 removal efficiency (%), and four crucial input parameters were chosen as independent variables: GTC dosage (wt%), solution pH, initial dye concentration (IDC, mg L⁻¹), and irradiation time (min). Due to the one-factor-at-a-time (OFAT) nature of the experiments, which restricts the detection of true interaction terms and variable interactions, the dataset was expanded within the experimentally observed region through a nearest-neighbor-based interpolation approach modeled after the Synthetic Minority Over-sampling Technique (SMOTE) to address this limitation. This approach as reported in ref. 24, generated synthetic data points by interpolating between contiguous observations in the multidimensional feature space (GTC dosage, pH, starting dye concentration, and irradiation period), yielding a dataset of n = 1680. This method preserves local trends and increases data density for nonlinear modeling, but, due to the additive independence assumption inherent in OFAT data, it may introduce artificial smoothing and partially conceal experimental noise. To ensure robust model construction and reduce overfitting, the dataset was randomly divided into training (80%) and testing (20%) subsets, and model optimization was carried out using 5-fold cross-validation in conjunction with Randomized Search, Grid Search, and Optuna frameworks. Description of Machine learning model applied in the present study is presented in Table S3. Multiple statistical metrics were used to evaluate model performance, including R², RMSE, MAE, MAPE, MedAE, Scatter Index, and Log-Cosh Error. The nearly-perfect predictive performance observed for certain models reflects the internally consistent nature of the augmented dataset and controlled experimental conditions; however, external validation with independent datasets is required to confirm generalizability under real-world conditions. In this study python 3.12.12 with google collaborative (google Colab) notebook with standard libraries such as pandas, matplolit.pyplot, numpy, sklearn, optuna, tensorflow.keras, MinMaxscaler among others.

Table 1. Batch studies data collection from reaction optimization.

GTC dose (wt%)	RB5 removal (%)	pH	RB5 removal (%)	IDC (mg L−1)	RB5 removal (%)	Reaction time (min)	RB5 removal (%)
0	15.6	1	100	10	90.0	10	24.5
2	84.2	3	99.5	20	89.7	20	52.5
4	85.3	5	94.6	30	72.4	30	58.2
6	86.2	7	89.7	40	62.5	40	62.7
8	87.5	9	65.0	50	54.8	50	70.3
10	89.7	11	49.4	—	—	60	78.9
20	71.5	—	—	—	—	70	82.2
—	—	—	—	—	—	80	89.7

The observed 15.6% RB5 removal at 0 wt% GTC (BTC-3 calcined at 300 °C in Table 1) corresponds to experiments conducted under 80 min of light irradiation in the absence of photocatalyst loading. This removal is attributed primarily to photolysis and potential minor surface-assisted degradation on the concrete substrate rather than adsorption or photocatalytic activity. It is important to distinguish this baseline contribution from true photocatalytic degradation. Although dark control experiments (no irradiation) with catalyst loading gave (% removal) that enabled direct quantification of adsorption-only removal, such measurements were not included in the present study (specifically as ML parameter) and are recommended for future work to further refine the evaluation of photocatalytic efficiency in terms of adsorption studies.

2.6. Characterization

The morphological, structural, elemental, and surface charge features of undoped titania, GT photocatalysts, and their related concrete photocatalysts were thoroughly studied to reveal structure–property relationships that influence photocatalytic activity. The morphological properties were investigated using scanning electron microscopy. The composition and phase transition of the photocatalyst were evaluated using X-ray diffraction (XRD). The electrokinetic and colloidal stability of GT photocatalysts calcined at different temperatures (300, 400, and 500 °C) was assessed via dynamic light scattering (DLS). Zeta potential measurements were used to evaluate the dispersion stability and surface charge characteristics in aqueous systems. Autocorrelation functions (sometimes called correlograms) were used to assess temporal stability and Brownian motion behavior. The point of zero charge (PZC) of the modified GT photocatalyst was estimated using the pH drift method to better understand the pH-dependent surface charge behavior that is significant to dye adsorption and photocatalytic performance. PZC measurements were also performed to further understand the electrostatic interactions between the catalyst surface and the RB5 dye molecules. To assess the recombination of electron–hole pairs, photoluminescence (PL) spectroscopy was used to record the PL spectra of calcined GT photocatalysts (at 300 °C, 400 °C, and 500 °C) using a Fluorescent Spectrophotometer (RF-6000, Shimadzu, Japan). The GT photocatalysts' surface chemical composition and oxidation states were investigated using XPS analysis.

2.7. Phytotoxicity test

The phytotoxicity of untreated RB5 dye and treated wastewater was evaluated using a mung bean (Vigna radiata) seed germination bioassay. Seeds were surface sterilized with 2% sodium hypochlorite for 2 minutes and rinsed three times with distilled water. Four treatments were tested in triplicate: distilled water (negative control), 0.1% CuSO₄ (positive control), untreated RB5 dye, and treated wastewater. Ten seeds were placed in Petri dishes lined with filter paper and moistened with 5 mL of test solution. Dishes were incubated in darkness at 25 ± 2 °C for 4 days. Germination percentage and root length (mm) were recorded on day 4. One-way ANOVA followed by Tukey's HSD post-hoc test (α = 0.05) was used for statistical analysis.

3. Results and discussions

3.1. Photocatalytic degradation studies (screening and optimization)

3.1.1. Screening and optimization of GQD/TiO₂ photocatalysts

To assess the impact of calcination temperature and graphene quantum dot (GQD) loading on the visible light-driven photo decolorization of RB5, screening studies were carried out. Fig. 2a shows that pristine TiO₂ has limited photocatalytic activity in visible light, indicating its broad bandgap and quick charge-carrier recombination. In contrast, including GQDs dramatically improves RB5 removal efficiency, demonstrating the beneficial impact of GQD modification. The enhanced photocatalytic performance is attributed to improved interfacial charge transfer and suppression of electron–hole recombination, facilitated by GQD-mediated photosensitization and electron-acceptor behavior, as evidenced by reactive species trapping experiments. The maximum RB5 photo decolorization effectiveness (98.7% within 80 min) among the compositions under investigation was achieved with a modest GQD loading of 1 mol%; however, activity gradually decreased with additional GQD content. RB5 elimination dropped to about 93% at moderate loading (5 mol%), and further performance degradation occurred at higher loadings (10–20 mol%). This pattern suggests that an ideal GQD concentration is needed to balance charge separation and light capture. Excessive GQD loading may cause nanoparticle agglomeration, shield active TiO₂ sites, and introduce recombination centers, reducing charge mobility and photocatalytic performance. The impact of calcination temperature on photocatalytic performance was significant (Fig. 2b). The 1GT-3 had the highest RB5 degradation efficiency due to optimum crystallinity, maintaining anatase TiO₂ phase, and efficient GT interfacial contact. While higher temperatures (500 °C) probably cause particle sintering, reduced surface area, and partial degradation of GQD surface functions, all of which reduce photocatalytic activity, calcination at 400 °C may minimize defect formation and interfacial coupling.²⁵

Fig. 2. Screening and optimization studies of GQD/TiO₂: effect of (a) GT loading (mol%), (b) calcination temperature (°C), (c) 1GT-3 dose (g L⁻¹), (d) pH, (e) irradiation time (min), (f) IDC (mg L⁻¹).

The effect of operational factors on RB5 photo decolorization with the optimized 1GT-3 photocatalyst was investigated subsequently. Fig. 2c shows that efficiency rose with catalyst dosage and peaked at 1 g L⁻¹ before declining. At higher dosages, catalyst aggregation and excessive turbidity enhance light scattering and restrict photon penetration, limiting the accessibility of photoactive sites. Fig. 2d illustrates the influence of solution pH on RB5 elimination. Higher degradation efficiencies were obtained in acidic to near-neutral environments (pH 1–5), where electrostatic interactions between the positively charged photocatalyst surface and the anionic RB5 molecules promote adsorption and interfacial redox processes.³ Photocatalytic activity dramatically declined in alkaline conditions, most likely due to competing adsorption of hydroxide ions on active sites, decreased production of hydroxyl radicals, and electrostatic repulsion. Reaction time experiments (Fig. 2e) demonstrated rapid RB5 degradation in the early phases, followed by a progressive plateau after roughly 80 minutes, indicating active site saturation and the creation of a dynamic adsorption–degradation equilibrium. Heterogeneous photocatalytic systems exhibit this pattern, which validates the effective use of photogenerated reactive species in the initial phases of reaction.⁵

Increasing the initial RB5 concentration (IDC) from 10 to 50 mg L⁻¹ resulted in a gradual decrease in RB5 removal efficiency (Fig. 2f). Excess dye molecules reduce photocatalytic efficacy at higher IDC by scavenging reactive oxygen species, attenuating light penetration, and competing for active sites. The optimized 1GT-3 photocatalyst exhibits outstanding RB5 photo decolorization under visible light with a catalyst dose of 1 mg L⁻¹, reaction duration of 80 min, near-neutral pH, and IDC ≤ 20 mg L⁻¹. The improved performance emphasizes the need for regulated GQD loading, proper thermal treatment, and balanced operational conditions in optimizing photocatalytic efficiency through improved separation of charges, surface interactions, and light harvesting.²⁶

3.1.2. Screening and optimization of GQD/TiO₂-loaded concrete (GTC)

The screening and optimization of the GT-loaded concrete (GTC) system identified the optimized conditions that maximise RB5 photodecolorzition (Fig. 2). Increasing the 1GT-3 loading from 0–10 wt% has improved efficiency from 21.2% to 89.5% as an optimal performance (Fig. 3a). Calcination temperature was also critical: 10-1GT-3 showed the highest activity of 89.5%, whereas TC (TiO₂ only concrete) remains lower, confirming the enhanced photocatalytic contribution of GT (Fig. 3b). Optimizing the GT dose revealed a similar trend of RB5 removal, though with reduced efficiency in concrete. Increasing GTC content shows a simultaneous increase in RB5 removal up to 10 wt% (Fig. 3a). pH studies showed near-complete removal under acidic to neutral conditions, while alkaline pH reduced efficiency, likely due to electrostatic repulsion that hinders dye adsorption (Fig. 3c). The reaction time profile showed steady performance, from 20 minutes to 80 minutes of light irradiation (Fig. 3d), achieving an efficiency of 89%. IDC had a strong effect: maximum removal occurred at 20 mg L⁻¹, while higher concentrations from 30–50 mg L⁻¹ lower efficiency through active site saturation and reduced light penetration (Fig. 3e).

Fig. 3. Screening and optimization studies using GTC: effect of (a) GT loading onto concrete (wt%), (b) calcination temperature (°C), (c) pH, (d) reaction time (min) (e) IDC (mg L⁻¹).

3.2. Kinetic modeling, mechanistic analysis, stability and recycling studies

The photocatalytic performance of GT/GTC is influenced by its surface chemistry and reaction kinetics; corresponding results are shown in Fig. 4, as demonstrated by kinetic modeling, scavenging analysis, LCMS, recyclability, and stability studies. The photocatalytic degradation kinetics of the GT photocatalyst follow a pseudo-first-order model, as presented in Fig. 4a, while Fig. 4b displays a Langmuir–Hinshelwood (LH) plot showing a linear relationship between the IDC and the reaction rate. This confirms that the degradation of RB5 follows surface-controlled pseudo-first-order kinetics typical of heterogeneous photocatalysis, with a coefficient of determination, R² = 0.9461 (Fig. 4b-1), an inset plot of the reciprocal of the apparent rate (K_app) of photodegradation vs. IDC.

Fig. 4. Photocatalytic degradation kinetics (a) pseudo-first order, (b) LH model (b-1) inset plot of reciprocal of apparent rate (K_app) of photodegradation vs. IDC, (c) reactive species trapping experiments, and (d) %RB5 vs. %COD removal (e) recyclability (f) stability studies and (g) potential RB5 intermediates generated during GT-assisted photodegradation determined by LC-MS.

HPLC analysis was used to monitor molecular changes in RB5 during visible-light photocatalysis over GTC at pH 7. HPLC chromatograms (Fig. S5) were taken at 254 and 365 nm to compare the dye before and after irradiation (0 min). The untreated RB5 solution exhibited strong, well-defined peaks mainly between 22 and 36 minutes, which corresponded to intact aromatic and azo-linked chromophoric components. Strong π-conjugation and molecular stability of the parent dye were validated by the high peak density and cumulative area at 365 nm. After 90 minutes of visible light irradiation, the chromatographic profile changed significantly, with the loss of classic RB5 peaks and the emergence of new peaks at shorter (<10 min) and intermediate (10–25 min) retention durations. Concurrently, a significant fall in peak intensity at 365 nm revealed successful azo chromophore destruction and extended conjugation, indicating RB5 fragmentation into smaller and more polar intermediates instead of simple decolorization. The individual byproducts were not analytically characterized, but retention-time alterations clearly indicate gradual oxidative breakdown. The loss of high-retention peaks and the appearance of early-eluting products are indicative of azo bond cleavage, followed by oxidation and partial ring opening of aromatic moieties, resulting in lower-molecular-weight oxygenated molecules. The lack of persistent intermediate accumulation implies continual oxidative transformation rather than a single-step conversion.

LCMS was used to identify photocatalytic degradation intermediates of the RB5 and Fragmentation pathways, as well as byproducts as presented in Fig. 4g. The suggested RB5 degradation mechanism reveals a multistep fragmentation process that is initiated by photogenerated reactive species. The first step includes dehydration of the parent molecule, which results in the creation of the initial intermediate fragment. The dye structure undergoes desulfonation and oxidative instability as sodium ions (Na⁺) are released from both sulfonated terminal groups. In the next stage, a crucial side-chain fragment (–CH₂–CH₂–OSOH) is removed from the molecular framework, followed by dehydrogenation, resulting in an unsaturated alkene moiety inside the remnant structure. As the reaction progresses, an extra sodium ion and sulfur oxide species (SO_x) are lost, indicating sulfonate functionalities cleave gradually. One aromatic substituent (the second branch associated with the aromatic ring system) is then separated as a distinct fragment, indicating that the dye's conjugated system has broken down. Continued irradiation results in a highly deconstructed aromatic residue that resembles a biphenyl-based core structure while retaining little sulfonated activity. The terminal degradation stage is characterized by the presence of a biphenyl-type intermediate with residual sulfonic acid groups and sodium sulfonate moieties, which are then mineralized to low-molecular-weight species.

Reactive species trapping experiments (Fig. 4c and Table S6) demonstrate diverse inhibition patterns that help to understand the stage-specific roles of reactive oxygen species in RB5 degradation. In the absence of efficient radical quenching (EDTA), total elimination (100%) was achieved in 80 minutes, demonstrating that photogenerated holes contribute indirectly through ROS formation rather than direct oxidation. The addition of isopropanol (IPA) lowered degradation to 72.13%, demonstrating that ˙OH radicals play a substantial but not dominant role. Benzoquinone (BQ) significantly reduced degradation to 34.86%, indicating that ˙O₂⁻ radicals are the major species responsible. The inhibition seen with AgNO₃ (48.20%) supports the role of photogenerated electrons in aiding ˙O₂⁻ production through oxygen reduction. All of these findings point to a stage-dependent mechanism in which ˙OH radicals cause rapid azo bond (–N N–) cleavage, which destroys chromophores and decolorizes them, while ˙O₂⁻ radicals control the oxidative degradation of aromatic intermediates, ring opening, and mineralization that follows. HPLC data confirms this view, with RB5 peaks persisting for 22–36 minutes and early-eluting intermediates accumulating for <10 minutes under ˙OH and ˙O₂⁻ scavenging conditions, indicating incomplete breakdown processes.¹¹ COD analysis (Fig. 4d) indicates substantial mineralization (82.1% at 80 min and 93.1% at 100 min for GT, whereas 77.61% at 80 min and 87.31% at 100 min for GTC), indicating the prominent involvement of ˙O₂⁻ in deep oxidation processes. The Ti–O–C-rich heterointerface, along with surface hydroxyl groups and oxygen vacancies, allows for efficient charge separation and prolonged ROS generation under visible light, resulting in both quick decolorization and considerable mineralization. This cooperative ˙OH/˙O₂⁻ mechanism presents a self-consistent explanation that reconciles scavenging, HPLC, and COD results.²⁷ The contribution of GQDs is thus attributed to improved charge separation and electron transfer, allowing for persistent ROS formation under visible light, evident by accelerated peak decay and the formation of low-retention intermediates in the absence of scavengers. True photocatalytic degradation of RB5 over GTC at neutral pH and visible light is directly demonstrated by comparing HPLC analysis.

Recyclability and stability studies are presented in Fig. 4(e) and (f). Recyclability studies of the GT and GTC systems over 6 cycles showed only a slight decline in performance, from 89.7% in the first cycle to 80.63% in the sixth. This modest reduction demonstrates good structural stability, highlighting GTC as a durable and cost-effective photocatalyst for sustained wastewater treatment applications. GTC system stability experiments with GT/GTC exposed to synthetic wastewater with common co-pollutants like NaCl is presented in Fig. 4f. A three-cycle recyclability experiment was carried out with 10 mM NaCl as a sample co-pollutant in order to evaluate catalyst stability under more practical circumstances. Over three cycles, GT fell from 92.5% to 86.3% and GTC from 83.5% to 73.7%, indicating a slight drop in degrading efficiency. This decrease is explained by ionic interference, which limits photocatalytic activity by having chloride ions partially occupy active sites and compete with reactive oxygen species. Despite this, the catalyst retains significant activity, showing good stability under saline conditions. Although real textile wastewater with surfactants and organic matter would give a more thorough examination, NaCl efficiently simulates salt effects found in textile effluents. GTC-based photocatalytic systems are most suited for tertiary polishing stages, where low organic load and turbidity limit fouling and allow for effective destruction of residual dyes and emerging pollutants.

3.3. Integrated experimental design and predictive data-driven modeling framework

3.3.1. Machine learning statistical descriptive analysis of the dataset

As a foundational step for ML workflows and data quality assessment based on our experimental raw data, descriptive analysis was performed. The result indicates that the operational parameters' performance outcomes exhibit moderate variation, with most variables showing slight left skewness and relatively flat platykurtic distributions. Both the GTC (average 7.1 wt%) and pH (average 6) show comparable levels of dispersion, suggesting consistent experimental control. In contrast, the initial dye concentration (IDC) displays a much wider spread, reflecting the broader range of loading conditions used in the study. Reaction time varies substantially, ranging from short to extended duration. RB5 removal efficiency averaged 71.7% with moderate variability and interestingly showed a negative skew, demonstrating that higher removal values occurred more frequently than lower ones. The quartile distributions for all parameters show steady, organized elevation with no sign of extreme outliers. Collectively, the dataset shows a well-defined parameter range and controlled variability as obtained from experiments (Table 2), making it a solid basis for statistical analysis and photocatalytic performance.

Table 2. Descriptive analysis of parameters.

Parameters/description	Mean	Standard deviation	Variance	Skewness	Kurtosis	Minimum	1st Quartile (Q1)	Median	3rd Quartile (Q3)	Maximum
GTC dose (wt%)	7.1	6.1	37.6	1.0	0.1	0.0	2.0	6.0	10.0	20
pH	6.0	3.4	11.7	0.0	−1.3	1.0	3.0	6.0	9.0	11
IDC (mg L−1)	30.0	14.1	200.1	0.0	−1.3	10.0	20.0	30.0	40.0	50
Reaction time (min)	45.0	22.9	525.3	0.0	−1.2	10.0	25.0	45.0	65.0	80
RB5 removal (%)	71.7	13.2	174.6	−0.6	0.2	21.2	63.6	73.2	81.3	100

3.3.1.1. Correlation analysis of parameters

The heatmap (Pearson correlation) brings to light the important connection between the process parameters and how effectively RB5 is removed. Both the GTC dose and the reaction time have positive correlations with removal performance. This means that when we increase the catalyst dose and allow for longer reaction time, we generally see better RB5 degradation results. IDC and pH show a negative correlation with the removal efficiency of RB5, this suggest that lower pH levels and reduced IDC boost their photocatalytic activity. The zero correlation variables are independent within the experimental setup, and each plays its own role in how the system behaves. Overall heatmap clearly shows that catalyst dosage and reaction time are key factors, followed by IDC and pH define levels (Fig. 5a). The Box plot gives a snapshot of variability in the experimental parameters and RB5 removal efficiency. The GTC doe and pH have narrow interquartile ranges, which tells us that these conditions were tightly controlled during the experiments. The IDC and reaction time show wider distributions, reflecting more variation in those experimental settings. RB5 removal does show some variability, with a few lower outliers indicating that degradation can sometimes be less effective under certain conditions. For most parameters, the mean and median values are closely aligned, tis suggest relatively symmetric distributions, while the skew in RB5 removal can be traced back to those lower outliers. Collectively, the plot emphasizes the stable control of catalyst dose and pH, the broader range in IDC and reaction times, and how these factors work together to influence photo-decolorization (Fig. 5b).

Fig. 5. (a) Heatmap of correlation and (b) Box plot visualization.

3.3.1.2. Loss functions for statistical metric indicator performance

The predictive performance of the ML models was effectively enhanced through systematic hyperparameter optimization (presented in Table S7). Among the selected ML models, the CB algorithm delivered the most predictive outcomes after cross-validation, characterized by perfect R² scores of 1.00 on both training and testing sets and consistently low error values across all performance indicators within a standard acceptable range, as seen especially with test metrics RSME (.109), MAE (0.059), and MAPE (0.101%) (Table 3). This result demonstrates their strong generalization capabilities and high accuracy in estimating the target variable. Conversely, LSTM exhibited the least performance, characterized by substantial metric magnitude error, RSME (6.574), MAE (5.383), and the lowest explained variance R² (0.764), after cross-validation, indicating inadequate fit and limited predictive capacity, confirming its status as the least performing among all models, followed by the CNN model. While other neural networks (ANN) produced moderate results, and others like RF, bagging, GBDT, LGBM, also demonstrated high efficacy with test R² values (0.9–1.000) and moderate errors within an acceptable range.

Table 3. Loss functions for statistical metric indicator performance.

N°	Model/metrices		R 2	RMSE	MAE	MAPE	MedAE	SI	LogCosh
1	CB	Train	1.000	0.072	0.043	0.072	0.027	0.001	0.003
		Test	1.000	0.0109	0.059	0.101	0.032	0.002	0.006
2	RF	Train	0.999	0.433	0.286	0.452	0.186	0.006	0.079
		Test	0.989	1.371	0.840	1.415	0.506	0.019	0.470
3	Bagging	Train	0.999	0.388	0.231	0.367	0.126	0.005	0.062
		Test	0.986	1.624	0.893	1.536	0.352	0.023	0.564
4	HGB	Train	1.000	0.195	0.146	0.217	0.114	0.003	0.018
		Test	1.000	0.248	0.180	0.274	0.132	0.003	0.029
5	LGBM	Train	1.000	0.127	0.088	0.136	0.063	0.002	0.008
		Test	1.000	0.279	0.163	0.269	0.097	0.004	0.033
6	GBDT	Train	1.000	0.122	0.089	0.139	0.066	0.002	0.007
		Test	1.000	0.191	0.138	0.223	0.098	0.003	0.018
7	GPR	Train	0.994	0.945	0.666	0.997	0.398	0.013	0.318
		Test	0.982	1.756	1.232	1.885	0.773	0.024	0.771
8	ANN	Train	0.994	0.972	1.771	1.136	0.645	0.014	0.356
		Test	0.992	1.168	0.915	1.366	0.764	0.016	0.475
9	LSTM	Train	0.785	6.093	4.859	7.520	4.032	0.085	3.214
		Test	0.764	6.574	5.383	8.211	4.591	0.092	4.729
10	CNN	Train	0.999	0.388	0.315	0.450	0.269	0.005	0.071
		Test	0.999	0.411	0.337	0.479	0.306	0.006	0.079

3.3.1.3. Regression plots for predictive modelling analysis

For the 10 selected models (Fig. S8), show a comparison between the model's predictive performance. Fig. 6 plots predicted RB5 photodegradation efficiencies against the experimental values for the best and least performing models, respectively. The predicted and observed efficiencies for most models are close to the diagonal line, suggesting that the ML algorithms are capable of properly capturing the behavior of photocatalytic degradation. Based on the performance metrics across the ten evaluated regression models, the CB model demonstrates the strongest capability, achieving near-perfect agreement between predicted and experimental values for both train and test sets, as indicated by an R² of 1.00 and exceptionally low error metrics (Table 3). This suggests it generalizes exceptionally well without significant overfitting. Other neural and tree-based ensemble models (LGBM, GBDT, RF, Bagging, HGB) also exhibit excellent performance with R² values of 0.9–1.00 and moderate error, though with slightly higher test-set error compared to CB. In contrast, ANN shows good but higher performance with R² of 0.992 for the test dataset, while the LSTM models are the least in this context, with the lowest test R² (0.856) and the highest errors, indicating a poorer fit and limited generalization for the regression tasks. Therefore, CB emerges as the most robust model.

Fig. 6. Best and least performing regression plots for the ML models (a) CatBoosting, (b) long-short-term memory.

While the dataset is primarily tabular, the reaction time variable introduces a temporal dimension to RB5 degradation. By grouping observations with identical physicochemical parameters, the degradation process was structured into discrete sequences, allowing LSTM models to attempt to capture kinetic dependencies and nonlinear patterns. However, as this sequential representation is only an approximation, the LSTM underperformed relative to conventional ANNs and tree-based ensembles. Consequently, the LSTM's role in this study is exploratory, whereas models like CatBoost proved more effective and fundamentally suited for the dataset's tabular nature.

3.3.1.4. The 5-k-fold cross-validation performance analysis for each model

A 5-fold cross-validation through the hyperparameter tuning using seven statistical metrics or loss function-based performance metrics indicators (R², RMSE, MAPE, MedAE, SI, LogCosh) was used to elucidate the predictive performance accuracy of the 10 ML models. Fig. 7 shows that cross-validation of CB models exhibited excellent goodness of fit, suggesting marginally better stability and minimization, through its test errors. In addition, LGBM, Bagging, HGB, GBDT, CNN and RF also demonstrated good R² values, but with slightly higher metric errors. ANN models performed moderately in their predictions, while LSTM showed extremely high error metrics performance. Each metric indicator's performance is well observed for each model as seen.

Fig. 7. Cross validation comparative metrics plots for the 10 ML model (a) CatBoosting, (b) random forest, (c) bagging, (d) hist gradient boosting, (e) ligh gradient boost machine, (f) gradient boost decision tree, (g) Gaussian process regression, (h) artificial neural network, (i) long-short-term memory, (j) convolution.

3.3.1.5. Predictive strength performance for best and worst models

Predictive strength performance is presented in Fig. 8. The predictive strength for CB shows a consistently close alignment between actual and predictive values across the entire test set, with only small pointwise deviations, indicating strong generalization and high predictive reliability even as the CB values fluctuate over a wide range (Fig. 8a). On the flip side, LSTM model shows visibly larger discrepancies between the actual and predicted series, including several pronounced over- and underestimations, which suggests that it struggles to fully capture the underlying data dynamics and is more sensitive to local variability (Fig. 8b). These plots demonstrate that the CB model provides a more accurate and stable representation of the target response, whereas the LSTM model, although able to follow the overall trend, delivers comparatively weaker predictive strength and higher residual variability, justifying the selection of CB as the preferred predictive framework in this study.

Fig. 8. Predictive strength performance for best (a) CB model and worst model (b) long-short-term memory model.

3.3.1.6. SHAP features importance and impact analysis

The Shapley Additive Explanation (SHAP) features importance analysis, demonstrating that GTC is the most influential predictor, followed by reaction time. The IDC is moderately positive, and pH can either enhance or reduce prediction depending on its value as ranked by mean absolute SHAP values, highlighting their relative contributions to models' performance (Fig. 9a). The SHAP summary features further show that higher GTC and reaction time levels strongly increase the predicted response, whereas higher IDC and pH play a moderate role or generally reduce it at high concentration and high pH (Fig. 9b). This directly aligns with correlation analysis in (Fig. 5a). Together, these results elucidate both the relative importance and directional effect of the input parameters on the model predictions and output.

Fig. 9. (a) SHAP features importance and (b) SHAP features impact.

3.3.1.7. The radar multi-metric and King-Gupta efficiency (KGE) analysis

The multi-metric radar plot enables a concise comparison of the 10 models' studies across multiple goodness-of-fit and error criteria, where optimal performance corresponds to high R² and minimal error values. Among all models, CB demonstrates the strongest overall performance, forming the largest radar envelope with near-unity R² and low errors, followed by CNN, LGBM, HGB, RF, GBDT, bagging, and GPR with slightly lower performance metrics compared to CB (Fig. 10a). KGE complementarily highlights CB as the most balanced model, achieving a perfect KGE value of 0.98 (α and β), indicating an optimal representation of correlation, variability, and boas, whereas LSTM exhibits variability and a low KGE of 0.92 (α and β) (Fig. 10b). Collectively, this visualization indicates that CB in conventional accuracy metrics CB offers the most physically consistent and well-balanced predictive behavior according to KGE.

Fig. 10. Showing multi-metric radar plot (a) and KGE analysis for best and worst model (b).

3.3.1.8. Taylor plot and empirical cumulative distribution function (EDF) analysis

The ECDF analysis provides a non-parametric assessment of prediction error distributions, showing that models with left-shifted curves yield a higher proportion of smaller errors. In this comparison, the CB model demonstrates the best performance, as its ECDF rises most rapidly and reaches a cumulative probability of 1.0 at the lowest error magnitude, indicating consistently minimal errors. This is closely followed by other models and lastly LSTM exhibiting broader error distributions (Fig. 11a). Complementarily, the Taylor diagram, which jointly evaluates correlation, centered RSME difference, and standard deviation, places the CB model closest to the reference point, confirming it superior correlation, minimal RSME and best agreement with observed variability, thereby establishing CB as the most accurate and statistically robust model while LSTM though close to the point but has high variability and RSME error (Fig. 11b).

Fig. 11. (a) Empirical cumulative distribution function and (b) Taylor plot analysis.

3.3.1.9. Partial dependence plots (PDPs) analysis

The PDPs were used to illustrate how the photodegradation of RB5 dye varies with respect to different input parameters, thereby providing insight to the model's behavior with respect to the single feature changes. The PDPs analysis reveals a pronounced nonlinear effect of all input variables on the model response. IDC (Fig. 12a) exhibits a clear optimum, with predictions rising sharply from 10–20 mg L⁻¹ and then declining beyond this point, indicating a critical sensitivity window between an optimum at 20 mg L⁻¹. The GT loading onto concrete (Fig. 12b) shows a strong threshold behavior, where increasing the dose up to 2 wt% yielded substantial removal and a continuous simultaneous increase to 10 wt%. The reaction time shows a gradual yet consistent rise as the removal of the RB5 process is continuous, with extended reaction times positively impacting the pollutant efficiency for up to 80 minutes of reaction time (Fig. 12c). The pH levels exhibit a more complex, nonlinear response. There are favorable conditions for RB5 removal in a stronger acidic environment and around pH 7, but efficiency declines in alkaline conditions, underscoring the negative impact that high pH levels have on RB degradation (Fig. 12d).

Fig. 12. PDPs showing the effect of (a) IDC, (b) GTC dose, (c) reaction time, and (d) pH.

3.4. Multiscale physicochemical characterization and mechanistic correlation with photocatalyst efficacy

These characterization techniques, when combined, provide a comprehensive understanding of the morphology, crystallinity, elemental composition, dispersion stability, and surface charge properties of photocatalysts and their embedded concrete structures, indicating a direct relationship between material structure and photocatalytic efficiency.

3.4.1. Morphology and phase structure

Scanning electron microscopy (SEM) was employed to examine the surface features of bare TiO₂ and GT nanocomposites. As illustrated in Fig. 13a₁, pure TiO₂ consists of predominantly spherical nanoparticles that are uniformly distributed but aggregated. This clustering is typical for anatase TiO₂, resulting from high surface energy and the natural tendency of particles to minimize free energy through interaction. In contrast, the GT nanocomposite (Fig. 13a₂) exhibits a modified surface where TiO₂ particles are more interconnected and better dispersed. The addition of GQDs appears to suppress aggregation, leading to a rougher and more heterogeneous surface texture. The successful integration of GQDs is further confirmed by the EDS spectrum (Fig. 13c), which identifies carbon, titanium, and oxygen as the primary elemental components of the hybrid material.

Fig. 13. Photocatalysts characterization: (a-1) SEM for TiO₂ (a-2), SEM for GT-3 (a-3), EDS GT-3, (b-1 to b-3) HRTEM images, (c) TEM particle size distribution (nm).

The structural properties and dimensions of the GT nanocomposite were analyzed using high-resolution transmission electron microscopy (HRTEM) and selected area electron diffraction (SAED). TEM images (Fig. 13-b1 and 13-b2) show that the material is composed of spherical to slightly elongated nanoparticles forming clusters. The SAED pattern (Fig. 13-b3) displays concentric diffraction rings with bright spots, identifying the polycrystalline nature of anatase TiO₂.²⁸ These rings correspond to the (101), (103), (200), (105), and (213) planes, matching JCPDS card No. 00-004-0477 and correlating with XRD findings (Fig. 13d). HRTEM lattice fringes further confirm the anatase phase, while faint diffraction spots related to the (002) planes of GQDs verify the presence of GQD domains. The simultaneous appearance of TiO₂ and GQD diffraction features confirms the formation of a high-purity crystalline GT heterostructure. Particle size measurements (Fig. 13c) reveal a uniform distribution between 8 and 40 nm, supporting the incorporation of GQDs into the TiO₂ framework.^21,29

To determine TiO₂ polymorphs and investigate the effects of GQD inclusion on crystallinity, phase stability, and lattice structure, diffraction patterns, the structural integrity and phase compatibility of TiO₂-loaded concrete (BTC) and GT-loaded concrete (GTC) were examined using X-ray diffraction (XRD), verifying that the photocatalyst immobilization did not negatively affect the cementitious matrix or generate undesired secondary phases. XRD patterns of bare TiO₂ and GQD-modified TiO₂ calcined at 300–500 °C with lattice parameters a (3.7 Å) b (3.7 Å), and c (9.4 Å), belonging to the tetragonal system show consistent anatase-phase peaks around 2θ = 25.3, 38.1, 48.2, 54.2, 62.7° are assign to (101, 103, 200, 105 and 213) belong to cystal planes of anatase TiO₂ phase (JCPDS N°. 21-1272) with slight shifts in 2θ values and peaks broadening in GT samples, indicating lattice distortion and reduced crystallite size due to GQD doping Fig. 14a). For concrete composites, BTC exhibits characteristic TiO₂ diffraction peaks. In contrast, GTC samples (incorporated with pre-calcined GT NCs into the concrete and labelled as GTC-3, GTC-4, GTC-5) with lattice parameters a (4.9 Å), b (4.9 Å), and c (17.0 Å) in the Rhombohedral system show additional peaks at 2θ = 23.0, 29.5, 36.0, 39.5, 43.2, 47.1, 47.6 63.9 and 68.5°, corresponding to crystal phases (012, 104, 110, 113, 202, 024, 018, 300, 210 suggesting enhanced crystallinity and possible formation of new phases or improved dispersion of GQD/TiO₂ within the cementitious matrix (Fig. 14b). The appearance of these peaks is explained by the characteristic system of CaCO₃ rhombohedral trigonal cell system as a reference for XRD phase indexing and phase confirmation, synergistic interaction of the GQD-modified catalyst and the concrete composition environment, in which GQDs behave as effective nucleation sites, promoting the crystallization of hydration products. Thus, a sharp peak at 29.5° is associated with calcite (CaCO₃), which has dual functions in the concrete-based samples: first, as the main mineral constituent of the recycled concrete powder (obtained from recycled concrete aggregates) utilized as a carrier, and second, as a byproduct of cementitious matrix densification. The roles of CaCO₃ include acting as a carrier phase, densification agent, and dispersion stabilizer.³⁰ The presence of distinct calcite peaks indicates that GQDs promote a more crystalline and refined microstructure, which may improve the mechanical integrity and carbonation resistance of the photocatalytic concrete layer in addition to improving the dispersion of the GQD/TiO₂ phase and preventing bulk material from masking it.^21,31 The elevated calcite (CaCO₃) peaks are ascribed to the role of GQDs as nucleation sites that promote the crystallization and densification of cement hydration products. While this improves microstructural characteristics, the very alkaline and calcium-rich environments of cementitious systems may have an impact on the implanted GT nanocomposite's long-term durability. Prolonged contact with high pH levels can cause surface hydroxylation of TiO₂ and partial oxidation/functionalization of GQDs. However, the strong Ti–O–C interfacial bonding, as validated by XPS investigation, provides robust chemical coupling that contributes to the composite's structural integrity and photocatalytic activity. Additionally, the cement matrix acts as a protective medium, decreasing nanoparticle leaching and increasing stability.³² The observed operational stability of GTC, with 89.7% degradation efficiency in the first cycle and 80.65% after six reuse cycles, suggests a progressive performance reduction and strong structural durability under repeated photocatalytic operation. These findings indicate good short- to medium-term durability; however, more studies incorporating accelerated aging and real-world environmental exposure are needed to thoroughly evaluate long-term performance.

Fig. 14. (a) XRD for TiO₂ and GT (b), and XRD for TiO₂-loaded concrete (BTC) and GT-loaded concrete (GTC).

3.4.2. Colloidal stability, surface charge carrier, and interfacial electrostatics

This section examines growth in hydrodynamic size during calcination, trends in zeta potential, and the effects of heat on dispersion stability (as determined by a correlogram). It also examines the point of zero charge (PZC) at approximately 6.36 and its implications for RB5 adsorption, as well as the electrostatic interaction framework. Dynamic light scattering (DLS) and zeta potential measurements were used to investigate the colloidal stability and surface electrokinetic behavior of TiO₂ and GT photocatalysts (Fig. 15). To investigate dispersion behavior in aqueous media, hydrodynamic particle size distribution was measured, and zeta potential analysis revealed surface charge features and suspension stability under experimental pH values relevant to photocatalytic degradation studies. These factors are essential for understanding catalyst–pollutant interactions, aggregation patterns, and effective surface availability in photocatalytic processes.

Fig. 15. Colloidal behavior, (a) size distribution, (b) zeta potential, and (c and d) hydrodynamic size.

The dynamic light scattering analysis (DLS) shows that as the calcination temperature of GT increases from 300 °C, 400 °C, and 500 °C, the particle size distribution also expands. This suggests that higher temperatures lead to thermal-induced agglomeration and growth. TiO₂. Which keeps a narrow size profile, and the GQD-modified samples display a wider distribution, indicating that the addition of GQD affects thermal stability and particle coalescence (Fig. 15a). Zeta potential measurements show a gradual shift from positive to nearly neutral values as the calcination temperatures increase, with GT calcined at 300 °C (GT-3) showing the highest surface charge at +20 mV, while GT calcined at 500 °C (GT-%) nears electrostatic instability at increased temperatures, suggesting a decrease in colloidal stability and elevated temperature. Fig. 15b shows that GT-3 has the highest positive surface charge (∼+20 mV), indicating improved colloidal stability. Higher temperatures decreased surface charge and colloidal stability, possibly due to loss of surface functional groups. Correlogram analysis further backs these observations. TiO₂ shows a quick drop in the correlation coefficient, stabilizing around zero after 10² µs (Fig. 15c), whereas GT-3 maintains a higher initial correlation of 0.8 before it starts to decay, indicating improved dispersion and dynamics stability to GQD functionalization (Fig. 15d).

To better understand the electrostatic interactions between the catalyst surface and dye molecules, surface charge properties were examined using point of zero charge (PZC) measurements. The PZC of the GT combination was found to be around pH 6.36 (Fig. 16a). Below this pH, the catalyst surface is positively charged, facilitating electrostatic interaction with anionic RB5 dye molecules and increasing adsorption and degradation efficiency. The GQD/TiO₂ surface is positively charged below this pH, encouraging electrostatic interaction with anionic RB5 dye molecules and improving adsorption and degradation. This pH-dependent performance is consistent with prior TiO₂-based photocatalysis research.

Fig. 16. (a) Surface PZC, and (b) PL correlation with photoactivity of GT.

Photoluminescence (PL) spectroscopy was used to assess the charge carrier recombination behavior of GT photocatalysts calcined at 300 °C (1GT-3), 400 °C (1GT-4), and 500 °C (1GT-5). All TiO₂-based photocatalysts display a broad emission band in the visible region (≈350–550 nm), which is typically attributed to radiative recombination of photogenerated electron–hole pairs and defect-related surface states (see Fig. 16b). Defect engineering corresponds with recent discoveries where the synergistic effect of oxygen vacancies and piezoelectric polarization significantly improves the photocatalytic reduction performance of Cr(vi) in TiO₂-based materials.³³ 1GT-5 had the highest PL intensity among samples, indicating rapid charge carrier recombination due to low crystallinity and lower electronic coupling between GQDs and TiO₂ lattices. Increasing calcination temperature to 400 °C (1GT-4) reduces PL intensity, indicating enhanced crystalline and stronger interfacial contact between GQDs and TiO₂. This enables electron transport from the TiO₂ conduction band to GQDs acting as electron sinks. 1GT-3 (300 °C) has the lowest PL intensity, indicating the most effective suppression of electron–hole recombination. Improved TiO₂ crystallinity and ideal interfacial charge transport routes brought about by increased heat treatment are responsible for this characteristic. However, extremely high calcination temperatures may cause partial GQD graphitization or aggregation, explaining why 1GT-3, despite having the lowest PL intensity, may not necessarily exceed 1GT-4 in photocatalytic efficiency. In GQD-modified and graphene-based TiO₂ systems, photocatalytic activity is maximized by balancing crystallinity, defect density, and carbon-semiconductor interfacial interaction. PL results substantially correlate with photocatalytic performance trends, suggesting that suppressed recombination and increased charge carrier lifespan are primary contributors to higher visible-light activity in GT systems.

Integrating PL and HPLC data yields a clear mechanistic picture: GQDs allow rapid interfacial electron transfer, extending charge carrier lifetimes and boosting ROS formation, which causes successive azo bond breaking and aromatic oxidation. Radical scavenger studies reveal that ˙OH and ˙O₂⁻ radicals are the principal active species, with minor roles for direct hole oxidation and adsorption. The absence of persistent intermediates indicates continual oxidative transformation, emphasizing the effectiveness of GQD/TiO₂ for visible-light-driven dye mineralization. GQD incorporation improves charge separation and accelerates photocatalytic degradation, making it a highly effective visible light photocatalyst for sustainable wastewater treatment. This is supported by low PL intensity, enhanced ROS formation, and HPLC-confirmed molecular fragmentation.

3.4.3. Surface chemistry and interfacial bonding

The surface chemistry, electrical structure, and interfacial interactions of 1GT-3 were studied using XPS and FTIR to determine the origin of its improved visible-light photocatalytic activity. The XPS survey spectrum (Fig. 17a) reveals the presence of Ti, O, C, and trace N elements, indicating that GQDs were successfully immobilized on the TiO₂ surface with no visible impurity phases. The absence of extraneous peaks suggests great material purity and a well-defined heterointerface. The Ti 2p spectra (Fig. 17b) show two distinct peaks at ∼458.5 eV (Ti 2p_3/2) and ∼464.2 eV (Ti 2p_1/2), which correspond to Ti⁴⁺ in anatase TiO₂. A small decrease in binding energy compared to pristine TiO₂ shows electronic interaction between TiO₂ and the π-conjugated GQDs network, indicating charge redistribution and electron density transfer.²⁶

Fig. 17. XPS analysis of (a) GQDs/TiO₂ spectra of (b) Ti 2p, (c) C 1s, (d) O 1s and (e) FTIR spectra.

The C 1s spectra (Fig. 17c) confirm the inclusion of GQDs, with deconvoluted peaks at ∼284.6 eV (sp² C–C/C C), ∼286.0 eV (C–O/C–N), and ∼288.3 eV. The presence of oxygen-functionalized carbon species promotes the development of Ti–O–C interfacial connections, which increase visible light absorption and allow for efficient charge delocalization across the carbon-semiconductor heterojunction. These interfacial connections are essential for stabilizing photogenerated electrons and promoting the production of ROS under visible light irradiation.³⁴ In the O 1s spectrum (Fig. 17d), at a binding energy of ∼529.6 eV, Ti–O lattice oxygen was identified, corresponding to the presence of Ti–O–Ti bonds characteristic of anatase TiO₂.¹⁴ The XPS results show the creation of a highly linked GQDs/TiO₂ heterointerface rich in oxygen vacancies, surface hydroxyl groups, and heteroatom-functionalized carbon species. The GQDs/TiO₂ nanocomposite outperforms pristine TiO₂ systems in terms of visible-light photocatalytic performance due to enhanced light harvesting, charge separation, and reactive oxygen species production.

The XPS results confirm a chemically bound Ti–O–C heterointerface between GQDs and TiO₂, which plays a significant role in interfacial charge transfer.³⁵ The observed negative shift in Ti 2p binding energy suggests electron transport from π-conjugated GQDs to the TiO₂ conduction band. The deconvoluted C 1s spectra also suggests the presence of oxygen-functionalized carbon molecules that aid in the production of covalent Ti–O–C bonds. Interfacial bonds facilitate electron transport, allowing for direct electronic interaction between GQDs and TiO₂ and fast charge delocalization. The heterojunction between GQDs and TiO₂ leads to band alignment and work function discrepancies, creating an electric field that separates photogenerated charge carriers.³⁶ DFT investigations on similar GQD/TiO₂ systems demonstrate that C–O–Ti connections produced by oxygen vacancies and surface functional groups considerably boost visible light absorption and promote electron transport from GQDs to TiO₂ while reducing recombination.^36,37 Furthermore, the presence of oxygen vacancies and surface hydroxyl groups enriches the interfacial environment by creating active sites for charge trapping and reactive oxygen species generation. The Ti–O–C heterointerface has two functions: (i) expanding visible-light harvesting through π-conjugated GQDs, and (ii) enhancing efficient charge separation and transport by interfacial electron transfer and built-in electric field effects.²⁶ This synergistic process is compatible with the improved photocatalytic activity and ROS-mediated degradation pathways found in the current investigation.

Complementary evidence of 1GT-3 interfacial alteration is provided by FTIR spectra (Fig. 17e). The broad absorption band in the 500–700 cm⁻¹ area is ascribed to Ti–O–Ti lattice vibrations, whereas sp² C–C stretching from graphitic domains of GQDs is responsible for the formation of a separate band at ∼1630 cm⁻¹. The broad band between 3000 and 3500 cm⁻¹ is attributed to O–H stretching vibrations, indicating improved surface hydroxylation and hydrophilicity. The lack of novel Ti–O related vibrational modes shows that GQDs are integrated predominantly by surface anchoring and heterojunction creation rather than bulk lattice substitution, which allows for interfacial charge separation without disrupting TiO₂ crystallinity.

The synergistic surface chemistry of GT resulted in remarkable photocatalytic activity, with 98.7% (1GT-3) and 89.7% (1GTC-3) RB5 decolorization achieved after 80 minutes of visible-light irradiation, as well as substantial mineralization (COD). HPLC analysis and radical quenching experiments show mechanistic specificity, including suppressed azo bond cleavage and incomplete aromatic ring oxidation under ˙OH and ˙O₂⁻ scavenging. The primary cause of azo bond scission is ˙OH radicals, but the subsequent oxidative fragmentation and deep mineralization of aromatic intermediates are largely attributed to ˙O₂⁻ radicals. The GT heterointerface, enriched with Ti–O–C linkages, surface hydroxyl groups, and oxygen vacancies, offers effective visible light harvesting, quick charge separation, and a cooperative ˙OH/˙O₂⁻ radical route (PL-XPS-FTIR-HPLC correlation). This integrated surface chemistry outperforms carbon-modified TiO₂ systems in terms of both dye decolorization and COD removal.

The UV-vis diffuse reflectance spectra (DRS) of 1GT-3 reveal a significant increase in light absorption across the visible region compared to pristine TiO₂, with a distinct red shift of the absorption edge. 1GT-3 exhibits prolonged absorption well into the visible range (Fig. 18a), demonstrating good optical sensitization by GQDs. In contrast, bare TiO₂ typically absorbs in the ultraviolet area due to its broad inherent bandgap. With an estimated optical bandgap of E_g = 2.7 eV, the accompanying Tauc plot (αhν)^1/2versus photon energy (Fig. 18b), reveals a considerable bandgap narrowing for 1GT-3. GQDs introduce π-conjugated carbon states and nitrogen-related electronic levels, creating intermediate energy states between TiO₂'s valence and conduction bands, leading to a reduced bandgap. As a result, lower-energy visible photons can efficiently excite electrons, allowing for photocatalytic activity when exposed to visible light. 1GT-3's increased visible-light absorption and bandgap narrowing correspond directly to its higher photocatalytic efficacy. The lower energy excitation allows for increased charge-carrier production rates, while the intimate Ti–O–C interfacial coupling promotes quick electron transfer from TiO₂ to GQDs, reducing recombination, as indicated by PL quenching data. As a result, 1GT-3's optical characteristics provide the foundation of its improved reaction kinetics and radical-driven degrading efficiency.

Fig. 18. DRS analysis of GQDs/TiO₂ (a) absorbance spectra (b) Tauc plot.

3.5. Comparative discussion of experimental, mechanistic, ML performance with previous studies

This study integrates physicochemical, scavenger, kinetic, and HPLC/LCMS investigations to understand the visible-light-driven photocatalytic process of the GT or GTC system. When illuminated, GQDs act as electron reservoirs and photosensitizers, boosting electron transfer from TiO₂ and inhibiting electron–hole recombination. Surface hydroxyl groups and oxygen vacancies encourage the creation of ˙OH, while the accumulated electrons convert dissolved oxygen to produce ˙O₂⁻ radicals. These reactive oxygen species work together to cleave azo bonds, open aromatic rings, and gradually oxidize RB5, eventually resulting in mineralization into CO₂ and H₂O. A comparison with previous studies shows that the current work advances photocatalytic concrete research by directly addressing long-standing challenges such as catalyst recovery, durability, and scalability, all of which have been identified as critical limitations in suspension-based photocatalytic systems.^38,39 While most reported photocatalytic concretes have been developed for self-cleaning surfaces and air purification, their use in aqueous wastewater treatment has been limited due to mass-transfer resistance, light attenuation within cementitious matrices, and low mineralization efficiency. This study's GQD-modified TiO₂ concrete achieves significant RB5 discoloration and COD reduction under visible light, overcoming the intrinsic limitations of immobilized photocatalytic systems.

Compared to previous photocatalytic concrete experiments, this work solves fundamental constraints in catalyst recovery, durability, and scalability that still exist in suspension-based systems. While most reported photocatalytic concretes are intended for self-cleaning or air purification, their use in wastewater treatment has been limited due to mass-transfer resistance, light attenuation inside cementitious matrices, and low mineralization efficiency.⁵ Incorporating GT into a concrete matrix provides structural stability while preserving strong Ti–O–C interfacial interactions, unlike conventional carbon-modified TiO₂ systems, where enhanced light absorption and charge separation are often offset by catalyst leaching and low reusability.⁴⁰ XPS data and strong PL quenching confirm suppressed charge recombination, allowing for persistent ROS formation over several cycles. The activity–stability trade-off that frequently restricts suspended photocatalysts is successfully resolved by this immobilization technique.⁴¹

This study distinguishes between genuine photocatalytic oxidation and adsorption-dominated color removal. The combination of COD reduction, HPLC-resolved degradation pathways, and radical scavenging studies indicates real oxidative mineralization rather than surface chromophore damage. The dominant roles of ˙O₂⁻ and ˙OH radicals, identified through benzoquinone and isopropanol quenching, are consistent with the bandgap narrowing (E_g = 2.7 eV) observed in DRS, and recombination suppression revealed by PL analysis, providing mechanistic coherence across optical, chemical, and kinetic observations.⁴² In cementitious systems, carbonation and CaCO₃ creation are frequently regarded as durability issues. However, it was discovered that the GT concrete's regulated CaCO₃ development improved photocatalytic stability by enhancing microstructural densification, catalyst dispersion, and alkalinity regulation. This synergistic effect, while rarely investigated in wastewater photocatalysis, is consistent with developing notions in multifunctional photocatalytic building materials.⁴³

Finally, previous research on photocatalytic concrete relied on one-factor-at-a-time experimentation or simplified kinetic descriptions such as pseudo-first-order or Langmuir–Hinshelwood models, which are insufficient to capture strongly nonlinear interactions between pH, catalyst loading, pollutant concentration, and irradiation time. To advance photocatalytic concrete research, this study integrates machine learning (ML) with mechanistic experiments. Traditional kinetic models and one-factor-at-a-time techniques fail to describe the nonlinear dynamics of immobilized photocatalytic systems. In order to attain almost perfect predicted accuracy while preserving physical interpretability, explainability techniques (SHAP and PDP) were combined with sophisticated machine learning models. This integrated experimental-mechanistic-ML strategy results in a scalable and rational way for optimizing photocatalytic concrete design, which is a significant step toward practical wastewater treatment applications.^44,45 To ensure methodological rigor, each ML model was provided with a categorical identification exclusively for performance comparison, rather than as an input feature. The coefficient of determination (R²) for the train and test datasets is displayed with each model represented by a column, to make performance comparison easier. No further encoding was applied, as these identifiers serve simply to mark the associated performance metrics and are not used as input features during model training. This integrated experimental-mechanistic-ML framework applies new ML-driven photocatalysis techniques to immobilized cementitious systems, providing a scalable avenue for rational design and optimization of photocatalytic concrete for wastewater treatment.

In addition to performance benchmarking, the comparative analysis was extended to evaluate material design strategies and the depth of mechanistic investigation (Table 4). Unlike many previous studies that primarily report degradation efficiency and catalyst reusability, the present work adopts an integrated framework combining advanced characterization and data-driven modelling to establish a comprehensive structure–performance–mechanism relationship. The GQDs-modified TiO₂ system was rationally engineered to enhance visible-light activity through Ti–O–C interfacial coupling with Graphene Quantum Dots, enabling improved charge separation and suppressed recombination. Mechanistic insights were derived from complementary techniques, including XPS for interfacial bonding analysis, photoluminescence spectroscopy for charge carrier dynamics, and chromatographic tracking of degradation intermediates of RB5. In parallel, ML-based predictive modelling was employed to quantitatively link operational parameters with photocatalytic performance, while interpretability tools provided insight into parameter influence and system behavior. This multi-level integration offers a more rigorous and holistic understanding compared to conventional studies, thereby advancing the rational design and optimization of photocatalytic systems for environmental applications.

Table 4. Comparative evaluation of carbon-doped TiO₂ for organic pollutant degradation^a.

Study	Synthesis method	Key techniques	Degradation efficiency (%)	Cyclic stability	Mechanistic depth	References
GT/GTC	SG	XRD TEM, XPS, PL, FTIR, DRS, ML	98.7 (RB5)	6th	Comprehensive performance: structure (XPS/PL) – performance (HPLC/LCMS/ML) correlation	This work
S,N-GQDs	HT	TEM, XRD, FTIR, Raman, DRS	H2 generation	3rd	Limited: carrier at interface, no ML/HPLC	46
CQD/TiO2	HT	XRD, SEM, TEM, XPS Raman, DRS	82.6 (MB)	4th	Light harvesting/recombination reduction, no ML/HPLC	47
CQD/TiO2	HT	TEM, PL, BET, DRS	99 (MO)	Not reported	Enhance charge separation, limited probes, no ML/HPLC	48
F-TiO2@N-GQDs	SG	XRD, FTIR, PL, XPS, SEM, EDS, DRS	90 (MB)	5th	Electron reservoir, reducing recombination, no ML/HPLC	10
GO/CNC/TiO2 QDs	Co-P	Sem TEM EDS, XRD, FTIR, DRS	95.5 (MB, MV, CF) low for antibacterial	Not reported	ROS generation, no ML/HPLC	49
TQDs/CC	HT	SEM, EDS, DRS	>90 (MO, RhB)	10th	Charge separation improvement	50
TiO2	—	Non	67 (MB)	Not reported	Limited	51
TiO2	—	SEM XRD	60 (TVOCs)	Not reported	Limited	52

^aSG = sol–gel; HT = hydrothermal; Co-P = Co-precipitation.

3.6. Phytotoxicity assessment

The phytotoxicity of untreated RB5 dye and treated wastewater was evaluated using a mung bean (Vigna radiata) seed germination bioassay. Germination percentage and root length were recorded after 4 days of exposure, as these parameters are reliable and sensitive indicators of phytotoxicity.⁵³ The results are illustrated in Fig. 1.

3.6.1. Germination percentage

As shown in Fig. 19a, the negative control (distilled water) exhibited a germination rate of 80.0 ± 10.0%, confirming normal seed viability under the experimental conditions. The untreated RB5 dye significantly reduced germination to 6.7 ± 11.5%, representing an approximately 92% inhibition compared to the control. Similarly, the positive control (0.1% CuSO₄) showed a germination rate of 16.7 ± 15.3%, confirming the expected toxicity of copper sulfate, which is known to disrupt membrane integrity and inhibit seed metabolism even at low concentrations.⁵⁴ In contrast, the treated wastewater demonstrated a germination rate of 80.0 ± 10.0%, which was nearly identical to the negative control. This finding strongly suggests that the treatment process effectively removed or degraded the phytotoxic components responsible for germination inhibition. The severe inhibition of germination by untreated RB5 dye is consistent with previous reports on the phytotoxicity of synthetic azo dyes. Azo dyes such as RB5 contain aromatic amine groups and sulfonate moieties that can interfere with water uptake across the seed coat and inhibit critical enzymes involved in the mobilization of stored reserves during germination.^55,56 The complex aromatic structure of RB5 dye is known to induce oxidative stress through the generation of reactive oxygen species (ROS) during its degradation, leading to cellular damage in exposed organisms.⁵⁷

Fig. 19. Phytotoxicity assessment of untreated RB5 dye and treated wastewater on mung bean (Vigna radiata) after 4 days of exposure. (a) Germination percentage, (b) mean root length (mm).

3.6.2. Root elongation

Root length is widely considered a more sensitive endpoint than germination percentage in phytotoxicity studies, as sub-lethal effects often manifest as growth inhibition even when germination occurs.⁵⁸ As shown in Fig. 19b, the negative control produced a mean root length of 0.401 ± 0.087 mm. Exposure to untreated RB5 dye resulted in severe root growth inhibition, with a mean root length of only 0.050 ± 0.087 mm, representing an 87.5% reduction compared to the control. The positive control (CuSO₄) exhibited a mean root length of 0.106 ± 0.092 mm, corresponding to 73.6% inhibition. The slightly higher root length in CuSO₄ compared to RB5, though not statistically significant, may reflect differences in the mechanism of toxicity between the heavy metal and the organic dye. Moreover, the treated wastewater yielded a mean root length of 0.395 ± 0.069 mm, which was statistically comparable to the negative control. This complete recovery of root growth indicates that the treatment process not only eliminated the lethal effects of the dye (allowing germination) but also removed sub-lethal inhibitors that would otherwise impair root elongation. Root elongation is particularly sensitive to chemical stress because root tips are sites of active cell division and elongation. Any interference with cell wall biosynthesis, microtubule formation, or hormone signaling can rapidly manifest as reduced root length.⁵⁹ The dramatic reduction in root length observed with untreated RB5 dye suggests that the dye or its components directly affected these processes. The complete absence of such effects in the treated wastewater confirms the effectiveness of the treatment.

One-way ANOVA revealed a statistically significant effect of treatment on both germination percentage and root length (Table 5). For germination, ANOVA yielded F(3, 8) = 34.82, p < 0.001, while for root length, F(3, 8) = 14.39, p = 0.0014. These results confirm that the observed differences among treatments were not due to random variation.

Table 5. One-way ANOVA summary for root length (mm) of Vigna radiata exposed to different treatments.

Source of variation	Sum of squares (SS)	Degrees of freedom (df)	Mean square (MS)	F-value	p-value
Between groups	0.31216	3	0.10405	14.39	0.0014
Within groups	0.05784	8	0.00723
Total	0.37000	11

Most importantly, the treated wastewater showed no significant difference from the negative control for both germination (p > 0.05) and root length (mean difference = 0.006 mm, p > 0.05). This statistical finding provides robust evidence that the treatment process successfully eliminated phytotoxicity from the RB5 dye solution. The complete statistical comparison is summarized by the significance letters (a and b) in Fig. 1, where shared letters indicate no significant difference.

To thoroughly assess the environmental safety of photocatalytically treated effluent, phytotoxicity experiments were conducted. A mung bean (Vigna radiata) seed germination bioassay was also used to measure phytotoxicity, and untreated RB5 showed severe inhibition of germination and root extension, confirming the dye's substantial phytotoxic effects. Treated wastewater significantly improved plant development characteristics, with germination and root length comparable to the negative control. These findings demonstrate that the GT photocatalytic system effectively prevented the acute phytotoxicity associated with RB5. The detoxification observed is due to the creation of a Ti–O–C interfacial structure between TiO₂ and Graphene Quantum Dots. This increases charge separation and promotes the synthesis of reactive oxygen species (˙OH and O₂˙⁻). These species aid in the breakdown of RB5 through azo bond cleavage, followed by the gradual oxidation of aromatic intermediates, which eventually leads to mineralization. The lack of inhibitory effects suggests that no persistent or hazardous byproducts were produced throughout the degradation process, validating the treated effluent's environmental safety. Furthermore, incorporating standardized bioassays like Vibrio fischeri would enhance environmental risk assessments. Beyond simple pollutant removal, this platform enables multifunctional applications such as simultaneous degradation and the synthesis of value-added chemicals, as seen in emerging systems like Ag/In₂S₃/BiOBr,⁶⁰ signaling a promising path for the development of sustainable, next-generation photocatalytic materials.

4. Conclusion

This study successfully developed and assessed a visible-light-activated GQDs/TiO₂-based photocatalytic concrete (GTC) system for sustainable wastewater treatment, using an integrated experimental and machine learning framework. The GQDs-modified TiO₂ demonstrated near-complete degradation of Reactive Black 5 (≈100%) in slurry systems and maintained high efficiency after immobilization within the concrete matrix (100% removal in 90 min and 89.7% in 80 min), confirming the feasibility of preserving catalytic performance in cementitious environments. Kinetic study showed that degradation followed the Langmuir–Hinshelwood pseudo-first-order model (R² = 0.9461), indicating surface reaction-controlled behavior. The improved activity was attributed to the formation of a Ti–O–C interfacial linkage between TiO₂ and Graphene Quantum Dots. This improved visible-light absorption facilitated charge separation, and suppressed electron–hole recombination, promoting hydroxyl radical (˙OH)-dominated oxidation pathways. These reactions allowed for effective azo bond breaking and subsequent intermediate oxidation, leading to substantial mineralization. Importantly, phytotoxicity testing with Vigna radiata showed complete recovery of germination and root elongation in treated samples compared to significant inhibition by untreated dye. These results demonstrate that the photocatalytic approach achieved not just decolorization but also effective detoxification without producing toxic or long-lasting byproducts. Advanced machine learning models, such as CatBoost and convolutional neural networks, showed near-ideal predictive performance (R² = 1.00) with negligible generalization error. The use of explainable artificial intelligence tools such as SHAP analysis and partial dependence graphs ensured the mechanical interpretability of model predictions, combining experimental observations with data-driven optimization. This study develops a strong material-process-modeling paradigm for the rational design of photocatalytic concrete systems, providing a scalable and climate-resilient solution for wastewater remediation under ambient settings. Future research should concentrate on the long-term durability and regeneration of photocatalytic concrete under real-world settings, validation using real textile wastewater matrices, and comprehensive identification of degradation intermediates utilizing advanced analytical techniques (e.g., LC-MS).

Author contributions

Nfor Elvis Nfor: formal analysis, investigation and writing – original draft, Nadia Riaz: supervision, conceptualization, resources, funding acquisition, methodology, writing – review and editing, Muhammad Faisal Javed: supervision, software, validation, visualization, writing – review and editing, Muhammad Saqib Khan: writing – review and editing, analysis, validation, Rizwana Sarwar, Muhammad Bilal: writing – review and editing, validation. Ajmal Khan, Ahmed Al-Harrasi, and S. Tasqeeruddin: writing – review and editing, funding acquisition, validation.

Conflicts of interest

The authors declare that they have no competing financial or personal interests that could have influenced the work reported in this paper.

Data availability

The data supporting the findings of this study are available within the manuscript and its supplementary information (SI). Supplementary information is available. See DOI: https://doi.org/10.1039/d6ra01915c.