The dye wastewater used for analysis and characterization was obtained from a local textile wastewater treatment plant (Pocheon, South Korea). The main components of dye wastewater are summarized in Table 1. Mean values were obtained from at least two measurements. The hydrogen peroxide (≥ 34.5%, H₂O₂) and methanol (MeOH, ≥ 99.5%, CH₃OH) were procured from Samchun Chemicals (Pyeongtaek, South Korea). Sodium sulfite anhydrous (≥ 96%, Na₂SO₃) was purchased from Daejung Chemicals (Siheung, South Korea).

Prior to the experiments, the dye wastewater was subjected to pretreatment using a microfiber filter (1.2 μm, Whatman GF/C) to eliminate any suspended solid particles. The UV/H₂O₂ process was conducted in a black acrylic photoreactor [36, 37] and UVC lamp(s) were used as a light source (TUV 4W G4 T5, Philips, USA). The UVC intensities were measured as 15.12, 26.56, 50.78, 65.86 and 80.30 W/m² using UV radiometers. The UV/H₂O₂ process was carried out for 120 min (UVC lamp number = 2, H₂O₂ concentration = 5 mM). Here, we focused on the effect of three crucial factors on the efficiency of color removal. UVC intensity (5.12, 26.56, 50.78, 65.86 and 80.30 W/m²), H₂O₂ concentration (0.2, 0.5, 2, 5, and 8 mM), and reaction time (10, 60, 120, 180, and 240 min) were adjusted in this study. For the color removal experiments, 50 mL of samples containing moderate amounts of H₂O₂ were placed under the UVC lamps in a quartz beaker with steady stirring using a magnetic bar. After the experiments, to quench the residual H₂O₂, 0.5 mL of Na₂SO₃ (1.6 M) solution was instantly spiked into the samples. Additionally, there was no obvious change in pH after the reaction (± 0.1). The color change was analyzed using a UV-VIS spectrophotometer (DR 3900, HACH), and the TOC concentration of the samples was measured using a Total Organic Carbon Analyzer (TOC-V CSN, SHIMADZU). The hydroxyl radical (•OH) scavenging test was performed by adding 20 mM MeOH in the solution. Moreover, the optical properties of the samples were analyzed to confirm the degradation of color using a UV-VIS spectrophotometer (DR 3900, HACH) at a wavelength of 720 nm. All the results of one-factor-at-a-time method were represented as the mean of duplicate experiments and the error bars represent the range of the values.

MLP is a type of ANN that consists of multiple hidden layers, including an input layer, hidden layers, and an output layer with full connectivity between adjacent perceptrons [38, 39]. Each perceptron in the MLP applies an activation function to the input from the previous layer to obtain an output. These activation functions enable the MLP to compute complex and non-linear relationships in the data, making it suitable for solving non-linear problems. During training, the backpropagation algorithm is employed to adjust the weights and minimize the error between the predicted and actual outputs. Therefore, the MLP model can provide more accurate predictions by capturing non-linear relationships compared to other approaches. The experimental conditions were determined using a central composite design (Fig. 1), and data were collected for different conditions. The input variables and their ranges were selected as follows: H₂O₂ concentration of 2, 5, and 8 mM; UV intensity of 15.12, 26.56, and 50.78 W/m²; and reaction times of 60, 120, and 180 min, respectively. Detailed information about the MLP application is presented in Table S1–S3. To avoid the weighted fitting of MLP to the specific input variable which have relative bigger absolute values, each input variables are normalized (Coded values, Table S1). Separate MLP models were trained for each target variable, including color removal rate (−) and TOC removal rate (−). The experimental conditions derived from the Central Composite Design (CCD) represent optimal conditions for investigating the impact of individual factors and their interactions. Consequently, when training the MLP using this data, there is a risk of underfitting if only a subset of the data is employed. Conversely, utilizing all available data introduces the potential for overfitting. To strike a balance between these concerns, our study adopts a strategy aimed at minimizing overfitting while maximizing the utilization of the entire dataset as Fig. S1. The optimal hyperparameters for the MLP models were determined through a grid search with 5-fold cross-validation using all available data and the MLP models with the optimal hyperparameters were trained using the entire dataset. The relative importance of the input variables was examined using Shapley Additive explanations (SHAP) based on the trained MLP models and the entire dataset. Additionally, to evaluate the performance of the MLP model, additional experiments (see section 3.5) were conducted under the following conditions: H₂O₂ concentration of 2.30, 2.60, 3.80, and 8.00 mM; UV intensity of 15.12 and 26.56 W/m²; and reaction times of 66, 90, 96, and 114 min, respectively. All these processes were conducted using the Python programming language using the ‘sklearn’ and the ‘shap’ packages.

The color removal efficiency of the UV/H₂O₂ process was investigated under various H₂O₂ concentrations (Fig. 2a). As illustrated in the figure, the color removal efficiency increases with higher H₂O₂ concentrations. The removal efficiencies of 0.2, 0.5, 2, 5, and 8 mM of H₂O₂ were 0.00 ± 1.23)%, 14.22 ± 3.24)%, 32.58 ± 0.00)%, 91.60 ± 0.01)%, and 95.08 (± 0.02)%, respectively. In the UV/H₂O₂ process, H₂O₂ molecules absorb photons, which results in the generation of hydroxyl radicals (•OH) [ɛ = 18.7 M⁻¹ cm⁻¹ at 254 nm [40]] (Eq. (1)).

In general, the generated hydroxyl radicals (•OH) (oxidation potential: 2.8 eV) [41] have high reactivity towards color-causing organic compounds [42]. Therefore, higher initial H₂O₂ concentrations (5 and 8 mM) resulted in higher color removal efficiency rates. To confirm the contribution of •OH in the removal of the color, MeOH was used as a •OH radical scavenging agent (k_•OH/MeOH = 9.7 10⁸ M⁻¹s⁻¹) [43]. As illustrated in Fig. 2c, the addition of MeOH exactly hindered color removal reducing the removal ratio from 91.60 ± 0.01)% to 15.14 ± 1.95)%. Additionally, photolysis alone did not remove the color of textile wastewater, suggesting that •OH played a major role in the color removal during the UV/H₂O₂ process. The optical properties of the textile wastewater samples before and after reaction were investigated by UV-Vis spectrum (Fig. 2d). In contrast to the before reaction, there was a substantial decrease in light absorption within the 400 to 700 nm range after the UV/H₂O₂ reaction. This observation suggests a complete removal of color through the UV/H₂O₂ process. An increase in H₂O₂ concentration also significantly enhanced the TOC removal efficiency (Fig. 2b). The TOC removal efficiencies were 0.00 ± 0.44)%, 1.03 ± 0.36)%, 0.00 ± 0.60)%, 2.92 ± 0.06)%, and 32.29 ± 0.01)% at H₂O₂ concentrations of 0.2, 0.5, 2, 5, and 8 mM, respectively. In contrast to the result of the color removal experiment, a noteworthy difference in TOC removal efficiency was observed between H₂O₂ concentrations of 5 and 8 mM. TOC concentrations were decreased only at 8 mM of H₂O₂ even though the color removal efficiencies of 5 and 8 mM H₂O₂ were similar. This indicates that the generation of •OH was not sufficient to further mineralize the organic compounds with 5 mM of H₂O₂ (2.92%). On the other hand, the higher H₂O₂ concentration (8 mM) generated more •OH for the mineralization of organic compounds, leading to higher TOC removal (32.29%).

The impact of UV intensity on the color removal efficiency was investigated by varying the number of UVC lamps (1, 2, and 3) and distances between lamps and the sample. The measured light intensities of different conditions were 5.12, 26.56, 50.78, 65.86 and 80.30 W/m², respectively. Light intensity directly affects color removal efficiency in the UV/H₂O₂ process. As shown in Fig. 3a, an increase in light intensity led to an enhancement in color removal efficacy, which was consistent with previous studies [44, 45]. The color removals for 5.12, 26.56, 50.78, 65.86 and 80.30 W/cm² were 49.61 (± 0.00)%, 91.60 (± 0.01)%, 94.66 (± 0.00)%, 97.71 (± 0.65)%, and 98.62 (± 0.65)%, respectively. In contrast, TOC degradation was negligible in all experimental conditions (Fig. 3b). The TOC removals were 0.44 (± 0.00)%, 2.92 (± 0.77)%, 2.49 (± 0.01)%, 2.82 (± 0.36)%, and 3.81 (± 1.04)% for 5.12, 26.56, 50.78, 65.86 and 80.30 W/m², respectively. This result suggests that the degradation products that were generated during the UV/H₂O₂ process contribute to the TOC concentration of the reacted samples (Eq. (2)).

Likewise, the increase in UV intensity in the UV/H₂O₂ treatment is effective for color removal in textile wastewater but there are still limitations regarding the mineralization of non-biodegradable organic compounds to CO₂ and H₂O.

The color removal efficiency of the UV/H₂O₂ process was also investigated at various reaction times (Fig. 4a). The color removal efficiencies were 3.21 (± 1.95)%, 40.00 (± 0.70)%, 91.60 (± 1.82)%, 99.22 (± 0.01)%, and 100.00 (± 0.00)% at reaction times of 10, 60, 120, 180, and 240 min, respectively. Longer reaction times significantly improved the color removal efficiency, with almost all of the color being removed at 180 min. As the reaction time increased, the frequency of reactions between OH radicals and organic compounds also increased, resulting in higher removal efficiency. The TOC removal efficiencies were 0.00 (± 0.37)%, 0.00 (± 0.66)%, 2.92 (± 0.77)%, 26.17 (± 0.01)%, and 32.31 (± 0.48)% at reaction times of 10, 60, 120, 180, and 240 min, respectively (Fig. 4b). A noticeable difference in TOC removal efficiency was observed at 240 min, which was consistent with previous studies [46, 47]. Therefore, our findings confirmed that longer reaction times improve TOC removal efficiency. These results further suggest that an increase in H₂O₂ concentration and reaction time affects TOC removal efficiency, whereas an increase in UV intensity has no significant effect.

The relative importance of the experimental conditions was analyzed using MLP. Fig. 5 illustrates the ratio of the mean absolute SHAP value for the experimental conditions, including H₂O₂ concentration, UV intensity, and reaction time. In terms of color removal, H₂O₂ concentration (43%) was the most important factor, followed by reaction time (34%) and UV intensity (23%) (Fig. 5a). Similarly, for TOC removal in the UV/H₂O₂ process, H₂O₂ concentration (50%) was the most important factor, followed by UV intensity (28%) and reaction time (22%) (Fig. 5b). Overall, H₂O₂ concentration emerged as the most significant factor [48, 49], and different impacts of reaction time and UV intensity were observed between color and TOC removal in the UV/H₂O₂ process.

Pair plots of the observed data and MLP predictions are shown in Fig. 6. The MLP demonstrated high predictability for all data, achieving the highest R² value of 0.9995 and the lowest mean square error (MSE) of 6.7×10⁻⁵ for color removal (Fig. 6a). In contrast, the R² value for the additional data was relatively low at 0.4172, indicating that the MLP model was overfitted. This underscores the imperative to identify supplementary strategies for averting overfitting when training an MLP model with the entire set of experimental conditions derived from the CCD. Additionally, the MLP showed considerably low predictability for TOC (Fig. 6b). Since TOC does not vary much with experimental conditions (Table S2), additional research is necessary to assess the predictability of the MLP for TOC through an appropriate fitting. The results of the comparison between observed data and MLP predictions for actual dyeing wastewater treatment confirmed that MLP is an appropriate predictive model for color removal.