Machine learning-driven performance prediction of Z-scheme g-C₃N₄/SnS₂ heterostructure photocatalyst for complete mineralization of indigo carmine and elucidation of degradation pathways

Abstract

The extensive use of synthetic dyes in industry has raised environmental concerns due to their high toxicity and persistence in water. In this study, a Z-scheme g-C₃N₄/SnS₂ heterostructure was synthesized via a facile one-pot thermal decomposition method with varying SnS₂ loading. The nanocomposites were thoroughly characterized using XRD, FT-IR, DRS, PL, Raman, XPS, FE-SEM, and TEM to confirm their structural, optical, and morphological structures. The photocatalytic performance was evaluated for indigo carmine degradation under natural sunlight. The nanocomposite GS5 sample (5% SnS₂) showed the highest efficiency, achieving 100% dye removal and 74.1% mineralization (TOC) of 10 ppm dye within 30 min at 1 mg/mL catalyst dosage, and 88.67% removal at 50 ppm dye concentration. The photocatalytic performance of the g-C₃N₄/SnS₂ nanocomposites was evaluated by varying key parameters such as pH, catalyst dosage, dye concentration, and regeneration cycles. Structural and optical analyses confirmed the formation of a well-coupled Z-scheme heterojunction, promoting charge separation and reactive oxygen species generation. The degradation mechanism, supported by GC-MS and scavenger studies, highlighted the dominant role of superoxide radicals. Machine learning models, including Random Forest, ANN, SVM, and XGBoost, successfully predicted photocatalytic efficiency, with Random Forest showing the highest accuracy (R² = 0.9734, error = 6.24). This study highlights the importance of light-driven photocatalysis, efficient Z-scheme heterostructures for improved charge separation, and the limited reports using machine learning with g-C₃N₄/SnS₂. It demonstrates the potential of g-C₃N₄/SnS₂ as a solar-driven photocatalyst for degrading recalcitrant dyes and the role of machine learning in predicting real-world performance.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-37528-5.

Introduction

Environmental pollution, particularly water contamination from industrial effluents, has become a critical global concern. Textile, paper, and pharmaceutical industries release large volumes of dye-containing wastewater, which poses a serious threat to the environment due to the high solubility, toxicity, and non-biodegradability of dyes^1–3. One such dye, indigo carmine, is extensively used in food, textile, and pharmaceutical industries because of its vivid color, stability, and low cost. However, it is toxic and has been reported to cause harmful effects, including allergic reactions, mutagenic and carcinogenic effects⁴. Therefore, effective treatment strategies are essential to eliminate such dyes and safeguard both the environment and living organisms.

Several techniques have been investigated for dye removal, including adsorption, ion-exchange, filtration, coagulation, and photocatalytic degradation^5–9. However, most of approaches only transfer contaminants to another phase rather than achieving complete mineralization, often leading to secondary pollution. In contrast, photocatalysis stands out as a sustainable and eco-friendly approach, where semiconductor materials harness light energy to generate reactive oxidizing species (ROS) that degrade pollutants into harmless byproducts such as CO₂ and H₂O, without generating secondary pollution. The fundamental photocatalytic mechanism involves several key stages: (1) light harvesting, where the photocatalyst absorbs photons, (2) charge excitation, where an electron is excited from the valence band to the conduction band, leaving a hole behind, (3) charge separation and transfer and (4) surface redox reactions. This process leads to the generation of highly reactive oxidizing species, which are the primary agents for degrading the pollutants^10–12. Photocatalysis offers additional advantages, including mineralization capability, operational simplicity, mild operating conditions, no secondary pollutants and low cost^7,13,14. Over the last decade, advanced oxidation processes (AOPs) and heterogeneous photocatalysis have emerged as effective approaches for wastewater treatment using photocatalysts.

Carbon-based materials have gained considerable attention as photocatalysts due to their abundance, tuneable electronic properties, and chemical stability¹⁵. Among them, graphitic carbon nitride (g-C₃N₄) is particularly promising because of its readily available precursors, ease of synthesis, layered structure, metal-free composition, visible-light activity, and structural tunability^16–18. g-C₃N₄ has been successfully applied in diverse fields, including hydrogen generation via water splitting, CO₂ reduction, photosensitive detectors, and the degradation of organic pollutants^19–22.

Despite these advantages, pristine g-C₃N₄ still faces challenges such as low surface area and fast recombination of photogenerated electron–hole pairs, which limit its photocatalytic efficiency^23,24. To overcome these limitations, strategies such as doping, morphological tuning, molecular structure optimization, and coupling with other semiconductors to form heterojunctions have been explored²⁵. Heterojunctions, in particular, enhance visible-light absorption and promote charge separation, significantly improving photocatalytic activity^26–29.

Recently, Z-scheme heterojunctions have emerged as a superior design compared to conventional type-II systems. Unlike type-II heterojunctions, which often reduce redox ability, Z-scheme systems preserve the strong oxidation and reduction potential of both semiconductors while facilitating efficient interfacial charge transfer. This mechanism allows photogenerated electrons in the CB of one semiconductor to recombine with holes in the VB of the other, maintaining strong redox activity. Several reports have demonstrated the advantages of Z-scheme systems for enhanced photocatalytic performance^30–32.

A wide range of g-C₃N₄-based heterojunctions have been developed in combination with other semiconductors^33–40. Among g-C₃N₄-based heterojunctions, tin disulfide (SnS₂) is especially attractive due to its narrow band gap (~ 2.2 eV), chemical stability, facile synthesis, and non-toxicity^41,42. The formation of heterostructures of g-C₃N₄ and SnS₂ broadens light absorption in the visible light region and enhances charge separation, resulting in improved photocatalytic performance^25,43. However, most studies focus on hydrogen evolution, CO₂ reduction, or general pollutant degradation, with only a few (9 out of 45) addressing dye degradation. To date, no reports exist on indigo carmine degradation using g-C₃N₄/SnS₂ nanocomposites. Moreover, conventional preparation methods rely on hydrothermal or solvothermal routes, requiring high temperature, pressure, and long reaction times, while simple, scalable, and cost-effective chemical mixing methods remain unexplored (Table 1).

Table 1.

Comparison of photocatalytic degradation of Indigo Carmine dye using different photocatalyst.

Sr. No	Catalyst	Concentration (ppm)	Dose (mg/ml)	Light source	TIME (min)	%removal	Technique used to propose mechanism	TOC (mg/mL) (Mineralization)	Ref.
1	CoFE2O4/SnO2	23	1	UV	120	55	-	-	51
2	Ni, C, N, S doped ZnO	15	0.2	Sunlight	120	99	-	-	52
3	Ni-BaMo3O10	20	1.42	Visible light	220	98	HR-QTOF, ESI/MS	70.92%	53
4	Al doped ZnO	10	0.15	Sunlight	140	97	-	-	54
5	Ag/ZnO	10	1	Visible light	120	95.71	-	-	55
6	TiO2 nanoparticle membranes	20	1.25	UV	180	100	-	-	56
7	Porphyrin conjugated TiO2	10	0.2	Visible light	120	99.77	-	-	57
8	Ag/Bi2O3/C	10	0.5	UV	18	98.6	LC-MS	-	58
9	Co3O4	10	1	Sunlight	240	99	1 H NMR	-	59
10	Fe-doped TiO2	10	0.5	Visible light	60	95	-	-	60
11	ZnS/Sm3+	10	1.5	UV	210	93	-	-	61
12	AgIO4/ZnO	23	1	Sunlight	120	98	-	-	62
13	WO3/CeO2	3	2	UV-vis	120	45	-	-	63
14	ZrO2/rGO	10	-	UV	60	85	-	-	64
15	Ag3VO4/CFB	24	0.2	UV	120	95	-	-	65
16	CdS/TiO2	10	1	Sunlight	300	88	-	-	66
17	a- Fe2O3/ bentonite	10	25	UV	120	18	-	-	67
18	Mn & S@ TiO2	20	1.5	Visible light	90	100	-	-	68
19	Se-doped ZnO	14	1	UV	480	96	-	-	69
20	ZnO/Nb2O5	46	0.35	UV	300	97	-	-	70
21	Bi2S3/ZnC02O4	40	1.5	UV-vis	90	77.4	-	-	71
22	Bi2O3/Bi2WO6	50	0.5	UV-vis	120	99.7	-	-	72
23	MnFe2O4@BC	30	1	Visible light	120	69.07	-	-	73
24	Pt-Pd-graphene	100	0.5	UV-vis	50	65	-	-	74
25	CuAl2O4	8	0.1	UV	25	99	-	-	75
26	g-C3N4/SnS2 (5% loading)	10	0.1	Sunlight	30	90	GC-MS (Anthrallinic Acid)	74.17%	This work
			1			100
		50	0.1			90

Additionally, advanced machine learning were employed to predict photocatalytic performance. Machine learning (ML) has rapidly become a transformative approach for predictive modelling and process optimization, with applications ranging from health care and finace to inductrial and healthcare systems⁴⁴. By offering data-driven insights, ML assists in improving efficiency, streamlining operations, and accelerate decision making. Popular algorithms such as support vector regression (SVM), Gaussian processes (GP), artificial neural network (ANN), and ensemble learning models are frequently utilized for training dataset. For example, Meng et al., applied the Catboost algorithm for high-throughput analysis to predict the degradation performance of TiO₂ doped photocatalysts in air pollutants, achieving an impressive R² values of 0.93⁴⁵. Similarly, Ganaprakasham et al. evaluated nine different ML models including multiple linear regression, polynomial regression, decision trees, random forest, extreme gradient boosting (XGBoost), adaptive boosting, and k-nearest neighbours (KNN), SVM, and ANN to predict the photocatalytic performance of BiVO₄ nanoparticles. Among these, XGBoost exhibited the best performance with an R² value of 0.85⁴⁶. Ahmed et al. employed HistGradient Boosting to predict the photocatalytic degradation of methylene blue in contaminated water, reported excellent accuracy with R² score 0.9905 and RMSE of 2.374⁴⁷.

In this study, g-C₃N₄/SnS₂ heterojunctions were synthesized via a facile one-pot thermal decomposition method with varying SnS₂ loadings (5–40%). The photocatalytic performance was evaluated for indigo carmine degradation under natural sunlight, achieving complete decolorization (100%) and 74.1% mineralization within 30 min at a catalyst dosage of 1 mg/mL. Mechanistic studies, supported by GC-MS and scavenger experiments, confirmed a Z-scheme charge transfer pathway with hydroxyl and superoxide radicals as the dominant reactive species. Advanced machine learning models, including Random Forest (RF), XGBoost (XGB), and ANN, were applied to predict photocatalytic efficiency. Integration of experimental data with ML enabled accurate forecasting of mineralization under diverse conditions, identified key parameters influencing performance, and provided guidance for process optimization.

This combined experimental-computational approach demonstrates a simple, scalable, and cost-effective strategy for developing high-performance g-C₃N₄/SnS₂ heterostructures, highlighting their potential for efficient degradation of recalcitrant dyes and sustainable environmental remediation.

Result and discussion

The results and discussion are divided into three sections. Section 2.1 focuses on the outcomes of characterization techniques, while Sect. 2.2 explores the photocatalytic results related to the degradation of IC. Section 2.3 discusses the machine learning predictions of the photocatalytic performance analysis.

Characterization results

The synthesized photocatalyst and nanocomposites were characterized for the phase, shape, functional groups, optical properties, and morphological characteristics. The results are discussed as follows.

Phase analysis

The XRD pattern of pure g-C₃N₄, pure SnS₂ and GS nanocomposites are shown in Fig. 1 (a). The g-C₃N₄ XRD pattern shows two distinct diffraction peaks at 12.8° and 27.6° corresponding to (100) and (002) planes, which are attributed to in-plane structural packing and interlayer stacking of the aromatic system, respectively. The XRD pattern of g-C₃N₄ match well with the standard JCPDS file no. 87-1526 confirming the presence of hexagonal phase. For pure SnS₂ three prominent diffraction peaks are observed at 28.3°, 32.1° and 49.7°, corresponding to (100), (101) and (110) diffraction planes, respectively. These peaks match well with the standard JCPDS card no. 23–0677, confirming the formation of hexagonal phase SnS₂. In the GS nanocomposites the characteristics peaks of both g-C₃N₄ and SnS₂ are clearly observed which indicates the successful synthesis of g-C₃N₄/SnS₂ nanocomposites. The intensity of (110) plan of SnS₂ increases from GS 5 to GS 40 as the loading of SnS₂ increases. Additionally, the (100) diffraction peak of SnS₂ at 2θ = 28.3° becomes more distinct from the (002) peak of g-C₃N₄ at 27.6° in GS 20 and GS 40 samples, further confirming the enhanced SnS₂ loading in the composites. The purity of the parent compounds in the samples was confirmed via Rietveld analysis as shown in Fig. S1. The Rietveld analysis confirmed the formation of pure phase of SnS₂ and g-C₃N₄. The rietveld analysis confirmed the presence of high intensity (002) plane indicated graphitic carbon like arrangement in g-C₃N₄ sample, while SnS₂ showed a purity of 68% indicating partial surface oxidation. In the case of GS-5 composite the presence of all SnS₂, g-C₃N₄ and graphitic carbon was observed. No additional impurity peaks are detected, demonstrating the phase purity and effective interfacial coupling between the two components. Moreover, a slight broadening and minor shift of the (002) peak of g-C₃N₄ toward lower angles in the composites suggests lattice strain and interfacial interaction between g-C₃N₄ and SnS₂. These results confirms that SnS₂ nanoparticles are well integrated with g-C₃N₄ layers without altering the crystal structure of either component, facilitating the formation of intimate interfacial contact, a crucial factor for efficient charge separation and transfer in Z-scheme photocatalysis. The XRD findings are further corroborated by TEM observations, which reveal uniform SnS₂ nanoparticle distribution over the g-C₃N₄ nanosheets, supporting the formation of a tightly coupled heterojunction structure.

Fig. 1.

(a) XRD pattern and (b) FT-IR spectra of pure g-C₃N₄, pure SnS₂ and GS nanocomposites.

Functional group analysis

Figure 1(b) shows the FT-IR spectra of g-C₃N₄, SnS₂, and GS nanocomposites. g-C₃N₄ exhibits characteristic peak at 811 cm^− 1 due to the bending vibration of the s-triazine ring, confirming the presence of the g-C₃N₄ framework. The absorption bands in the range of 1238–1456 cm^− 1 correspond to various stretching modes of the aromatic heterocyclic framework in g-C₃N₄. Additionally, the peaks observed between 1541 and 1631 cm^− 1 are assigned to C = N stretching vibrations. A broad band in the region of 3000–3350 cm^− 1 corresponds to the stretching vibrations of N–H and O–H groups, from residual surface hydroxyls or amine functionalities. For SnS₂, a distinct absorption band around 625 cm^− 1 corresponds to the Sn–S stretching vibration. In the GS nanocomposites, all samples exhibit the characteristic peaks of g-C₃N₄, indicating that the graphite-like structure of g-C₃N₄ is well preserved during composite formation. However, due to the relatively lower loading of SnS₂ in the composites, the characteristic bands are not clearly visible. Moreover, a slight shift in the C-N and C = N stretching bands is observed with increasing SnS₂ loading, suggesting strong interfacial interaction and possible electronic coupling between g-C₃N₄ and SnS₂, which enhance charge transfer efficiency and facilitate heterojunction formation, contributing to the improved photocatalytic performance of the composites.

Morphological analysis

The FE-SEM images g-C₃N₄, SnS₂, and GS nanocomposites are shown in Fig. S2. The g-C₃N₄ exhibits a typical layered and systematically arranged nanosheet-like morphology with porous texture, characteristic of graphitic structure. In contrast, SnS₂ displays a nanoflake-like fibrous structure. For the GS nanocomposites, GS-5 retains the nanosheet-like layered morphology like g-C₃N₄, indicating the dominant presence of g-C₃N₄. As the SnS₂ loading increases from GS-10 to GS-40, the morphology of the composites gradually shifts toward nanoflake-like structures, resembling that of SnS₂. This morphological evolution enhances the surface area, thereby providing more active sites for dye adsorption.

Further morphological investigation was conducted using TEM analysis as shown in Fig. 2. The TEM images confirm the sheet-like morphology of g-C₃N₄ and reveal disc-like structures resembling RBCs for the SnS₂ nanoparticles. In the nanocomposite, SnS₂ discs are clearly seen deposited on the g-C₃N₄ nano-sheets, confirming successful heterojunction formation.

Fig. 2.

(a) TEM images of (a) pure g-C₃N₄, (b) SnS₂, (c) GS nanocomposites, (d) HR-TEM image, and (e-f) elemental mapping image of GS nanocomposite.

The SAED image of the g-C₃N₄-SnS₂ sample (Fig. 2c, inset) indicates a polycrystalline nature. The observed diffraction spots and rings were indexed to the (110), (101), and (100) planes of SnS₂ and the (002) plane of g-C₃N₄. HR-TEM images further confirm the presence of both parent materials in the composite structure. The elemental mapping analysis of the GS-5 sample confirmed the presence of C, N, Sn, S in the composite. Figure 2 (e-f) depicts the homogeneity of the dispersion of SnS₂ over g-C₃N₄ in the nanocomposite sample. This uniform elemental distribution suggests strong interfacial interaction and effective heterojunction formation between SnS₂ and g-C₃N₄, which is beneficial for enhancing the photocatalytic performance of the GS nanocomposites.

Optical analysis

The optical properties of the prepared samples were investigated using UV–Vis diffuse reflectance spectroscopy (UV-DRS) and photoluminescence (PL) spectroscopy. The UV–Vis absorption spectra of g-C₃N₄, SnS₂, and GS nanocomposites are shown in Fig. S3(a). As the SnS₂ loading increases, the GS nanocomposites show a gradual red shift in the absorption edge toward longer wavelengths, indicating enhanced visible-light absorption.

Figure S4 shows Tauc plot of g-C₃N₄, SnS₂ and GS nanocomposites were studied using DRS. The bandgaps were observed to be 2.35 and 2.81 eV for SnS₂ and g-C₃N₄ respectively. For GS nanocomposites, two band edges were observed, corresponding to the presence of both g-C₃N₄ and SnS₂, confirming the successful formation of the composites. The band gaps of the GS-5, GS-10, GS-20 and GS-40 nanocomposites were found as 2.60, 2.58, 2.45 and 2.35 eV, respectively. No significant change in the bandgap of g-C₃N₄ was observed in the GS nanocomposites, as it primarily serves as the substrate. The reduction in bandgap energy of the GS nanocomposites might be due to the strong interfacial interaction of g-C₃N₄ and SnS₂, hence facilitates electronic coupling and charge transfer between the two semiconductors.

The PL spectra of g-C₃N₄, SnS₂ and GS nanocomposites are shown in Fig. S3(b). It shows that SnS₂ exhibits weaker emission with three peaks at 420, 490, and 530 nm, reflecting its higher charge separation efficiency. GS nanocomposite with 40% (GS-40) loading shows SnS₂ type behaviour while other three composites show emission like g-C₃N₄ indicates that the composite with less SnS₂ loading show g-C₃N₄ type behaviour. In general, a decrease in PL intensity indicates suppression of charge-carrier recombination and an increased carrier lifetime. However, it is also important to note that an increase in PL emission can, in some cases, suggest suppression of non-radiative recombination pathways, leading to higher radiative efficiency and potentially enhanced catalytic activity.

According to available reports, higher PL intensity may indicate that a greater proportion of excitons recombine radiatively rather than through non-radiative pathways associated with defects and trap states. This observation indirectly points to the formation of a defect-free or low-defect photocatalyst. Therefore, in certain cases, a photocatalyst with higher PL emission can be considered more optically efficient due to the reduction of non-radiative recombination, facilitating more efficient charge-carrier transfer for photocatalytic reactions.

XPS analysis

XPS analysis was carried out to examine the surface chemical composition and bonding configuration of the photocatalyst, as shown in Fig. 3. The survey spectra of SnS₂ and g-C₃N₄-SnS₂ composite confirm the presence of Sn, S, C and N in the composites (Fig. 3a). The C1s spectra exhibits two peak at 284.4 eV and 287.8 eV corresponding to the sp² C–C bond of graphitic carbon and sp² bonded carbon in the triazine rings, respectively (Fig. 3b). A slight shift of these peaks toward higher binding energies (eV) indicate interfacial interactions and bonding between the parent components in the composite. The N1s spectrum of pure g-C₃N₄ (Fig. 3c) displays three main peaks at 399.1, 400.1, and 401.9 eV, assigned to sp² hybridized N in C = N-C in the triazine ring, tertiary nitrogen -N-(C₃) and protonated nitrogen species (N-H) species, respectively. In the g-C₃N₄/SnS₂ nanocomposite, the corresponding binding energies shift slightly to 399.4, and 401.6 eV, with a more pronounced shift observed in the g-C₃N₄/SnS₂ heterostructure, indicating a strong interfacial interactions formed during the composite synthesis. The Sn 3d spectra of pure SnS₂ and the GS5 composite (Fig. 3d) shows two peaks at 493. 8 and 485.4 eV, corresponding to Sn 3d_3/2 and Sn 3d_5/2, respectively. A decrease in the peak intensity in the composite is attributed to the low SnS₂ loading (5%) on the g-C₃N₄ sheets. Similarly, the high-resolution spectrum of S 2p (Fig. 3e) can be deconvulated into two peaks at 162.0 and 163.1 eV, assigned to S 2p_3/2 and 2p_1/2. In the composite g-C₃N₄/SnS₂ these peaks shift to 161.5 and 162.5 eV, respectively. The observed shift towards lower binding energies in both S 2p and Sn 3d spectra after SnS₂ loading indicate the formation of heterojunction interface, likely involving Sn-N and S-C binds⁴⁸.

Fig. 3.

(a) Survey spectra of GS-10 and SnS₂, (b) C 1s, (c) N1s, (d) S 2p and (e) Sn 3d spectra of g-C₃N₄, SnS₂ and g-C₃N₄-SnS₂.

Photocatalytic studies

The prepared photocatalysts and nanocomposites were analysed for adsorption, degradation, % removal, kinetic studies, performance assessment, and the results are discussed in detail as follows.

Adsorption, photocatalytic degradation and % removal studies of IC dye

The photocatalytic degradation of IC was investigated using the prepared GS nanocomposites, g-C₃N₄ and SnS₂ as the photocatalysts under natural sunlight irradiation. Figure 4 (a) shows bar graph illustrating the adsorption and degradation efficiencies of IC using g-C₃N₄, SnS₂, and GS nanocomposites. It can be seen that the g-C₃N₄ exhibits 2.43% adsorption and 80.48% degradation, while SnS₂ demonstrates 100% adsorption with no subsequent degradation. In the case of nanocomposites GS5, GS10, GS20 and GS40 show 7.31%, 26.82%, 29.26% and 53.65% adsorption and 92.68%, 60.47%, 50.24% and 25.26% degradation, respectively. Among all the samples the GS 5 nanocomposites show the maximum degradation efficiency as compared to parents’ materials and other nanocomposites. Figure 4 (b) shows the % removal of IC which was found to be 82.91%, 100%, 99.99%, 87.29%, 79.5%, and 78.91% for g-C₃N₄, SnS₂, GS 5, GS 10, GS 20 and GS 40 nanocomposites, respectively.

Fig. 4.

(a) Adsorption and degradation bar graph (b) % removal of IC using g-C₃N₄, SnS₂ and GS nanocomposites.

Figure 5 shows the digital image illustrating the color change of the catalysts before and after the photocatalytic degradation of IC. It can be concluded that in g-C₃N₄, GS5 and GS10 exhibits no significant colour change suggesting mineralization of the dye. In contrast, GS 20 and GS 40 shows change in colour suggesting adsorption behaviour like SnS₂.

Fig. 5.

Digital image shows the change in the colour of g-C₃N₄, SnS₂ and GS nanocomposites before (B.T) and after (A.T) the treatment with IC.

Kinetic study

The kinetic study of the synthesized samples for adsorption and photocatalytic degradation of IC was monitored at regular time intervals. The experimental data were fitted into the first and second order kinetics models using Eqs. 1 and 2. Fig. S5 shows the absorbance versus time plot for the adsorption and degradation of IC using g-C₃N₄, SnS₂, GS 5, GS 10, GS 20 and GS 40, respectively. Fig. S6 (a-d) shows the kinetic data fitted to 1st order and 2nd order kinetics model. The corresponding rate constants (K) and correlation coefficients (R²) derived from these fittings are summarized in Table S1. From the graph it can be concluded that the higher value of R² for adsorption shows that it follows second order reaction and for degradation of IC follows first order kinetics, as higher R² value in the first-order model fitting.

Performance assessment

The conclusion from above study indicate that the GS 5 nanocomposite exhibits the maximum photocatalytic performance among all the tested samples, demonstrating superior photodegradation capability. Therefore, further studies by varying parameters (dosage and pH), and scavenger analysis were performed using GS5 photocatalyst. These experiments involved a 30-minute dark to evaluate adsorption, followed by a 30-minute exposure to natural sunlight to evaluate degradation efficiency.

Dosage study

A dosage study was conducted to determine the optimal amount of GS 5 photocatalyst required for maximum removal of IC. Various dosages of GS 5, ranging from 0.1, 0.2, 0.5, 1, and 2 mg/mL, were examined. Figure 6 (a) shows a bar graph illustrating the % removal of IC at different catalyst dosages. The removal efficiency increased from 89.27% to 99.9% when the dosage was raised from 0.1 to 1 mg/mL, due to the high concentration of active sites for photocatalytic degradation. Increasing the dosage from 1 to 2 mg/mL did not lead to a significant improvement, indicating that the catalyst had reached its saturation point. Consequently, a dosage of 1 mg/mL was chosen as the optimum dosage.

Fig. 6.

Bar graph representation of performance assessment (a) with different amounts of dosage (b) pH study (c) scavengers’ study in the presence and absence of scavengers and (d) concentration studies.

pH study

Figure 6 (b) shows the effect of pH on the photocatalytic degradation of IC dye. The pH experiment was conducted at different pH levels 2, 4, 6.27 (pH of 10 ppm solution of IC), 9 and 11. The % removal of IC was found to be 100, 99.2, 99, 99 and 99% at pH 2, 4, 6.27, 9 and 11, respectively. Although the removal efficiency remained high across all pH levels, the highest degradation was observed at pH 2, suggesting slightly enhanced performance under strongly acidic conditions. These results indicate that the GS 5 photocatalyst is effective over a wide pH range but shows optimal activity in an acidic medium. Therefore, acidic conditions are slightly more favourable for the maximum removal of IC dye.

Scavenger study

The primary ROS were identified through the scavenger analysis, as shown in Fig. 6 (c). Different scavengers such as isopropanol (IPA), ammonium oxalate (AO), AgNO₃ and benzoquinone (BQ) were used for hydroxyl, holes, electrons and superoxide radicals, respectively. From the results, it can be observed that the BQ shows a significant decrease in degradation efficiency to 75.37% while IPA shows a minor decrease to 95.18%. In contrast, A.O. and AgNO₃ show no change, with a degradation efficiency of 100% suggesting that holes and electrons do not contribute to the mineralization process. The scavenger studies results confirm that superoxide radicals () are the primary reactive species responsible for dye degradation, while •OH radicals play a minor role for the photocatalytic degradation of IC.

To gain deeper insights into the performance of the best photocatalyst, GS-5, in comparison with g-C₃N₄, a comparative study was conducted using IC dye solutions of varying concentrations (10–50 ppm). As shown in Fig. 6 (d), GS-5 shows potential for degrading higher concentrations of the dye, achieving 88.67% removal at 50 ppm and up to 91.44% removal at 40 ppm. In contrast, g-C₃N₄ exhibited only 62.77% removal at 50 ppm, indicating lower efficiency at higher dye concentrations.

As discussed in the literature review, among 45 papers, the majority employed separate synthesis followed by physical mixing of the parent materials to form nanocomposites. In contrast, the present study utilized a thermal decomposition approach, allowing SnS₂ to integrate directly with the 2D nanosheets of g-C₃N₄. For comparative analysis, the GS 5 nanocomposite was prepared using two methods: chemical thermal decomposition (GS-5 C) with 5% SnS₂ loading, and physical mixing (GS-5P) with the same loading. Figure 7 shows the comparative performance of the nanocomposites prepared via these two approaches. It was observed that the composite synthesized by the one-pot chemical thermal decomposition method (GS-5 C) significantly outperformed GS-5P, despite both having the same SnS₂ content. This highlights the effectiveness of the chemical integration approach in enhancing photocatalytic performance.

Fig. 7.

Comparison of the removal efficiency of GS-5 C and GS-5P towards the removal of IC dye.

Recycle, regeneration and stability study

To assess structural stability, the regenerated photocatalyst was characterized using XRD and FT-IR analyses before and after the recycling tests. As shown in Fig. 8 (a) and (b), there were no major changes in the crystal structure or functional groups of the catalyst after regeneration. These results confirm that the structural integrity of GS 5 is preserved, indicating good stability and reusability of the photocatalyst. The reusability of the photocatalyst was evaluated through recycling experiments conducted for the degradation of IC dye. As shown in Fig. 8 (c), the GS 5 photocatalyst maintained good photocatalytic activity over five successive cycles, with degradation efficiency decreasing slightly from approximately 99% to 86%. This indicates that GS 5 exhibits good reusability. After each cycle, the photocatalyst was regenerated by washing with acetone and subsequently dried at room temperature overnight.

Fig. 8.

(a) XRD spectra, (b) FT-IR spectrum before and after degradation of IC using GS 5 and (c) recycle study of GS 5 for the degradation of IC.

Mechanism for photocatalytic degradation of IC

Based on the scavenger studies, Mott–Schottky analysis (Fig. S7), the g-C₃N₄/SnS₂ heterostructure follows a direct Z-scheme charge transfer mechanism. The calculated conduction band (CB) and valence band (VB) potentials of g-C₃N₄ are − 0.482 eV and − 3.292 eV, respectively, whereas for SnS2 they are 0.109 eV and 2.241 eV vs. NHE. The suitable band alignment enables efficient charge separation and enhanced redox ability. The EIS results as shown in Figure S8 indicate a significant decrease in charge-transfer resistance upon formation of the heterojunction compared to the parent compounds. This clearly suggests an efficient charge-transfer mechanism in the GS-5 sample, which correlates with its highest photocatalytic activity toward the degradation of Indigo Carmine.

Under natural sunlight irradiation, both semiconductors generate photoinduced electron–hole pairs as shown in Eq. (1). In the Z-scheme mechanism, the photogenerated electrons in the CB of SnS₂ recombine with the holes in the VB of g-C₃N₄ Eq. (2). Consequently, the electrons in the CB of g-C₃N₄ and holes in the VB of SnS₂ remain active and participate in redox reactions. The electrons in g-C₃N₄ reduce O₂ to form superoxide radicals (), while the holes in SnS₂ oxidize H₂O or OH- to form hydroxyl radicals (^•OH)^49,50. These reactive oxygen species (ROS) act as the primary oxidative agents responsible for the degradation of indigo carmine (IC) dye into harmless products such as CO₂ and H₂O^30,31 Eqs. (3–7). The proposed mechanism is illustrated below in Fig. 9, highlighting the Z-scheme charge transfer pathway in the g-C₃N₄/SnS₂ heterostructure.

1

2

3

4

5

6

7

Fig. 9.

Photocatalytic degradation mechanism of g-C₃N₄/ SnS₂ nanocomposites.

Degradation pathways of IC by GC-MS study

The degradation of Indigo Carmine was further investigated for the identification of metabolites using GC–MS. The GC-MS spectra of metabolites observed during photocatalytic degradation of the indigo carmine at different intervals are shown in Fig. S9. The molecular ion peak of the IC dye was observed at m/z = 421. After 5 min, partial cleavage of the parent compound formed intermediates at m/z = 374 and m/z = 358, indicating hydroxylation and oxidative ring modification. At 10 min, peaks at m/z = 248 and m/z = 218 were detected, corresponding to 2-(2-amino-5-sulfophenyl)-2-oxoacetic acid and 2-amino-5-sulfobenzoic acid, suggesting further breakdown produced smaller aromatic compounds. After 15 min, indole-based fragments like 2,3-dioxoindoline-5-sulfonic acid (m/z = 229) were observed. Finally, after 30 min, low-molecular-weight fragments were observed, such as 2-aminobenzoic acid (m/z = 138) and 6-amino-2,3,4-trihydroxybenzoic acid (m/z = 184), confirming significant mineralization. These results confirm a stepwise degradation mechanism involving oxidative demethylation, deamination, desulfonation, and aromatic ring cleavage, finally leading to the formation of simpler, less toxic intermediates. Based on the GC-MS spectra of the metabolites, the proposed degradation pathway of the transformation products is shown in Fig. 10.

Fig. 10.

Degradation pathway of the transformation products formed during photocatalytic degradation of IC.

A comprehensive literature review was carried out to evaluate recent advancements and challenges in the photocatalytic degradation of organic pollutants, with particular focus on Indigo Carmine (IC) dye. Table 1 shows that the g-C₃N₄/SnS₂ heterostructure developed in this work delivers competitive and often superior performance compared to existing catalysts.

Among the various catalysts such as Bi₂O₃/Bi₂WO₆, Ni-BaMo₃O₁₀, Co₃O_4, and Mn & S@TiO₂, high degradation efficiencies have been achieved; however, these systems typically require longer reaction times (90–240 min) under visible or UV–vis light. In contrast, the g-C₃N₄/SnS₂ (GS) heterostructure achieves 90 to 100% removal within just 30 min under natural sunlight for 10 ppm IC, which is significantly faster than reported systems that require 480 min. The GS catalyst shows a minimum dosage (0.1 mg/mL) compared to the commonly reported 1–2 mg/mL, highlighting the high intrinsic activity of the g-C₃N₄/SnS₂ composite, while also improving cost-effectiveness and minimizing secondary sludge generation. Additionally, the GS photocatalyst maintains excellent degradation efficiency (~ 90%) at higher dye concentrations (50 ppm), outperforming many reported photocatalysts that lose efficiency due to active site saturation or light reduction at higher concentrations.

Importantly, mechanistic insights obtained through GC-MS analysis identified anthranilic acid as a stable intermediate, and total organic carbon (TOC) removal reached 74.17%, mineralization. These results not only verify efficient degradation but also highlight the eco-friendliness and robustness of the GS catalyst. Such detailed mechanistic evidence under low catalyst loadings and short irradiation times is rare in existing literature.

A comparison of g-C₃N₄/SnS₂ photocatalysts reported for various dyes (Table S2) indicates that the nanocomposite synthesized in the present work exhibits superior performance, particularly for indigo carmine (IC) degradation. While earlier studies on methylene blue, methyl orange, and rhodamine B report high degradation efficiencies (90–98.7%), they typically require longer irradiation times (60–360 min) and/or higher catalyst loadings (0.1875–1 mg/mL). In contrast, the thermal decomposition-derived g-C₃N₄/SnS₂ in this study achieves complete IC removal (100%) within just 30 min at 1 mg/mL for 10 ppm concentration and maintains a high efficiency (88.67%) even at 50 ppm IC, demonstrating excellent activity at both low and high pollutant loads. Moreover, previous reports mainly focus on cationic dyes, whereas the present work targets the more recalcitrant anionic dye indigo carmine, underscoring the robustness and versatility of the prepared catalyst. The combination of short reaction time, high degradation efficiency across a wide concentration range, and applicability to challenging dye structures establishes the present g-C₃N₄/SnS₂ system as a potential candidate for the environmental remediation applications.

Machine learning models for predictive analysis

Machine learning models have enabled us to identify patterns in the data and predict the performance of the catalyst. These models are trained using sample datasets, learning the correlations between the variables. Once trained, they are evaluated on new, unseen data to make accurate predictions. In this research, machine learning methods are employed solely for predictive modelling to assess the photocatalytic performance of the g-C₃N₄-SnS₂ photocatalyst. The photocatalysis data from Sect. 5.2 was used as sample dataset and used for running simulations of the ML models. The experimental features are listed in Table S3. Pearson Correlation coefficient were evaluated to quantify the relationship between parametric variables (Fig S10). The coefficient values, provide a precise extent of the interdependence on variables. Coefficient value of 0.79 indicated strong positive correlation of % removal with the Exposure time, while concentration with a coefficient value of 0.046 showed slight co-relation as shown in Fig.S8. Hence, feature selection techniques identified two important variables, exposure time and concentration of the dye as correlated with removal efficiency.

Based on the degradation results and dataset obtained, four ML models namely Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), XGBoost (XGB) were trained and evaluated for predictive analysis of removal efficiency. The models were trained for 80% of the dataset and tested on the remaining 20%, with evaluation metrics including coefficient of determination (R²), root mean squares error (RMSE) mean absolute error (MAE), and standard deviation of errors. The hypermeters significantly influence the capability of models to learn patterns from data. For ANN, the number of hidden layers, neurons per layer, activation function, learning rate and iterations were optimized to improve the performance, Similarly, for SVM, kernel type, cost parameter and gamma are the key hyper parameters, while in Random Forest, the number of trees, maximum depth and boosting rounds were tuned. Hyper parameters optimization strategies such as random search or grid search can be employed, although further tuning would enhance the performance of models such as XGB and ANN.

Figure S10 shows the ML results and residual analysis of the RF, KNN, SVM, ANN and XGB models. Table S4 and Fig. 11 show the comparative performance of models based on the R2, MAE and RMSE values. Random forest achieved the highest test R² of 0.898, lowest RMSE (10.95) and MAE (6.45, indicating its robustness and superior predictive accuracy. ANN also performed strongly with R² of 0.8915. though it’s RMSE (11.30) and MAE (8.29) were slightly higher, suggesting slightly less stability compared to RF. XGBoost achieved a test of R² of 0.8787 with MAE (7.62) and RMSE (11.95) bus exhibit slight overfitting as indicated by its perfect trading R2 of 1.0.

Fig. 11.

(a-d) Bar graph comparison and (e) normalized heat map of model performance based on valuation metrics (R², MAE, RMSE & Std. Dev.).

On the other hand, KNN and SVM performed poorly and showed weaker predictive ability, with high error values, severely under fitting the dataset. The residual analysis confirmed these trends: RF residuals were tightly clustered near zero, ANN residuals were centered with more variance at higher values, and XGB residuals were close to zero with occasional outliers, while SVR and KNN showed systematic bias and poor fit at lower values. Furthermore, performance of the model was evaluated using five normalized metrics: R² values (train & test), mean absolute error, root mean squared error, standard deviation. The result of these evaluations are presented in a heatmap as shown in Fig. 11(e). Overall, RF was identified as the best performing model, balancing accuracy and generalization without overfitting, XGBoost showed competitive performance but signs of overfitting, while ANN demonstrated strong potential with slight room for further hypermeter optimization.

Conclusion

The present work demonstrates the successful synthesis of g-C₃N₄/SnS₂ nanocomposites by a novel, scalable one-pot thermal decomposition method offering a facile chemical mixing route for heterojunction formation not previously reported. The synthesized samples were extensively characterized by XRD, FTIR, UV-DRS, FE-SEM, and PL spectroscopy to confirm their structural, morphological, and optical properties. Photocatalytic performance was evaluated under natural sunlight, where the GS5 nanocomposite exhibited the highest efficiency (100%), achieving complete degradation of indigo carmine dye (dosage = 1 mg/mL) within 30 min. The kinetics, pH, catalyst dosage, scavenger and regeneration cycles were performed to evaluate performance stability. Scavenger experiments identified superoxide radicals as the primary active species, while GC–MS analysis confirmed the breakdown of indigo carmine into simpler intermediates, supporting the proposed Z-scheme degradation mechanism. The GS5 photocatalyst maintained excellent reusability over five successive cycles, with only a slight decline in efficiency (from ~ 99% to 86%).

Furthermore, predictive modelling was performed using machine learning algorithms such as RF, SVM, KNN, ANN, and XGB, to evaluate and simulate predictive accuracy. Among them, RF outperformed all others, achieving the highest test R² (0.8982) and the lowest error values, closely followed by XGB and ANN, which also demonstrated strong predictive ability. In summary, ensemble-based models, particularly Random Forest, provided the most reliable predictions, and the use of such ML models not only authenticates the robustness of the dataset but also offers a background for improving the predictive analytics in similar applications.

This work addresses a significant research gap in the degradation of indigo carmine using g-C₃N₄/SnS₂ nanocomposites and introduces a cost-effective and sustainable solar-driven approach for water purification. The enhanced photocatalytic activity is attributed to the synergistic effect of the 2D/2D g-C₃N₄/SnS₂ heterojunction, which operates via a Z-scheme mechanism and shows excellent potential for wastewater treatment. The integrated experimental–computational strategy offers a simple and scalable route for designing efficient photocatalysts aimed at sustainable environmental remediation.

Materials and methods

Materials

Thiourea (SRL 99% pure), diphenyl ether (Sigma-Aldrich, 99%), methanol (Finar), indigo carmine (SRL), tin chloride pentahydrate (Sigma-Aldrich,98%), p-benzoquinone (Sigma-Aldrich 98%), isopropyl alcohol (Finar 99%), ammonium oxalate (Finar), silver nitrate LR (ACS). All chemicals were used without any modification and the preparation of dye solution was done using distilled water.

Synthesis of g-C₃N₄

g-C₃N₄ was synthesized via a thermal polymerization using thiourea as a single precursor. In this synthesis, 5 gm of thiourea was finely ground using a mortar and pestle, then transferred to a crucible. The powder was then subjected to thermal polymerization in a muffle furnace at 550 °C with a heating rate of 5 °C/min for 4 h. After cooling to room temperature, the resulting product was collected.

Synthesis of g-C₃N₄/SnS₂(GS) nanocomposites

g-C₃N₄/SnS₂ nanocomposites were synthesized via a facile thermal decomposition method. In a typical synthesis, 100 mg of g-C₃N₄ was dispersed in 10 mL of diphenyl ether in a round-bottom flask. Different amounts of tin (IV) chloride pentahydrate (1 mM) and thiourea (2 mM) were then added to the suspension. The reaction was heated upto 200 °C, and maintained for an hour. After cooling to room temperature, the mixture was precipitated with excess methanol. The obtained precipitate was centrifuged, washed with methanol, and dried overnight in an oven at 80 °C. The samples were named as GS-5, GS-10, GS-20 and GS-40 with loading of 5%, 10%, 20% and 40% w/w of Sn to g-C₃N₄, respectively. SnS₂ nanoparticles were synthesised using same approach in absence of g-C₃N₄.

Characterization

The synthesized g-C₃N₄/SnS₂ heterostructures were analyzed using different analysis instruments to determine the phase composition, functional group, morphology and optical properties. Rigaku Mini flex 600-C X-ray diffractometer with Cu Kα radiation (scan range: 5°-80°, scan rate- 5° /minute) was used for the phase identification. Chemical functionalities were examined by Fourier-transform infrared (FT-IR) spectroscopy using a PerkinElmer Spectrum 2 spectrophotometer (ν̄ − 4000–500 cm⁻¹) range. Optical absorption properties were examined using a PerkinElmer Lambda 365+ (scan range- 200–1100 nm). Photoluminescence (PL) spectra were acquired at room temperature using a PerkinElmer LS 55 fluorescence spectrometer, recorded within the 360–710 nm range to evaluate the recombination behaviour of photogenerated charge carriers. ZEISS Ultra 55 was used to analyse the morphology and observe surface texture and structural features. Transmission electron microscopy (TEM) images were recorded on JEOL 2100 TEM (JEOL, USA). XPS was performed on a Thermo Fisher NEXSA (Al Kα, 1486.6 eV) at 12 kV, 6.5 mA, pass energy 50 eV, under 9 × 10⁻⁸ mbar. Binding energies were referenced to C 1s (284.5 eV); peaks deconvoluted using CASA with Shirley background. To monitor absorbance changes during the photocatalytic degradation process, UV-Vis spectroscopy was performed using a PerkinElmer λ365 spectrophotometer.

Photocatalytic studies

The photocatalytic degradation of g-C₃N₄, SnS₂ and GS nanocomposites was monitored using UV-Visible spectrophotometer. A typical degradation experiment involved the addition of 1 mg/mL of photocatalyst to a 10 ppm solution of IC, followed by sonication for uniform dispersion. The solutions were then kept in the dark to achieve adsorption–desorption equilibrium. After equilibrium, 30 µL of H₂O₂ was added as an initiator, and the suspension was irradiated to sunlight for half an hour. At regular time intervals, aliquots were withdrawn, centrifuged to remove photocatalyst, and the supernatant was analysed using a UV-Visible spectrophotometer. The degradation of IC was monitored by measuring the change in intensity of the characteristic absorption peak at λ_max = 611 nm. For the kinetics study, the collected data were fitted to both first order and second-order models using the corresponding equations:

For first order kinetics;

1

For second order kinetics;

2

To investigate the influence of pH, degradation studies were performed under different pH conditions (2, 4, 9, and 11) to identify the optimal pH at which the catalyst exhibits the highest photocatalytic efficiency. The pH of the IC solution was adjusted using 1 M HCl and 1 M NaOH solutions. A study was performed to optimize the catalyst dosage for the best removal of IC. Degradation experiments were conducted using varying catalyst dose ranging from 0.1 to 2 mg/mL to assess their effect on the photocatalytic efficiency of IC. The role of ROS was investigated using scavengers such as isopropyl alcohol (IPA) for hydroxyl radicals, ammonium oxalate (AO) for holes, silver nitrate (AgNO₃) for electrons, and benzoquinone (BQ) for superoxide radicals. Additionally, the photochemical stability and reusability of the photocatalyst were evaluated through, a regenerability study.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

HP, PM, JP: data collection, analysis and interpretation of results draft manuscript preparation, GP & RG: software modelling, manuscript preparation, IT, NT, MAM SM & SS: Resources and investigation, manuscript preparation, RG: Supervision, Review and revision of the manuscript, Conceptualization. All authors reviewed the results and approved the final version of the manuscript.

Funding

This work was supported by the Deanship of Scientific Research, Vice Presidency for Graduate Studies and Scientific Research at King Faisal University, Saudi Arabia (Grant No. KFU253313).

Data availability

The data will be made available on request.

Declarations

Competing interests

The authors declare no competing interests.