ZnO-Based Photocatalysts: Synergistic Effects of Material Modifications and Machine Learning Optimization

Abstract

ZnO-based photocatalysts have attracted significant attention for their potential use in advanced oxidation processes for environmental remediation. However, critical challenges, such as rapid charge carrier recombination and narrow light absorption, and poor long-term stability necessitate material modifications to enhance performance. This review provides a comprehensive and critical analysis of recent developments in ZnO-based photocatalysts, including heterojunctions with metal oxides, carbon-based hybrids, metal/non-metal doping, and metal–organic framework materials. Furthermore, emerging trends, such as the integration of atomistic calculations and machine learning (ML) techniques in material design, property prediction, and the optimization of photocatalytic performance, are critically examined. These modern computationally driven approaches provide new insights into band gap engineering, charge transport mechanisms, and the optimization of synthesis parameters, thereby accelerating the discovery of high-performance ZnO-based photocatalysts. However, their practical integration remains limited due to the availability of high-quality datasets and the lack of interdisciplinary methodologies. The review also discusses key research gaps, including emerging environmental applications, as well as stability and scalability challenges, providing a roadmap for future research in data-driven photocatalysis. By evaluating current research, this review aims to provide a foundation for the modification of next-generation ZnO-based photocatalysts for environmental applications.

1. Introduction

Soon after the first report of the photocatalytic oxidation behavior of ZnO [1], significant attention was devoted to the photocatalytic properties of ZnO in various applications. ZnO, as a semiconductor, possesses a unique photocatalytic ability derived from its capacity to absorb irradiation and generate reactive oxygen species. Due to its unique physicochemical characteristics and antimicrobial activity, ZnO has emerged as an alternative to TiO₂ in photocatalytic reactions [2,3]. ZnO is among the leading metal oxide semiconductors used in advanced oxidation processes (AOPs) due to its cost-effectiveness, high redox potential, non-toxic nature, and overall environmental safety [4]. Heterogeneous photocatalytic oxidation, a specific type of advanced oxidation process (AOP), utilizes ZnO to efficiently treat organic compounds in aqueous solutions. This method is effective for treating wastewater, drinking water, surface water, and groundwater [5]. ZnO has many positive characteristics, but this photocatalyst faces several challenges that prevent its wider practical application. Solar light utilization by ZnO is reduced because of its large band gap, which allows it to capture only about 5% of ultraviolet light [6]. Additionally, ZnO faces challenges in practical applications due to the rapid recombination of photogenerated electron–hole pairs (e⁻–h⁺), which occurs because of the strong electrostatic attraction between them [7]. Hence, to improve the photocatalytic performance of ZnO, strategies like the synthesis of binary and ternary ZnO-based heterogeneous structures or nanocomposites have been reported [8,9].

Binary photocatalysts represent luminescent substances that are designed to enhance the visible light absorption characteristics of semiconductor host photocatalysts. In this way, the range of potential applications of photocatalysts could be expanded. They act on the basic principle of photoinduced redox reactions on excited frontier orbitals. These catalysts include dye/semiconductor systems, organic/semiconductor systems, and semiconductor/semiconductor systems [10,11]. Ternary photocatalysts involve doping or heterojunction materials combined from two different binary photocatalysts. Ternary photocatalysts usually exhibit higher activity than binary photocatalysts. Synthesis techniques include various doping methods and heterojunction combinations that enhance their photocatalytic activity [10,12,13].

While some of these strategies have led to significant improvements in photocatalytic performance, the lack of standardization in experimental conditions in different studies makes direct comparisons difficult. A significant number of factors influence the synthesis and properties of the modified ZnO photocatalyst [3], which is a time-consuming, labor-intensive, and unaffordable process. To minimize the number of experiments while maximizing the information obtained from the process, machine learning (ML) analysis can be utilized. This approach allows for the analysis and optimization of various parameters in the synthesis and photocatalytic degradation process, including pH, irradiation time, catalyst dose, and initial material concentration [14]. Additionally, screening of ZnO modifications and predicting the degradation efficiency of pollutants using ZnO-based photocatalysts can be achieved with an ML approach [15]. Despite enormous progress, challenges in building universally applicable models remain due to limited datasets and inconsistent reporting standards. Therefore, the future direction in this field should focus on the development of standardized open access databases that combine experimental and theoretical data on ZnO-based photocatalysts.

In this review, we provide a comprehensive introduction to the key features of ZnO and its modifications, focusing on photocatalytic properties. Understanding the various possible modifications, including heterojunctions with metal oxides, carbon-based hybrids, metal/non-metal doping, and metal–organic framework materials, is crucial for overcoming the limitations of pure ZnO in AOPs and enhancing its effectiveness in degrading a wide range of compounds. Several review articles [16,17,18,19] have devoted attention to the individual aspects of ZnO-based photocatalysts. However, only a few have integrated material design strategies with advancements in data-driven approaches, such as ML methods. A combination of traditional modification techniques with computational tools provides a comprehensive overview of the current state and future direction of ZnO research. Furthermore, this research aims to present recent trends in ML methods for ZnO design, property prediction, and estimation of photocatalytic degradation efficiencies. By bridging the gap between experimental strategies and ML optimization, this review highlights synergies that have not been comprehensively addressed earlier in the literature. Also, emerging environmental applications and the stability and scalability of ZnO are comprehensively discussed, providing a path for future discoveries in data-driven photocatalysis. Finally, a better understanding of the ML approach in ZnO-based photocatalysts can enhance its practical application for water purification.

2. Fundamental Aspects of ZnO Photocatalysis

There has been considerable interest in the growth of semiconductor nanostructures for various photocatalytic applications over the past few decades. ZnO is well-suited for photocatalytic applications due to its wide energy band gap of 3.37 eV at room temperature [20]. The basic physical parameters of the ZnO structure are shown in

Table 1. The basic physical parameters of the ZnO structure [20,23,27].

Physical Parameters	Values
Stable phase at 300 K	Wurtzite
Lattice constants	a = b = 0.32495 nm; c = 0.52069 nm
Melting point	1975 °C
Density	5.66 g/cm3
Band gap (direct)	3.37 eV
Refractive index	2.01
Hole effective mass	0.59
Electron effective mass	0.24
Exciton binding energy	60 meV
Static dielectric constant	8.656

. The band gap energy of ZnO also indicates that e⁻–h⁺ can occur with a minimum number of photons. ZnO has excellent thermal, mechanical, and chemical stability. Essentially, ZnO is an n-type semiconductor due to the usual dominance of group VI atoms as compensation donors [21]. At standard pressure and temperature, the crystal structure of ZnO is predominantly the wurtzite type, characterized by a hexagonal lattice, which plays a crucial role in determining its physical properties, including optical and electronic characteristics [22]. Another widely studied polymorph in which ZnO can also crystallize is the cubic zinc blende structure [23]. The structure of wurtzite is closely related to that of hexagonal zinc blende and can be thought of as two interlaced face-centered cubic lattices, with the base above and below to form a unit cell. The structures of wurtzite and zinc blende both have the same type of unit cell as diamond. The crystal structure of wurtzite is noncentrosymmetric, meaning that the lattices can support both pyroelectric and piezoelectric activity [23,24]. Surface inversion asymmetry does not exist in cubic lattices, so centrosymmetric crystals do not have this property [25]. Similarly, the other polymorph structure of ZnO is the rock salt cubic structure. Variations in crystallinity, surface polarity, and defect concentration can affect charge separation and reaction kinetics. Therefore, the crystal structure of ZnO can directly affect its photocatalytic activity. The features of the wurtzite phase are absent in cubic forms, zinc blende, and rock salt, potentially limiting their photocatalytic behavior [26].

The hexagonal wurtzite structure (Figure 1a) is the most well-known crystal structure of ZnO, characterized by its unique electronic and geometric properties. The wurtzite structure features a hexagonal crystal structure, with Zn and O atoms arranged in tetrahedral coordination. This arrangement contributes to the good piezoelectric, electronic, and optical properties of ZnO. These properties make this crystal structure suitable for different applications in materials science and engineering [28]. The lattice’s parameter is the fundamental characteristic of the hexagonal wurtzite structure [29]. Under mechanical stress, ZnO photocatalyst can generate an electrical charge because the lattice structure does not show a center of inversion symmetry, which is essential for piezoelectricity [30]. Hydrothermal and sol–gel synthesis methods are used to control the orientation and morphology of ZnO nanostructures. Hydrothermally synthesized ZnO nanorods exhibit preferential growth along the c-axis, which is consistent with their hexagonal structure [31]. Also, the presence of growth additives, such as hexamethylenetetramine, could affect morphology by favoring polar growth in the c-axis direction while suppressing lateral growth. Therefore, nanostructures with increased surface area can be obtained for catalytic applications [32]. Although the wurtzite structure dominates in ZnO-based photocatalysts, this form is not always the most efficient one. Its strong polar surfaces can lead to surface recombination of charge carriers, which are undesirable, and therefore can make the design of large-scale materials very complex. Despite the literature data on wurtzite ZnO, a comprehensive structure–performance relationship, especially in comparison with other polymorphs, is still required in the literature [17,18,19].

The zinc blende structure (Figure 1b) is also known as the sphalerite or cubic crystal structure. In this structure, the atoms are arranged in a face-centered cubic lattice, with each atom tetrahedrally coordinated to four neighboring atoms. This arrangement of atoms has unique properties that make zinc blende materials suitable for applications in optoelectronics, photonics, and nanotechnology. The relatively high symmetry is the most crucial feature of the zinc blende, and it is important for its electronic properties [33]. Zinc blende materials have a direct energy band gap, which is essential for light-emitting applications.

Furthermore, under various temperatures and pressures, the zinc blende’s structure exhibits phase stability and remains stable at room temperature. However, transitions to other forms, such as wurtzite, can occur under high temperatures (above 100 °C) or specific environmental stress conditions [34]. Also, the size plays a significant role in determining the stability of zinc blende structures. Studies show that zinc blende nanowires can maintain their crystalline structure even at very small diameters (below 50 nm), where phases like wurtzite would usually be expected. This behavior is attributed to increased surface energies and quantum confinement effects, which can modify electronic properties useful for nanoscale devices [35,36]. Also, the crystalline structure of ZnO is suitable for specific optoelectronic and quantum applications. However, zinc blende photocatalytic activity is less studied and understood compared to wurtzite [37]. Additional experimental and computational studies are required to assess the activity of this phase in photocatalytic reactions.

The rock salt structure (Figure 1c) is not a stable structure of ZnO, as it prefers the wurtzite crystal structure. In the wurtzite phase of ZnO compounds, the cation has an anionic coordination number of four. However, research shows that ZnO compounds synthetized under high pressure prefer rock salt structures that have an anionic coordination number of six. According to calculations, the cubic arrangement is the most stable for ZnO. At ambient conditions and 0 K, the ground state structure of ZnO corresponds to wurtzite, a phase derived from the stacking of hexagonal close-packed layers of Zn²⁺ and O^2– ions [38]. A phase change from wurtzite to rock salt is predicted to occur at extremely high pressures (above 60 GPa). The crystal structure, equation of state, phonon dispersion curves, and elastic properties of rock salt ZnO have been calculated using density functional theory (DFT). The rock salt phase of ZnO is energetically stable compared to the wurtzite phase in the pressure range of 84–230 GPa [39]. The pressure-induced transitions and the effect of octahedral coordination on ZnO’s properties can be studied using the rock salt phase [40]. However, due to limitations, such as extreme conditions required for its synthesis and stabilization, the application of ZnO’s rock salt phase in photocatalysis is poorly explored.

It is essential to note that wurtzite ZnO has a paper-like lattice that shares similarities with the nominal rock salt arrangement, provided that this cubic lattice is rotated, as Zn prefers octahedral coordination. Thus, it can be expected that deviations from the randomized rock salt configuration will take longer to become significant in the formation of rock salt ZnO. Bulk ZnO is moderately ductile compared to other wurtzite compounds. Hence, breaking adjacent bonds to form different ionic coordination at the surface interface would not necessarily represent a significant energy difference between materials with the bonding configuration [41].

ZnO is a wide band gap semiconductor that has attracted substantial attention in the field of photocatalysis. It has a high exciton binding energy, thermal stability, and ease of growth. ZnO has numerous applications, including UV detectors, laser diodes, and photocatalysis. However, most of its applications are limited due to the challenging and complex nature of producing p-type ZnO. This is the consequence of the ZnO tendency to exhibit n-type conductivity, which is mainly caused by native defects, such as oxygen vacancies and zinc interstitials. The electronic properties of ZnO have usually been explained by invoking intrinsic defects [42]. The frequently observed conductivity of n-type ZnO has been attributed to the presence of oxygen vacancies. However, previous theoretical studies related to an oxygen vacancy defect with deep levels below the conduction band (CB) minimum have indicated that further investigation of the origin of the unintended n-type conductivity of ZnO is required [43].

Materials’ interactions with light can be explained using optical studies. Semiconductors are characterized by their intrinsic and extrinsic properties. The extrinsic optical properties of semiconductors depend on defects or dopants introduced into the semiconductor, which can develop in discrete electronic states within the valence band (VB) and the CB. The intrinsic properties of semiconductors are related to the interaction between electrons in the CB and holes in the VB, as well as exciton effects resulting from the Coulomb interaction [44]. Techniques like photoluminescence, reflection, transmission, cathodoluminescence, and optical absorption are used to study the optical transitions in ZnO semiconductors. Among the mentioned techniques, photoluminescence is widely used to study the optical behavior of ZnO nanostructures. The photoluminescence spectra of different ZnO nanostructures exhibit both UV and visible emission bands, resulting from defects, interstitials, antisites, complex defects, and vacancies. The defect-related emissions are very significant for determining the quality, defect density, and synthesis conditions of ZnO photocatalysts. However, inconsistencies in emission profiles highlight the need for further investigation. A semiconductor with high band gap energy and a direct band gap is a luminescent material. After being excited to an energy higher than the band gap energy, the recombination of e⁻ and h⁺ can produce luminescence. The simple model of optical absorption is represented by several low-lying direct transitions covering the energy gap of 3.44 eV in the Zn-orbital mean potentials. Also, the exciton energy band of the same materials is predicted by the model to be 4.0 eV [45]. The good optical material has a low energy band gap in the visible spectrum. As mentioned, ZnO has a direct band gap energy of 3.37 eV, while the e⁻–h⁺ binding energy is 60 meV at room temperature. This shows that ZnO is not a good optical material. In the visible spectrum, the photon energy ranges from 1.68 to 3.10 eV. ZnO cannot absorb photons in this energy range because the photon energy is lower than the band gap energy [17]. The density of states plays a crucial role in understanding the physicochemical characteristics of materials in materials science, such as, for example, describing peaks in optical or vibrational spectra, e.g., the electronic density of states near the Fermi level directly affects the light absorption and charge carrier dynamics of ZnO-based photocatalysts.

Semiconductor nanocatalysts are considered some of the most suitable materials for removing various organic pollutants present in the environment. ZnO exhibits a large bandwidth, excellent photoabsorbance in the UV region, and efficient charge-carrying capacity. ZnO has also proven to be one of the most effective photocatalysts for the photoinduced removal of various types of industrial pollutants and hazards [46]. However, the practical use of ZnO is limited by its low quantum efficiency under solar irradiation. Consequently, the hybridization of ZnO with visible light-active materials or the development of heterojunction-based systems is a promising strategy in this research area. The rapid charge separation responses are generally attributed to the thin electron–hole diffusion distance and, mainly, the unique structural and surface properties. The size of ZnO particles is equal to the quantum power. This affects the textural characteristics and the recombination of photoinduced e⁻–h⁺ pairs in the ZnO semiconductor. In photocatalytic processes, the rapid separation of e⁻–h⁺ pairs and their low recombination play a crucial role in achieving high catalyst activity. ZnO is widely used in AOPs, primarily in photocatalytic reactions. In AOPs, the primary reactive species, hydroxyl radicals (^•OH), are utilized for the degradation and mineralization of a wide range of organic pollutants into CO₂, H₂O, and inorganic ions. Besides TiO₂, ZnO, with its electrical properties and photoperformance, represents one of the most often used photocatalysts nowadays [47,48,49].

AOPs are techniques used for the oxidation of toxic pollutants by semiconductors, which are activated through the absorption of light. Nowadays, AOPs are widely used due to their oxidation ability, as they are able to degrade and mineralize the organic pollutants present in the environment. The most studied AOPs are direct photolysis, photocatalysis, ozonization, sonolysis, and the Fenton process. Photocatalysis is a process in which a semiconductor material is activated upon the absorption of light. After light absorption, a highly reactive species, ^•OH, is generated, which can oxidize the organic pollutants found in wastewater [50]. Despite AOPs’ efficiency, scalability and energy consumption remain considerable challenges in industrial applications.

The absorption of radiation facilitates the generation of charge carriers in ZnO. This leads to the formation of photogenerated e⁻–h⁺. Upon the absorption of photons with an energy equal to or higher than the band gap of ZnO, electrons are promoted from the VB to the CB, leaving a positively charged hole behind. In this way, e⁻–h⁺ is generated. The formation and recombination of e⁻–h⁺ play a main role in the efficiency of ZnO in photocatalytic reactions [51]. e⁻–h⁺ also determines the electronic, optoelectronic, and photocatalytic properties of ZnO. The main challenge is separating e⁻–h⁺ to extend their reactivity and lifetime. When a photon is absorbed, electrons are promoted from the VB to the CB, resulting in the formation of free charge carriers. The formation of photogenerated charge carriers is important for ZnO applications in photocatalysis, UV photodetectors, and light-emitting devices. However, a challenge in optimizing these processes is understanding the interaction between photocarriers and their environment, particularly in light of the recombination process [52]. Their recombination process can determine the efficiency of photogenerated charge carriers. There are two different mechanisms of recombination: biomolecular and unimolecular recombination. Biomolecular recombination involves the interaction between free electrons and holes, leading to the formation of e⁻–h⁺ or neutral atomic configurations. On the other hand, unimolecular recombination occurs when a surplus of charge carriers is present, leading to spontaneous breaking up in the absence of complementary charge carriers [53]. Understanding the dominance of one recombination type over the other in different ZnO morphologies or defect environments is still an open question in this field. e⁻–h⁺ recombination typically results in a multiexponential decay of the photocurrent response. The formation of interstitial electrons, subpicosecond recombination mediated by deep traps, picosecond recombination through defects at the film/substrate interface, microsecond temporal recombination at grain boundaries, and millisecond charge trapping on oxide surface states through exothermic transfer of holes to adsorbed species correlate with the evolution of the photocatalytic reaction. The time-resolved intensity and decay of the fundamental absorption can be primarily determined through shallow electron recombination assisted by traps [54].

Photocatalysis is a process in which a photon’s energy generates charge carriers in a semiconductor catalyst that subsequently drives redox reactions. This phenomenon occurs naturally in processes like photosynthesis and is fundamental to several applications, including the degradation of pollutants, water splitting, nitrogen fixation, organic synthesis, and carbon dioxide reduction [55]. These processes are classified into heterogeneous and homogeneous photocatalysis. In heterogeneous photocatalysis, the reactions take place over a solid surface and involve physical adsorption followed by photocatalytic reactions that convert the reactants to the final products. The mechanism relies on light absorbed by the semiconductor surface to generate e⁻–h⁺. The e⁻ then interacts with adsorbed species, degrading them into a reactive intermediate, while the h⁺ degrades a different adsorbed species into a reactive intermediate. In homogeneous photocatalysis, both the reactants and the catalysts are dissolved in the solution. The activation of the catalyst through irradiation and its interaction with the reactants occur at a molecular level [56,57].

The mechanism of photocatalytic degradation using ZnO is represented in Figure 2. When ZnO is exposed to UV radiation, e⁻–h⁺ is generated. The h⁺ from the VB can oxidize water or hydroxide ions to form ^•OH. The e⁻ represents reducing species, which can be used to reduce the ZnO CB edges to (i) generate superoxide radical ions (${\mathrm{O}}_2^{•-}$) from O₂, (ii) reduce O₂ to ${\mathrm{O}}_2^{•-}$, (iii) react with dissolved O₂ in the aquatic solution to form ${\mathrm{O}}_2^{•-}$, and (iv) transform O₂ molecules into H₂O₂. The ^•OH and ${\mathrm{O}}_2^{•-}$ play a major role in photocatalytic reactions, where the concentration of ${\mathrm{O}}_2^{•-}$ has been shown to be at least 20 times higher than the concentration of ^•OH. The high concentration of superoxide radicals in the reaction system indicates that free electrons are selectively formed on this semiconductor [58]. This suggests that in ZnO-based materials, the reduction caused by conduction band e^– dominates over oxidation reactions, especially under UV irradiation. Also, when superoxide radicals react with H₂O₂, ^•OH and singlet oxygen (¹O₂) are generated. Dominant charge carriers in ZnO-based photocatalytic processes are ^•OH, h⁺, and ${\mathrm{O}}_2^{•-}$, which play a significant role in the photodegradation of pharmaceutical compounds and other organic pollutants, transforming them into CO₂, H₂O, and inorganic ions [59]. There are several advantages to using semiconductor metal oxides for the degradation of organic pollutants compared to other oxidants. The primary advantage is the direct oxidation and degradation of harmful organic pollutants by the reactive oxygen species generated [60]. The effectiveness of ZnO is strongly dependent on operational conditions, such as pollutant structure and concentration, pH, and irradiation. These dependencies often inhibit practical applications in large-scale systems and demand further optimization of photocatalytic processes. The photocatalytic activity of ZnO in photodegradation processes has encouraged the application of ZnO photocatalysts in the synthesis of cost-effective and environmentally friendly materials for the degradation of organic pollutants in wastewater [61].

The exceptional multifunctional properties of ZnO photocatalysts result from the ability of crystallinity, morphology, and defects to control ZnO surface properties, including sensitivity, selectivity, and response/recovery time. Numerous studies [62,63,64] have reported that ZnO with different morphologies and severe defects likely exhibits photocatalytic activity in the visible spectrum. Over the past few decades, research has focused on investigating the factors related to ZnO’s performance, including crystallinity, morphology, and defects [62].

Crystallinity is one of the most overvalued properties of ZnO photocatalysts, both in the scientific field and in highly efficient and expensive applications. This property corresponds to the alignment and periodic distribution of atoms in the lattice, as well as the heterogeneities of length and size. Phenomena related to crystallinity include phase transitions within the operating temperature range, coupling between the electronic behavior of the newly synthesized film and the substrate, and the energy gaps resulting from transitions between allowed and disallowed states. As mentioned, ZnO has three crystal forms that can change the electronic, optical, and chemical properties. These properties can determine the optimal conditions to obtain the most desirable photocatalyst performance [65]. Advanced applications of ZnO, including various morphologies and structural characteristics, have created a gap between commercial oxides and the use of synthesized composites. The underutilization of commercial ZnO has contributed to this gap [66]. This is particularly relevant when keeping in mind that commercial ZnO, despite its low cost and availability, often does not provide good results due to poor morphological control and defect density [36]. In photocatalytic processes, one-dimensional order is a required property that allows for the creation of Dirac fermions, which have longer binding energies than small polarons. This property depends on structural quality characteristics. When one-dimensional order occurs, high crystalline quality is important for good photocatalytic performance. It is essential to note that commercial ZnO exhibits certain shape defects that contribute to one-dimensional order and reduce efficiency in boundary areas [67].

The optical properties of ZnO have been studied, as they are closely related to the defect structure properties of ZnO nanomaterials. The optical properties of the studied ZnO nanowires and the electrical properties of metal–semiconductor–metal UV photodetectors based on ZnO nanowires were presented. Visible photoluminescence was observed to be between 2.9 and 3.3 eV [68]. The photoemission of a nanostructured ZnO thin film was also studied by Tüzemen et al. [69]. The one-dimensional ZnO thin film showed a narrowed energy gap of 3.30 eV. The nanostructured 3D ZnO exhibited weak photoluminescence at room temperature, likely due to the recombination of localized states with the VB, which is attributed to the large surface-to-volume ratio. An ultraviolet photodetector with tailor-made ZnO nanostructures synthesized through electrodeposition was made [70]. Its optical and electrical properties had a strong relationship with its morphology. These findings confirm the strong correlation between optical behavior and nanoscale morphology. However, additional research is required to explain how specific defect types or structural parameters influence emission properties.

Charge mobility and separation of ZnO semiconductors can be affected by defects. Abundant defects in ZnO can enhance the separation of photoinduced e⁻–h⁺. The structural stability of the material can be improved by transforming the material’s defect structure into a more perfect one. A specific concentration of a single-component porogen can cause defects in the ZnO porous structure, resulting in fewer holes and electron-trapping materials. Additionally, the material can be modified by incorporating another component through solid solution processing. After obtaining a high-stability material, the transformed crystal structure can reduce the formation of surface hydroxyl groups and harmful photoreductive species. In this way, the intrinsic properties of ZnO are utilized, which improves its potential applications in photocatalysis [71]. Thus, achieving an optimal defect density remains a key challenge that requires precise synthesis control and advanced characterization techniques.

Crystallinity vs. Morphology. The growth conditions during ZnO synthesis determine its stability, morphology, shape, and defects, resulting in the synthesis of materials with different properties and performances. Physicochemical properties and characteristics depend on morphology, stability, and defects. Two factors that affect the morphology of ZnO nano-microstructures are nanoparticle morphology and crystallinity. These factors appear to have a notable influence on the properties of ZnO. By altering the shape and modifying ZnO crystals, the number and type of defects that cause intrinsic and extrinsic reactivity can be altered. Effective stabilization of crystal forms by controlling size and shape could improve ZnO activity. The application and stabilization of different ZnO crystal forms, as well as the change in the number of associated defects, require morphology control and ZnO’s compliance with specific conditions [72].

Defects vs. Performance. It is reported that an excess of defects beyond an unknown general defect limit and certain specific defects could be detrimental, reducing the lifetime and mobility of carriers, as well as the quantum yield of the corresponding processes. ZnO degrades as defects increase. However, it has not been proven that its lower performance in some processes is due to defects. Experiments show that poor-quality ZnO nanomaterials have more defects than those of higher performance. Interestingly, the higher-quality samples contain a larger functional surface area with fewer defects, which is responsible for the high performance of ZnO in various applications [71,73].

3. Strategies for ZnO Modification in Composite Photocatalysts

3.1. Metal–Oxide Composites

Integrating ZnO with other materials to boost its performance by forming heterojunctions with other semiconductors is a practical approach, as it enhances charge separation and broadens light absorption [74]. ZnO, when combined with other semiconductors, such as TiO₂, SnO₂, Fe₂O₃, CuO, Co₃O₄, and MnO₂, is among the most promising materials for photocatalytic processes. However, the synthesis of such ZnO-based photocatalysts poses some challenges, such as obtaining a desirable particle size and phase composition, which are crucial to ensuring efficient charge transfer and stability. Also, obtaining a uniform distribution of components while preventing undesirable secondary phases or agglomeration requires optimized synthesis conditions [75,76,77].

Binary heterojunctions with related band gaps, such as TiO₂/ZnO and SnO₂/ZnO, exhibit superior photocatalytic effects compared to individual materials like TiO₂, SnO₂, and ZnO. This improvement is due to their ability to enhance the lifetime of photo-excited e⁻–h⁺, facilitate interfacial charge transfer to adsorbed substrates, and improve the photostability of the hybrid nanocomposites [78,79]. Coupling ZnO with semiconductors with different band gaps, such as Fe₂O₃, CuO, and MnO_2, induces a shift of the band gap absorption to longer wavelengths and is beneficial for inhibiting photon-induced e⁻–h⁺ recombination [80,81]. However, although these systems show promising properties, their efficiency may vary depending on the synthesis method, interface quality, and environmental conditions [82].

The specific type of heterojunction formed depends on how the energy bands of the materials align with each other. It is crucial to understand these alignments, as they determine the mechanisms of charge separation, recombination, and photocatalytic activity. Heterojunction architectures, such as types I, II, and III, utilize different band alignments to control charge transfer, recombination, and performance. The Co₃O₄/ZnO composite belongs to the type I heterojunctions (Figure 3a), which have straddling band structures that trap e⁻–h⁺ in one material, limiting charge movement due to the lack of a built-in electric field. While they are easy to synthesize, they exhibit low efficiency in photocatalytic applications [83]. In contrast, type II heterojunctions (Figure 3b), such as those in ZnO/In₂O₃ photocatalysts, feature staggered band alignment that promotes effective charge separation. Electrons and holes are spatially separated between two materials, which enhances carrier lifetimes and improves photocatalytic performance [84]. In type III or broken gap heterojunctions, the energy bands of the two semiconductors do not overlap, making efficient charge transfer and separation nearly impossible. As a result, current research on photocatalytic material enhancement primarily focuses on type I and type II heterojunctions due to their more favorable charge dynamics [85]. A p-n heterojunction (Figure 3c) formed between p-type oxides, such as ZnO and CuO, generates a built-in electric field at the interface due to Fermi level equilibration. Electrons flow from the n-type to the p-type material, while holes move in the opposite direction, creating a depletion region. This built-in field promotes effective charge separation and reduces recombination, thereby enhancing overall efficiency [86]. The Z-scheme heterojunction (Figure 3d) pairs ZnO, an oxidation semiconductor with a low VB position, with semiconductors with high CB gap positions, such as CeO₂. When exposed to light, the electrons in the CB of the ZnO recombine with the holes in the VB of the CeO₂. This process creates a Z-shaped flow of charge. As a result, highly reactive e⁻–h⁺ is left in the ZnO and the CeO₂, respectively, which helps to maintain a strong redox potential [87]. The S-scheme heterojunction (Figure 3e) couples a reduction photocatalyst with a higher Fermi level, such as CdS, with an oxidation photocatalyst with a lower Fermi level, like ZnO. The advanced composite modifications lead to the synthesis of highly active photocatalysts, but these methodologies are still in the stage of evolving.

When they come into contact, electron transfer occurs until Fermi levels align, generating a built-in electric field and band bending at the interface. Under radiation exposure, electrons in the ZnO’s CB recombine with holes in the CdS’s VB through this field, leaving behind strong redox-active electrons in the CdS and holes in the ZnO. This selective recombination boosts charge separation and enhances photocatalytic efficiency. ZnO is sometimes employed as a reducing agent to create an S-scheme heterojunction with semiconductors, such as WO₃, Bi₂O₃, CeO₂, and TiO₂ [88,89].

3.2. Carbon-Based Composites

Over the last decade, carbon nanostructures have demonstrated their significant role in advancing nanocomposites. Their integration of ZnO and carbon nanostructures has enhanced the intrinsic properties of ZnO. This revealed new properties essential for environmental and energy-related applications. Research has shown that the heterojunction of ZnO with carbon materials enhances the performance of nanocomposites in various applications, including photocatalysis [90], gas sensing [91], photovoltaic devices [92], and fuel cells [93]. In particular, carbon-based/ZnO photocatalysts significantly enhance photocatalysis due to their synergistic combination of ZnO’s semiconducting nature and carbon’s high surface area, excellent conductivity, remarkable chemical stability, exceptional mechanical strength, and environmental friendliness and availability [94]. These properties improve photocatalytic efficiency and extend the practical usability and lifespan of ZnO-based photocatalysts. Based on carbon dimensionality, carbon-based/ZnO photocatalysts can be classified into four types: 0D, 1D, 2D, and 3D [95].

Graphene, an example of a 2D carbon-based nanomaterial, has been of considerable interest due to its excellent physicochemical characteristics compared to other carbon materials [96]. Graphene is a sp²-hybridized hexagonal crystalline material in a single-atom-thick sheet [97]. Two key strategies have been developed to create graphene-based/ZnO photocatalysts, ZnO particles coated with graphene and graphene layers enhanced with ZnO, both aimed at achieving optimal photocatalytic properties [98]. Recent studies have demonstrated that the graphene–ZnO nanocomposite can degrade 74% of methylene blue dye after 180 min of exposure to solar irradiation. This effectiveness is attributed to graphene acting as an electron mediator and transporter, which improves charge carrier dynamics and inhibits recombination (Figure 4) [99]. Further studies have shown that ZnO–graphene nanocomposites possess a porous web structure, which enhances the adsorption and mass transfer of dyes and oxygen. This exceptional photocatalytic activity is also attributed to the lower band gap energy of the prepared nanocomposite. Graphene facilitates the separation of photogenerated charge carriers and reduces charge carrier recombination, which ultimately leads to an improvement in photocatalytic activity [100].

Carbon nanotubes (CNTs), which are considered 1D carbon-based nanostructures, are another excellent material for generating ZnO photocatalysts with enhanced catalytic properties. Outstanding electrical properties and the hollow tubular structure of these materials create a conductive framework that improves ZnO’s photocatalytic activity. Also, the significant electrical and conductive characteristics enhance the composites’ ability to adsorb contaminants [90] physically. CNTs are cylindrical structures composed of rolled graphite sheets, with diameters ranging from 3 nm to 30 nm and lengths extending to several millimeters. There are three main types of CNTs: single-walled, double-walled, and multi-walled nanotubes [101]. Research has shown that the corrosion resistance of CNTs helps to prevent the photo-corrosion of ZnO, thereby enhancing its photocatalytic efficiency [102].

Additionally, the high aspect ratio and large specific surface area of CNTs enhance the adhesion of ZnO on their surface with increased CNT loading, which helps to prevent the formation of aggregates [103]. However, at higher concentrations, the CNTs may begin to aggregate, which can reduce their ability to interact uniformly with the ZnO matrix. Incorporating excessive amounts of CNTs can lead to saturation of the active sites, resulting in a decrease in both photocurrent and photoresponsivity, which highlights the importance of optimizing the CNT concentration [104].

Carbon Quantum Dots (CQDs) are 0D carbon-based nanomaterials with unique structural and electrical features that distinguish them from other quantum dot families. CQDs are monodisperse and usually spherical particles primarily composed of sp²/sp³ carbon, with diameters of less than 10 nm [105]. Their surface defects create a distinct up-converted photoluminescence effect that can transform ultraviolet light into visible light. This property enables CQDs to broaden the light absorption spectrum of ZnO-based photocatalysts, enhancing solar utilization efficiency [105]. For instance, CQDs have been demonstrated to enhance the performance of ZnO-based photoelectrodes by improving visible light absorption and reducing electron–hole recombination. They act as electron shuttles by accepting photogenerated electrons from ZnO and rapidly transferring them to another location. This process enhances charge separation and prevents the recombination of e⁻–h⁺, thereby improving photocarrier transport [106]. The photocatalytic behavior of ZnO photocatalysts can be significantly tailored through modification with carbon-based materials. The fine-tuning of charge dynamics, light absorption, and surface interactions is enabled by their structural diversity, ranging from 0D CQDs to 2D graphene. A critical assessment of the current literature [17,107,108] shows that graphene provides high-level charge mobility and large surface contact, while CNTs enable structural integrity and electron conduction paths and CQDs contribute to band gap modulation and spectral sensitivity. Despite these promising advances, challenges, such as reproducibility, interfacial engineering, and optimal carbon loading, remain underexplored and require further investigation [106].

3.3. Metal and Non-Metal Doping

One practical and commonly used approach to enhance ZnO’s photocatalytic activity is the doping of metals and non-metals [109]. Doping is a fundamental process in materials engineering that involves adding foreign atoms, metallic or non-metallic, into the lattice structure of a host material to create controlled changes in the material’s electronic, optical, or structural characteristics. Incorporating these dopants through substitution or interstitial placement disrupts the regular periodicity of the ZnO lattice, leading to localized charge imbalances, lattice strain, or the introduction of new energy states within the band structure [74]. These modifications change the intrinsic properties of ZnO and tailor it for different photocatalytic applications. As previously mentioned, a significant drawback of using ZnO semiconductors as photocatalysts is their low charge separation efficiency and wide band gap. To address this issue, doping has been utilized to raise the VB energy or decrease the CB energy of ZnO, thereby narrowing the band gap energy to the ultraviolet-visible region (Figure 5) [110]. The photoactivity of doped ZnO depends significantly on the preparation method, the nature of the doped ions, and their concentration [111]. In situ techniques, such as sol–gel synthesis [111,112,113], hydrothermal methods [114,115], or co-precipitation [116,117], have been employed to facilitate the simultaneous incorporation of dopants during the formation of ZnO.

Compounds like transition metals have shown advantages in tuning the electronic band structure of ZnO in photocatalytic applications by introducing localized states within the band gap. Metal dopants, such as Mn, Fe, Co, Ni, and Cu, have been used to decrease the band gap energy of ZnO photocatalysts by introducing new energy levels within the ZnO band gap. This led to the photocatalytic activation of ZnO in the visible light range [118]. Additionally, it was demonstrated that the incorporation of Fe resulted in the suppression of charge recombination within the ZnO composite by efficiently trapping CB electrons [119]. It was reported that visible-light-active Cu-doped ZnO has enhanced photocatalytic activity attributed to Cu-induced oxygen vacancies [120]. These oxygen vacancies, as a typical type of point defect, lead to enhanced light absorption, decreased charge carrier recombination, and increased active site availability [121]. The improved performance of Ag-doped ZnO in comparison to undoped ZnO can be attributed to the increased surface area, porosity, and surface morphology of the ZnO nanostructures when Ag is incorporated into its lattice structure [122]. The band structure modification enables visible-light absorption, a key aspect for efficient solar-driven photocatalysis.

Although metal dopants offer many advantages, non-metal dopants provide a lower cost than metal dopants [123]. Non-metallic ion doping of ZnO involves replacing oxygen vacancies in ZnO with non-metallic ions or introducing additional oxygen vacancies to create defects. This process reduces the energy band gap and enhances the optical response region, especially in the visible spectrum [124]. Similarly to metal dopants, non-metals, such as N, S, and P, can shift the band gap of ZnO [125,126]. These atoms can easily replace oxygen atoms in the ZnO lattice, further narrowing the band gap and decreasing the energy required for light absorption [127]. It was shown that P doping inhibits electron–hole recombination by introducing P⁵⁺ ions into the lattice, which act as electron traps, reducing recombination. Additionally, P doping increases surface oxygen concentration, where interstitial oxygen atoms near P act as polaronic electron traps, further suppressing recombination. Furthermore, S- and N-doped ZnO photocatalysts had better photoactivity than pure ZnO due to their lower crystallite size, lower band gap energy, and large pore size distribution [126]. The primary concern is the stability of these systems under irradiation, as well as their reusability.

As previously discussed, non-metal and metal doping has been widely used to improve the efficiency of ZnO photocatalysts. Researchers examining the influence of dopant concentration showed that increasing the dopant concentration enhances the photocatalytic activity of ZnO in degrading pollutants in aquatic environments [128,129]. However, recent research has suggested that excessive doping may reduce the photodegradation efficiency of ZnO [130,131]. This reduction is likely due to the creation of trapping sites for e⁻–h⁺, obstructing active sites and producing larger aggregated particles [132,133]. Additionally, despite their potential, doping ZnO has limitations, such as poor reaction selectivity, difficulties in achieving consistent synthesis, and increased production costs of synthesis methods. Therefore, it is challenging to synthesize ZnO nanomaterials that will degrade and mineralize organic pollutants with high efficiency and selectivity at low costs [134].

3.4. Metal–Organic Framework and Polymer Composites

Fixation of ZnO in a matrix material is an attractive approach in the fundamental search for advanced photocatalysts suitable for water remediation applications [135]. This type of modification addresses significant limitations of ZnO, such as particle aggregation, photo-corrosion, and low recyclability. MOFs and polymer composites have emerged as innovative matrix materials that combine with ZnO to enhance its properties. This combination leverages the synergistic effects of both ZnO and the composite matrix to effectively address environmental issues [136,137].

MOFs are porous, crystalline materials that can be structured in one, two, or three dimensions by connecting metal centers or clusters to organic linkers [138]. One of the most remarkable features of MOFs is their easily tunable pore size, which can range from meso- to micro-scale by modifying the organic and inorganic components [139]. Additionally, they have a high surface area, good chemical reactivity, and diverse structural possibilities [140]. Despite the advantages offered by MOFs, they have certain limitations in catalytic reactions. One significant drawback is the low concentration of active metal species within MOFs [141]. The introduction of active metal species, such as ZnO, through in situ MOF growth on a ZnO substrate has been reported as a way to enhance the performance of both MOF and ZnO. Incorporating zeolitic imidazolate framework ZIF-9, a type of MOF, significantly enhanced the band structure of ZnO, and the efficiency of e⁻–h⁺ separation was effectively enhanced during the photocatalytic degradation reactions [142]. Another way of forming ZnO–MOF hybrids is in situ ZnO growth on MOF structures, where ZnO is either surface-anchored or pore-encapsulated. The deposition of ZnO depends on its particle size and the diameter of the MOF pores. The excellent performance of the composites produced through this method is mainly due to the MOF’s ability to prevent the agglomeration of ZnO nanoparticles. [143]. In this way, MOFs can function as structural supports and pollutant pre-concentrators, increasing local reactant concentrations and improving degradation kinetics [144]. Yang and coworkers reported a new strategy to synthesize ZnO/ZIF-8 hybrid photocatalysts through partial framework transformation. In this approach, porous ZIF-8 can absorb organic molecules on its surface and within its pores, leading to the formation of a high-concentration layer of organic pollutants, wherein ZnO functions as a photocatalyst to degrade these compounds [145].

Polymers, having a wide variety of forms and properties, are currently being used to improve ZnO’s photocatalytic activity [146,147,148]. Di Mauro and coworkers reported excellent photocatalytic performance of poly(methyl methacrylate)/ZnO composite films in the removal of dyes and phenols from water. The presence of a polymer decreased the photo-corrosion of ZnO during the photocatalytic process. It lowered the cost by eliminating the need for a filtration step to remove particles from the water [149]. Polyvinyl chloride effectively modified the band gap of ZnO, enhancing its photocatalytic efficiency in dye degradation.

Furthermore, incorporating the polymer was shown to reduce the recombination of photogenerated charge carriers in ZnO [150]. Another beneficial property of introducing polymers into ZnO nano-photocatalysts is recyclability. Borjigin and coworkers showcased this recyclability through swelling tests and by conducting repeated degradation measurements using the same sample. After three cycles of photocatalytic reactions, the composites continued to exhibit excellent catalytic activity [151]. Both MOFs and polymers combined with ZnO show great potential as next-generation photocatalysts. However, challenges remain in achieving scalable, cost-effective synthesis methods and ensuring long-term durability under real-world conditions. Therefore, additional research is crucial.

4. Modeling of ZnO-Based Materials for Photocatalysis Based on Atomistic Calculations

The growing need for efficient and sustainable photocatalysts has led to increased use of atomistic modeling techniques in the development of ZnO-based materials. These simulations, grounded in quantum mechanical principles, such as DFT, enable researchers to explore how atomic-scale structural, electronic, and optical properties influence photocatalytic activity. By simulating doping effects, surface reactivity, defect formation, and band structure modifications, atomistic methods provide deep insights that are often difficult or costly to obtain experimentally. These simulations enable rapid hypothesis testing and screening of testing materials before synthesis, therefore accelerating their discovery and reducing experimental trial, cost, and error. These calculations serve as a powerful tool for predicting material behavior, guiding experimental efforts, and rationally designing enhanced ZnO-based photocatalysts.

4.1. Atomistic Calculations in Materials Science—Methodological Considerations

4.1.1. Overview of Atomistic Modeling Approaches

Atomistic calculations encompass a broad range of computational techniques that model materials at the scale of individual atoms and electrons, offering critical insights into their structural, electronic, and dynamic properties [152,153,154,155]. At the lower end of this spectrum are classical approaches, such as molecular mechanics (MM), which treat atoms as point masses and interactions using empirical force fields to approximate bond stretching, angle bending, torsional rotations, and non-bonded forces [156]. These methods enable the simulation of large systems with low computational cost but lack explicit treatment of electronic structure. At the higher end are quantum mechanical methods, including wavefunction-based approaches (e.g., Hartree–Fock, MP2, CCSD) and the widely used DFT, which balances computational efficiency and accuracy by modeling the electron density rather than many-electron wavefunctions. These quantum mechanical methods are essential for predicting electronic, optical, and catalytic properties without empirical parameters [157,158,159,160,161,162,163]. To explore time- and temperature-dependent behavior, molecular dynamics (MD) simulations can be performed using either classical force fields or quantum-based potentials, providing valuable insight into diffusion, thermal stability, and surface reactivity [164,165]. Collectively, these atomistic methods form a versatile toolkit for understanding and predicting the behavior of materials in photocatalysis. Each method has its advantages, and their combined use enables multi-scale modeling approaches that bridge the gap between predictive models and their real-world applications.

Between classical force field-based techniques and fully ab initio quantum methods lies a valuable class of semiempirical methods, which offer a compromise between computational speed and quantum-level accuracy [166,167]. These methods incorporate empirical parameters derived from experimental data or higher-level calculations to simplify the quantum mechanical treatment of electronic interactions. Popular examples include PM6, PM7 [168,169,170,171,172,173,174], and Grimme’s xTB (extended tight-binding) family of methods, such as GFN1-xTB and GFN2-xTB [175,176,177,178], which are specifically designed for efficient and reliable modeling of large molecules, nanomaterials, and condensed-phase systems. Just recently, Grimme’s group launched the next generation of the semiempirical extended tight-binding method, known as g-xTB [179]. This newly introduced method is a successor of GFN2, offering improved accuracy and robustness while introducing only a modest increase in computational cost, approximately 30% to 50% higher than GFN2. Despite this increase, g-xTB remains remarkably fast compared to conventional DFT methods. Semiempirical approaches are beneficial for pre-optimizing geometries in the early stages of materials screening, exploring conformational space, or estimating key electronic descriptors, like HOMO–LUMO gaps, dipole moments, and reactivity indices, at a fraction of the cost of DFT. In the study of photocatalysts, including ZnO-based systems, these methods can accelerate high-throughput screening, serve as a bridge toward more accurate DFT refinement, or provide input features for data-driven models. Their scalability and growing accuracy make them an increasingly important component of the computational toolkit in materials design and photocatalysis research. Despite their growing usability, semiempirical methods rely on the quality of the parameters and may lack transferability between chemical environments and material types.

In recent years, ML has introduced a transformative approach to atomistic modeling by enabling the prediction of material properties with accuracy approaching that of quantum mechanical methods but at a fraction of the computational cost [180,181,182]. Unlike traditional methods that solve complex equations derived from quantum mechanics (e.g., the Schrödinger equation), ML-based models use pre-trained functions that map atomic structures directly to properties like total energy, forces, band gaps, or adsorption energies. These models are trained on large datasets generated from high-level quantum mechanical (QM) calculations, allowing them to learn the underlying physics without performing explicit computations each time. As a result, once trained, ML models can evaluate millions of structures orders of magnitude faster than conventional DFT or wavefunction methods. This speed makes them ideal for high-throughput screening, molecular dynamics simulations, and real-time material discovery. However, their accuracy depends heavily on the quality and diversity of the training data, and they may struggle with transferability beyond their training domain [183,184]. Despite these limitations, ML-based atomistic models are rapidly becoming indispensable in materials science and photocatalysis, complementing traditional methods rather than replacing them and unlocking new scales of efficiency and insight. In this review, we will mention some specific models and tools obtained through ML methods.

In the context of photocatalysis, the application of atomistic calculations enables researchers to probe the fundamental mechanisms underlying light absorption, charge separation, surface reactions, and catalytic activity, all at the atomic and electronic levels. For example, DFT can be used to compute band structures, density of states, work functions, and adsorption energies, offering predictive insights into how doping, defects, and surface terminations influence photocatalytic efficiency. These calculations are especially valuable for ZnO-based photocatalysts, whose performance is susceptible to structural modifications, such as substitutional doping, vacancy formation, or heterojunction engineering. Atomistic simulations also aid in screening new materials before synthesis, reducing the reliance on costly and time-consuming experiments. By identifying promising compositions, surface orientations, or nanostructures, computational studies help accelerate the discovery of next-generation photocatalysts with enhanced light-harvesting capabilities, improved stability, and greater environmental compatibility. The predictive power of these models is very high, as the approximations and assumptions involved highlight the need for ongoing validation of experimental data. Furthermore, these insights often guide experimental design, leading to more focused and effective material development efforts.

4.1.2. Molecular vs. Periodic Modeling in ZnO Systems

In the context of atomistic modeling, it is essential to acknowledge that these methods have been developed for both molecular and periodic systems, each offering distinct advantages. Molecular-level calculations, such as molecular DFT, are particularly valuable for studying the local reactivity of individual molecules, providing insights into stability, reactive sites, and degradation pathways of pollutants or intermediates involved in photocatalytic processes. In contrast, periodic calculations, designed for crystalline or extended systems, enable the investigation of key material properties, such as band structure, density of states, and optical absorption, which are essential for understanding light–matter interactions. In photocatalysis, these features are directly linked to one of the most fundamental properties: the band gap, which determines a material’s ability to absorb and utilize sunlight for catalytic reactions. In Figure 6, we illustrate the periodic and nanocluster ZnO structures of its zinc blende form, highlighting how different modeling frameworks can be applied depending on the scale and objective of the investigation.

However, the boundary between molecular and periodic modeling is often more flexible than it appears. Researchers frequently adopt hybrid or approximate strategies that use features from both approaches to balance computational efficiency and accuracy. In some instances, molecular calculations can be used to approximate periodic systems, especially when simulating clusters of crystalline material. For example, ZnO nanoparticles or nanosheets can be modeled as finite ZnO clusters with fixed atomic positions or appropriately saturated edge atoms, enabling quantum-level insights into local bonding environments, surface states, and defect behavior. This approach allows researchers to gain relevant information about extended materials while maintaining the computational efficiency and methodological simplicity of molecular quantum chemistry packages.

Conversely, molecules can also be subjected to periodic boundary conditions, especially when studying adsorption phenomena and molecule–surface interactions or when embedding a molecule in a dielectric or bulk-like environment. By placing a molecule in a sufficiently large vacuum box within a periodic unit cell, researchers can simulate isolated systems while leveraging the benefits of periodic DFT implementation, which often incorporates advanced features, such as plane-wave basis sets, k-point sampling, and efficient parallelization.

By integrating molecular and periodic simulations, more robust and predictive insights can be obtained. This interchangeability and complementarity between molecular and periodic approaches expands the toolkit available to materials scientists working on photocatalysis. It allows them to tailor their models according to the nature of the problem, desired accuracy, and available computational resources, thereby enhancing the scope, depth, and flexibility of atomistic simulations in the design and understanding of ZnO-based and other photocatalytic materials.

4.1.3. Selection of Density Functionals and Band Gap Corrections in DFT

When discussing methodological setups in atomistic simulations applied to materials science, several important considerations must be highlighted. DFT has become the most widely adopted approach, primarily due to its favorable balance between computational cost and accuracy [185,186]. The versatility of DFT makes it applicable to a wide range of systems, from molecules and clusters to bulk crystals and interfaces [187]. However, DFT is not without limitations. A key challenge is the absence of a universal exchange–correlation functional that performs consistently well across all types of materials and systems. As a result, a wide variety of density functionals have been developed, each optimized for specific properties, bonding environments, or classes of materials. While this diversity allows for flexibility and tailored modeling, it also complicates the process of method selection, particularly for newcomers to the field. Choosing the appropriate functional requires a careful evaluation of the system under study and the properties of interest.

When using DFT to compute electronic properties, such as band structures, one must carefully consider its well-known limitations, particularly in the field of photocatalysis, where the band gap plays a decisive role in determining a material’s photocatalytic potential. Conventional functionals, most notably the Perdew–Burke–Ernzerhof (PBE) [188] functional under generalized gradient approximation (GGA), tend to systematically underestimate band gap values [189,190,191], leading to inaccurate predictions of photoactivity thresholds and misalignment with redox potentials.

To overcome this challenge, two main strategies have emerged. The first involves the development and use of more sophisticated exchange–correlation functionals, such as hybrid functionals, which incorporate a portion of exact Hartree–Fock exchange to improve accuracy. A prime example is the Heyd–Scuseria–Ernzerhof (HSE06) functional [192], which has demonstrated excellent performance in predicting electronic band gaps across a broad spectrum of materials. However, this increased accuracy increases computational cost, which may limit practical application in high-throughput screening or large-scale interface modeling.

The second strategy centers on band gap correction techniques applied to standard functionals. These include methods like DFT+U [193,194], which introduces a Hubbard-like correction to better describe localized electrons, particularly in systems involving transition metal oxides, and is widely used for its efficiency and simplicity. Herein, we illustrate the case of the band structure of the wurtzite form of ZnO (Figure 7).

For obtaining the band structures presented in Figure 7, we used Quantum Espresso 7.2 with the following computational details: PBE functional and GBRV [195] pseudopotentials with 40 Ry and 200 Ry cutoffs for the wavefunction and charge density, respectively. For Brillouin zone sampling, we used a grid placement distance of 0.05 Å, while all other values were set to default. As shown in Figure 7, uncorrected DFT calculations result in a severe underestimation of the band gap. Specifically, instead of the expected band gap value exceeding 3 eV, the calculation yielded a value of only 0.8 eV. This discrepancy highlights a fundamental limitation of the standard GGA functional in describing excited-state properties and semiconductor band gaps [196,197]. However, the inclusion of the DFT+U method [166,167] with Hubbard potential of 10 eV for the Zn 3d and 6 eV for the O 2p states (according to the work of Goh et al. [198]) significantly improved the results, producing a band gap in excellent agreement with experimental data with minimal additional computational cost. However, this method presents a challenge related to the fact that the values of the Hubbard potential are not known in advance for all materials and certain efforts must be invested in identifying the values that lead to an accurate band structure.

Another notable correction approach is DFT-1/2 [199,200,201], which addresses self-interaction errors by applying a pseudo-self-energy correction to atomic potentials. This method improves the estimation of the band gap without significantly increasing computational demands. What makes it attractive is that there is no need to try different values of parameters, such as Hubbard potentials. Instead, it works practically out of the box, making it an attractive alternative for high-throughput studies. Based on the dopant interactions, phase-dependent electronic behavior, and surface defects, its resilience is crucial for ZnO-based photocatalysts. In practice, researchers often strike a balance between computational cost and accuracy. For routine screening or structural studies, the PBE functional remains a practical choice, especially when combined with DFT+U or DFT-1/2 adjustments. For applications requiring higher accuracy, such as evaluating electronic transitions, photocatalytic thresholds, or defect levels, hybrid functionals like HSE06 are preferred, despite their increased computational requirements. These flexible strategies enable the tailoring of computational protocols to meet the precision needs of specific investigations involving ZnO-based photocatalysts.

4.1.4. Van Der Waals Forces and Noncovalent Interactions

In addition to the well-documented challenges of functional selection and band gap underestimation in DFT, another significant limitation arises from its insufficient treatment of noncovalent interactions. Standard DFT functionals, particularly those frequently used, such as PBE and B3LYP [202,203,204,205], often fail to account for long-range dispersion forces, commonly referred to as van der Waals (vdW) interactions. These forces play a crucial role in determining the structural, energetic, and dynamic behavior of molecular crystals, layered materials, adsorption processes, and biomolecular systems.

To overcome this deficiency, several correction schemes and specialized functionals have been developed. Among the most widely adopted approaches is Grimme’s empirical dispersion correction, the so-called “D” scheme, which introduces additive correction terms to the DFT total energy [206,207]. The “D” scheme has become a powerful tool due to its ease of integration with various exchange–correlation functionals and its negligible computational overhead. This method is both efficient and easy to implement and has demonstrated substantial improvements in the description of binding energies, geometries, and weak intermolecular forces in diverse chemical systems.

Over time, Grimme’s DFT-D approach has been improved (e.g., D2, D3, and the more recent D4 [207,208,209,210,211]), and, from the perspective of development, it is important to mention that it can be seamlessly combined with various functionals (e.g., PBE-D3, B3LYP-D3), allowing users to retain the favorable aspects of the base functional while enhancing performance in noncovalent regimes.

In parallel, functionals have also been specifically developed to account for dispersion interactions intrinsically. A typical example is the Minnesota family of functionals, where the goal was to incorporate medium-range dispersion through empirical fitting of a large number of functional parameters [212]. One of the best-known functionals from these efforts is M06-2X for the main-group periodic system of elements [213,214,215,216].

Choosing between empirical corrections, such as D3, and specially developed functionals depends on the specific requirements of the simulation. Empirical corrections tend to offer greater computational efficiency and flexibility, while specially designed functionals provide a more rigorous but computationally more expensive description of dispersion interactions.

For ZnO-based photocatalysts and similar systems where weak intermolecular forces or molecule–surface interactions are essential, such as in dye-sensitized solar cells, gas adsorption, or hybrid organic–inorganic interfaces, explicitly accounting for dispersion effects is essential to obtain accurate structural and energetic predictions. Incorporating these corrections leads to more realistic simulations of adsorption geometries, reaction intermediates, and modeling of surface-bound species.

4.2. Applications of Atomistic Calculations in ZnO-Based Materials for Photocatalysis

Various computational tools can be employed to investigate key properties that govern photocatalytic behavior, including geometry optimization, electronic structure, charge dynamics, and light–matter interactions. Below, we summarize the main categories of insights that can be obtained through atomistic modeling.

Structural optimization is a foundational step in most DFT studies. It involves relaxing atomic positions and, if necessary, lattice parameters to find the energetically most favorable configuration of ZnO, whether in bulk form, as a surface, or as a doped or defective nanostructure. Accurate geometries are essential for meaningful predictions of electronic and optical properties.

Adsorption energy calculations allow researchers to explore how reactants (e.g., water, O₂, dyes, or pollutants) interact with ZnO surfaces or doped variants. Using high-level methods, such as hybrid functionals (e.g., HSE06), researchers can predict adsorption strength and the likelihood of surface-mediated reactions, which is critical for understanding catalytic mechanisms.

Electronic structure analysis, including band structure and density of states (DOS), is key to evaluating the light absorption capacity and electronic transitions within ZnO. DFT and especially hybrid DFT methods can predict how doping or surface functionalization shifts the band edges, which is essential for aligning the conduction and VB with redox potentials relevant to water splitting or dye degradation. However, the choice of a functional significantly affects the results, introducing uncertainty.

Charge distribution and transfer studies, such as Bader charge analysis or electrostatic potential mapping, help identify how charge is localized or transferred within the material or between the catalyst and the adsorbates. These insights are crucial for evaluating the efficiency of charge separation and transport, which directly impact photocatalytic activity.

Predictions of optical properties using time-dependent density functional theory (TD-DFT) or dielectric function analysis can reveal absorption spectra and optical transitions. These simulations are particularly valuable for assessing how modifications to ZnO (e.g., through doping or nanostructuring) affect its ability to harvest visible light.

Defect formation and doping behavior play a major role in tuning the photocatalytic efficiency of ZnO. Atomistic calculations can determine defect formation energies, preferred charge states, and transition levels within the band gap. These parameters help predict which intrinsic or extrinsic defects are thermodynamically favorable under different synthesis conditions. For instance, oxygen vacancies, zinc interstitials, or substitutional dopants (e.g., N, S, or transition metals) can significantly alter the electronic structure, create mid-gap states, or facilitate charge separation, thereby modifying the photocatalytic response. DFT-based studies help clarify these effects and guide rational strategies for defect or dopant engineering.

Surface reactivity and catalytic mechanisms can also be explored using atomistic simulations. ZnO surfaces, especially under photocatalytic conditions, often undergo restructuring, hydroxylation, or adsorption of intermediate species. Using DFT in combination with nudged elastic band (NEB) or transition state search methods, researchers can map reaction energy profiles, identify activation barriers, and evaluate plausible reaction pathways for key processes, such as water splitting, dye degradation, or CO₂ reduction. This provides a molecular-level understanding of catalytic steps, including charge transfer, bond breaking, and recombination dynamics.

Thermal stability and dynamic behavior are increasingly being studied using ab initio molecular dynamics (AIMD) or semiempirical dynamics (e.g., with xTB or ReaxFF). These simulations enable the exploration of how ZnO nanostructures, doped surfaces, or interfaces behave under finite temperatures, pressures, or solvent exposure. AIMD simulations can capture temperature-induced bond fluctuations, migration of surface species, or structural degradation, which are critical for assessing the robustness of photocatalysts under real-world conditions.

4.3. Tools for Atomistic Calculations

Physicochemical concepts and target systems at the atomic level can be connected using atomistic computational tools via a computational platform. Each tool supports appropriate computational methods or enables specific computational functions. Modeling materials, molecules, and other atomic-scale systems and users preferences can be achieved using these packages. Expert knowledge of the computational methods and a vast understanding of the expected properties of the results are necessary in order to identify appropriate modeling tools [152,217]. With rapid advances in computational modeling and materials science, a vast number of software tools have become available—many of which are now optimized to run efficiently, even on standard desktop machines. These tools can be broadly classified into three categories, open source, free for academics, and commercial software, each offering distinct benefits depending on user needs and resources.

The most essential type of category when it comes to scientific software is open source, which provides free access to both the binaries and the source code, allowing for full customization and transparency. This category is widely adopted in academic research due to its flexibility and active user communities. When it comes to calculations on periodic structures, a notable example is Quantum Espresso (latest version 7.4.1) [218,219,220,221]. For calculations on molecules, notable examples include PSI4 (latest version 1.9.1) [222,223,224], Multiwfn (latest version 3.8) [225,226,227,228], and xTB (latest version 6.7.1) [175,176,177,178].

Quantum ESPRESSO is particularly relevant for ZnO-based materials, as it supports periodic DFT calculations and allows users to obtain band structures, density of states (DOS), optical absorption spectra, and charge distribution maps, all of which are crucial for photocatalysis studies. PSI4 offers a range of high-accuracy quantum chemical methods, while Multiwfn serves as an advanced wavefunction analysis toolkit capable of postprocessing a variety of electronic structure outputs. The xTB code, developed by Grimme’s group, is a popular semiempirical method tailored for large-scale screening and pre-optimization of molecules and clusters, including ZnO nanostructures. It also supports molecular dynamics (MD) simulations via its GFN-FF force field, an efficient classical force field tuned for broad chemical applicability.

Still in the area of open source software but for force-field-based simulations, LAMMPS (latest version 2025) [160,161] and GROMACS (latest version 2025.2) [162,163,164,165,166,167,168,169] are two of the most widely used MD engines. LAMMPS is especially suited for inorganic and hybrid systems, including materials like ZnO, offering broad support for various interaction potentials. GROMACS, designed initially for biomolecules, is also used in hybrid materials research, including photocatalytic water decontamination [229].

Another important class of software types comprises the so-called free for academic tools, which are available at no cost to academic users. These tools often come with powerful capabilities and extensive documentation, even if their source code remains closed. Among the most prominent is ORCA (latest version 6.1.0) [170,171,172,173,174,175,176,177], a quantum chemistry suite that supports a wide range of DFT and wavefunction-based methods, as well as time-dependent DFT (TD-DFT) and hybrid functionals. ORCA is particularly effective for modeling ZnO clusters or doped systems at the molecular level, especially where localized states or detailed optical transitions are of interest. The latest ORCA versions offer significantly improved performance and scalability, making them well-suited for advanced electronic structure investigations.

In recent years, a unique class of modeling tools has emerged, offering unprecedented simplicity and accessibility: online modeling platforms that require no local installation or maintenance. These platforms allow all computational tasks to be performed directly within a web browser, providing exceptional flexibility and ease of use. Among them, only one is currently free for academic users—atomistica.online (https://atomistica.online, accessed on 1 August 2025) [230,231]. This online platform is freely available upon registration. Atomistica.online enables users to perform semiempirical calculations entirely through their web browsers, supporting engines like xTB, g-xTB, MOPAC, and SHERMO [232]. The platform also offers additional utilities, including the generation of MD input systems using Packmol, construction of ionic liquid ion pairs, conformer generation using RDKit (latest version 2025.03.3), and access to several powerful features from Multiwfn. These include the creation of Reduced Density Gradient (RDG) scatter plots and the generation of cube files for visualizing RDG isosurfaces. The platform also offers the possibility of predicting properties of ionic liquids based on imidazolium cation using newly developed ML models (IonIL-IM-D1, IonIL-IM-D2, and IonIL-IM-V). All of these features are fully integrated into a browser-based interface, offering a seamless and efficient modeling experience for both novice and expert users. In its latest developmental phase, atomistica.online also offers calculators based on Meta’s Open Molecules 2025 (OMol25) and universal model for atoms (UMA) models.

Commercial software packages are full-featured suites designed for industrial and advanced academic research, typically offering premium features, user-friendly graphical interfaces, and dedicated support. Examples include the Materials and Biologics Science Suites by Schrödinger Inc. (New York, NY, USA, https://www.schrodinger.com), ADF by Software for Chemistry & Materials (Amsterdam, The Netherlands, https://www.scm.com), and QuantumATK by Synopsys (Sunnyvale, CA, USA, https://www.synopsys.com/manufacturing/quantumatk.html, accessed on 1 August 2025). These platforms offer a wide range of functionalities—from periodic and molecular DFT to MD simulations, force field generation, and sophisticated postprocessing tools. Schrödinger’s Materials Science Suite integrates high-quality DFT (via Quantum ESPRESSO), MD (via Desmond), and GUI-based workflow design. ADF includes the BAND module for periodic DFT calculations, while QuantumATK supports atomistic modeling with integrated builders, band structure tools, and ML-ready descriptor generation. When speaking of online modeling platforms that are commercially available, it is a must to mention Rowan (Boston, MA, USA, https://rowansci.com) and Samson (Grenoble, France, https://www.samson-connect.net).

Each category of software offers unique advantages. Open source tools promote transparency, adaptability, and collaboration. Free for academic platforms make advanced modeling accessible to a wide research community without licensing burdens. Commercial software, while requiring a financial investment, provides unmatched ease of use, integration, and customer support. Depending on the system’s complexity, computational demands, and research goals, researchers working with ZnO-based photocatalysts can select the most suitable combination of tools to ensure efficient and reliable modeling workflows.

4.4. Examples of Studies on ZnO-Based Photocatalytic Materials Complemented by Atomistic Calculations

When addressing the issue of band gap underestimation by conventional DFT methods, such as the widely used and computationally efficient PBE functional, it is essential to highlight the study by Goh et al. (2017) [198]. In this work, the authors performed DFT+U calculations using the VASP modeling package [180,181,182,183,184] on the wurtzite ZnO structure, exploring a broad range of Hubbard U values for both Zn and O atoms in order to identify the optimal parameters that best reproduce experimental observations. In addition to the band gap, one of the most critical properties, they also evaluated structural parameters and benchmarked their results against both experimental data and prior theoretical studies. The study concluded that two specific sets of U values yielded the best agreement with experimental results, U_d = 9.5 eV and U_p = 7.86 eV and U_d = 10.0 eV and U_p = 7.74 eV, where U_d refers to Hubbard potentials for d-states, while U_p refers to Hubbard potentials for p-states.

In the context of benchmarking various methods for obtaining a reliable electronic structure of ZnO-based materials, it is also important to highlight the review by Harun et al. (2020) [22], which summarizes key studies employing a range of computational approaches and modeling codes. The authors emphasized that the DFT+U method provides a computationally inexpensive yet effective strategy for enhancing the accuracy of band gap predictions, making it a valuable tool in modeling ZnO and related photocatalytic materials.

In a recent study by Wang et al. (2025) [233], a DFT study combining the PBE and HSE06 density functionals, along with Grimme’s D3 dispersion correction to properly account for van der Waals interactions, was conducted to explore two novel 2D ZnO monolayers, SOD-110 and SOD-111 [186]. These monolayers were derived from specific crystal planes of a porous ZnO phase. Structural optimizations revealed that both systems are thermodynamically and kinetically stable. SOD-110 exhibits a centrosymmetric network of Zn-O four- and six-membered rings, while SOD-111 adopts a bowl-shaped geometry resembling a Zn₁₂O₁₂ cluster. Furthermore, the study shows that substitutional doping with elements like Te, S, Se, and Cd enables precise tuning of the electronic structure and photocatalytic performance, offering valuable control over key structural and functional properties.

Lanthanum-doped ZnO (LZO) thin films were investigated from both experimental and computational perspectives in the study by Soussi et al. (2022) [187,234]. Atomistic calculations were performed using the Full Potential Linearized Augmented Plane Wave (FP-LAPW) method. Structural characterization via XRD revealed that the synthesized films exhibited a tetragonal wurtzite structure. Theoretical modeling employed the Tran–Blaha modified Becke–Johnson (TB-mBJ) approximation to analyze electronic and optical properties. The authors reported band structures for ZnO with varying La concentrations, showing that the band gap increases with higher La content—consistent with experimental observations. Furthermore, while the transmittance gradually increased with La concentration, the refractive index exhibited only a slight increase.

In a recent study, Kumar et al. (2025) [235] investigated the co-doping effects of Al and Tl on the optoelectronic properties of thin ZnO films using both experimental methods and DFT calculations based on the PBE functional, with D3 dispersion corrections included [188]. Among other findings, the authors reported key structural parameters derived from DFT, such as lattice constants, bond lengths, and formation energies. The formation energies ranged from −2.00 eV for Al-doped ZnO to 5.83 eV in the case of radiatively annealed ZnO films.

In their 2025 study, Salman et al. [188] investigated the influence of Mg doping on the structural and electronic properties of ZnO. Due to the close similarity in ionic radii between Mg²⁺ and Zn²⁺, Mg serves as an ideal dopant, enabling substitution within the ZnO lattice with minimal distortion. The authors specifically examined the Mg_0.5Zn_0.5O composition, in which two Mg atoms replaced two Zn atoms in the hexagonal wurtzite unit cell. Structural and electronic properties were evaluated using the hybrid HSE06 density functional, yielding results in close agreement with experimental data. The introduction of Mg caused only slight lattice distortion; the a parameter decreased from 3.2815 Å to 3.2830 Å, and the c parameter decreased from 5.2862 Å to 5.0950 Å, relative to pure ZnO values reported in reference [185]. HSE06 calculations also revealed a direct band gap of 2.435 eV for the doped structure.

Another noteworthy study focusing on the photocatalytic performance of ZnO-based materials and heavily relying on DFT calculations was conducted by Ullah et al. (2025) [189,236]. The authors investigated the effects of Mn, Co, and Al doping, as well as citrate capping, on the reactivity of ZnO nanoparticles and their efficiency in photocatalytic dye degradation. Their results demonstrated that the combination of Co doping and citrate capping notably enhanced photocatalytic performance. DFT calculations and a detailed analysis of the band structure played a crucial role in explaining this improvement by revealing the emergence of an additional absorption peak in the visible range due to Co doping, along with a broadening of the absorption spectrum attributed to citrate capping. These insights provided a mechanistic understanding of the observed enhancement in photocatalytic activity. The limited availability of systematic investigations of the individual influencing parameters brought attention to studies supplemented by atomistic calculations based on DFT. For ZnO, the computational methods have mainly been used to rationalize experimental data published on bulk and pure ZnO nanoparticles. However, the literature lacks calculations dedicated to modified materials with their photocatalytic applications [237,238].

5. Modeling of ZnO-Based Photocatalysts for Photocatalysis Based on Machine Learning

Parallel to advances in atomistic modeling, ML has emerged as a powerful data-driven approach to accelerate the discovery and optimization of ZnO-based photocatalysts. By analyzing existing experimental or simulated datasets, ML models can identify complex patterns and relationships between structural features and photocatalytic performance [239]. These models are capable of predicting key properties, such as band gaps, stability, or degradation efficiency, across a wide range of compositions and synthesis conditions. ML not only complements traditional simulations but also reduces the need for exhaustive trial and error experimentation, offering a scalable path toward intelligent material design.

5.1. Machine Learning Techniques in Materials Science

The capacity of ML algorithms to provide insight into large and complex datasets is a key advantage in materials science. ML algorithms, combined with the availability of large experimental and theoretical datasets and advances in high-performance computing, promise to revolutionize materials science, providing greater insight, better design, and guidance to complement and improve experimental and theoretical work [240,241]. ML methods are a set of computational techniques that enable systems to learn patterns or relationships from data without being explicitly programmed (Figure 8). These methods use algorithms to analyze input data, identify underlying structures, and make predictions or decisions based on new data [242].

ML is typically categorized into three main types:

• Supervised learning, where models are trained on labeled data;
• Unsupervised learning, which explores hidden patterns in unlabeled data;
• Reinforcement learning, which learns through trial and error by interacting with an environment.

The ML methods aim to generate an ML model, which is a mathematical representation trained on data that can make predictions or classifications. It is the outcome of applying an ML algorithm to a dataset. Once trained, an ML model can take unseen input data and output results based on the patterns it has learned [243].

ML has proven to be a unique computational approach able to optimize the search for novel materials with target properties [244]. This approach enables researchers to extract patterns, generate predictions, and direct the discovery of novel materials with improved efficiency [245,246]. Initially, materials development relied on time-consuming trial and error experimentation. This approach was subsequently complemented with various atomistic calculations, relying on molecular and quantum mechanics, including methods ranging from force field calculations to density functional theory and computationally demanding modern wavefunction methods [247].

In materials science, ML models are often used to predict properties like the band gap, stability, catalytic activity, or formation energy of materials by learning from previously calculated or measured datasets [248]. In these regards, ML models are typically trained on large datasets that include structural, electronic, thermodynamic, and mechanical properties [249]. These models can predict target properties, classify materials, or even suggest new candidates with desired functionalities. This shift toward data-centric research is particularly valuable in photocatalysis, where the performance of a material depends on a complex interplay of multiple factors, such as band gap energy, surface area, crystallinity, defect density, and charge carrier mobility [247,249,250,251].

The integration of ML into materials research also fosters interdisciplinary collaboration, merging the expertise of chemists, physicists, computer scientists, and engineers. With the increasing availability of open source tools and materials databases, such as the Materials Project [252], AFLOW [253,254,255], or the Open Catalyst Project [256], the barrier to entry has been significantly lowered, allowing a broader scientific community to harness the power of ML for materials discovery and optimization. In developing ZnO-based photocatalysts, the use of ML significantly contributes to overall cost and time effectiveness, particularly given the various methods to improve commercial ZnO discussed in Section 3. This section explores how ML is being applied to ZnO-related materials for photocatalytic applications, with a focus on predictive modeling, materials discovery, and performance optimization.

5.2. Machine Learning Applications in ZnO-Based Photocatalysts: From Property Prediction to Reaction Modeling

Before presenting concrete examples of ML applications in ZnO-based photocatalysis, it is helpful to consider the broader context in which ML methods can contribute to this research area. ZnO is a widely studied semiconductor material with diverse applications in environmental remediation, energy conversion, and antibacterial coatings, particularly due to its high photoreactivity, tunable electronic properties, and chemical stability. However, optimizing its performance for specific photocatalytic reactions involves navigating a complex design space, including variations in morphology, dopants, synthesis conditions, and reaction environments. To address these challenges, ML provides a robust framework by enabling data-driven prediction, correlation discovery, and guided materials optimization. In this section, we explore how ML can be employed to predict the intrinsic properties of ZnO photocatalysts, model photocatalytic degradation mechanisms, optimize composite formulations, and evaluate the sustainability of ZnO-based photocatalysts. In addition, sustainability assessment, such as life cycle impact estimation and resource efficiency, benefit from ML-driven approaches that integrate environmental criteria and performance metrics.

One of the primary applications of ML in ZnO research is the prediction of intrinsic material properties. The versatile properties of ZnO arise from its tunable electronic and optical features, which can be modified through methods like doping, morphological engineering, or composite formation. ML models trained on experimental datasets or results from DFT calculations can accurately predict properties like band gap energy, electron affinity, dielectric constants, mechanical strength, and optical absorption coefficients [257,258]. The potential antibacterial mechanism of ZnO nanoparticles with microbial cells was proposed by Chelghoum et al. [257]. In order to modify the photodegradation of quinoline yellow dye by ZnO nanoparticles, Gaussian Process Regression (GPR) combined with an improved Lévy flight distribution (FDB-LFD) algorithm was applied. High accuracy of the used methodology was provided by automatic relevance determination and an exponential kernel function. Precise optimization of photodegradation rate predictions was achieved using a MATLAB-based application, which was developed to integrate the GPR_FDB-LFD model and the FDB-LFD algorithm. The results of this study offered cost-effective and efficient methods, obtained through GPR and FDB-LFD approaches, in order to save resources and time during the photodegradation of dyes. Akyildiz et al. [258] applied ML algorithms, artificial neural networks (ANN), and Support Vector Regression (SVR) to predict the photocatalytic activity of atomic layer deposition (ALD)-coated textiles. ML algorithms were trained and tested using the k-fold cross-validation technique. The results showed that ANN and SVR methods can predict ALD-coated textiles’ photocatalytic activity in the degradation of methylene blue and methyl orange. The results indicated that the ALD-coated textiles’ photocatalytic activity can be predicted with the selected range of process parameters using suitable ML algorithms. Overall, these predictive models typically utilize descriptors derived from atomic composition, crystal structure, or synthesis parameters. They can be based on algorithms, such as Random Forests, Support Vector Machines, or neural networks. This approach enables researchers to screen large numbers of candidate structures before undertaking expensive experimental validation or high-level computations, thereby significantly reducing the resource investment required in early-stage materials design.

Another important application of ML methods, closely related to the predictive modeling discussed in the previous paragraph, is the optimization of material properties through detailed feature importance analysis. Once an ML model with satisfactory accuracy and stability is developed, it becomes essential to analyze how individual input features influence the model’s predictions. This analysis not only helps in understanding the underlying structure–property relationships but also highlights which features have the most significant impact on photocatalytic performance. One of the most widely used techniques for such an analysis is SHapley Additive exPlanations (SHAP) [259,260], which provides consistent and interpretable insights by quantifying the contribution of each feature to a specific prediction. For example, in an ML model trained to predict the degradation efficiency of dyes using ZnO-based photocatalysts, SHAP analysis might reveal that factors like particle size, surface area, and band gap energy are the most influential descriptors [261]. Such insights can guide researchers in optimizing synthesis protocols or selecting material modifications to enhance catalytic activity further.

To better illustrate these concepts, Figure 9 and Figure 10 present two visualizations generated using synthetic data that we created solely for explanatory purposes. In Figure 9, we display a traditional feature importance plot based on a Random Forest model.

Feature importance in ensemble tree models is calculated by measuring the average decrease in variance (for regression tasks) or impurity (for classification) whenever a feature is used to split a node across all trees. The higher the cumulative reduction in error associated with a feature, the greater its importance. In this example, feature 3, feature 5, and feature 1 clearly stand out as the most influential, while features 4, 2, and 6 show minimal impact. This ranking provides a global view of which features contribute most to the model’s predictive power.

However, this method does not provide insight into the direction or distribution of feature effects. To address this, Figure 10 shows the corresponding SHAP summary plot, offering a finer-grained understanding of model behavior.

SHAP values decompose each individual prediction into the sum of contributions from all features based on cooperative game theory. In this figure, each dot represents a SHAP value for one feature and one sample, indicating whether that feature increased or decreased the prediction and by how much. Features are annotated with the number of times they decreased (L) or increased (R) the predicted output across the dataset.

As seen in the SHAP plot, feature 3 consistently produces large shifts in model output, confirming its dominant influence. The blue-to-red gradient shows that higher values of feature 3 (in red) tend to increase the predicted outcome, while lower values (in blue) tend to decrease it. This indicates a positive correlation between feature 3 and the model’s prediction. In contrast, feature 5 shows the opposite pattern; higher values (red) are mostly associated with negative SHAP values, indicating that increasing this feature lowers the predicted outcome. This implies a negative correlation between feature 5 and the target property in the learned model.

Feature 1 also plays a major role and exhibits an apparent positive effect, although to a lower extent compared to features 3 and 5. Namely, higher values of this feature typically raise the predicted output, as shown by the clustering of red points on the positive SHAP side. Meanwhile, features 6, 2, and 4 show low SHAP magnitudes and more symmetric distributions centered near zero. Their feature values do not consistently influence the prediction in any direction, confirming their minimal predictive relevance, which is also reflected in their low importance scores in Figure 9.

Together, these complementary techniques illustrate not only which variables matter most to the model but also how they influence predictions. For example, feature 1 can be the band gap of materials, feature 2 can be nanoparticle size, etc. Such interpretability tools are crucial for guiding rational design and optimization strategies in ML-driven materials research, including ZnO-based photocatalysts.

In addition to predicting material properties, ML is increasingly used to model the photocatalytic degradation of organic molecules. ZnO-based photocatalysts are widely applied in the decomposition of dyes, pharmaceuticals, pesticides, and other persistent pollutants under UV or visible light [258]. ML techniques can capture the complex interplay between catalyst properties, reaction conditions, and pollutant structures. For example, with ML methods, one can predict degradation efficiency based on variables like surface area, pH, light wavelength, and initial pollutant concentration [258]. Moreover, classification models and clustering algorithms have been employed to identify patterns in degradation pathways and predict the formation of intermediates and final products [15]. These insights are especially valuable in environmental applications, where understanding by-product toxicity and degradation mechanisms is as critical as achieving high removal efficiency.

The synergy between ML and first-principles computational methods is another area of exciting development. By training models on data generated through DFT calculations, researchers can extend the predictive power of ab initio methods to larger systems or unknown chemical spaces with reduced computational costs [262,263]. A notable advancement in this domain is the OMol25 project [264], which involved over 100 million DFT calculations to create a comprehensive dataset encompassing a diverse range of molecular systems. This dataset enabled the training of the UMA [265], an ML model capable of predicting quantum mechanical properties with DFT-level accuracy, significantly accelerating computational chemistry research. An essential component of the UMA framework is the OC20 model, initially developed for catalysis science. The OC20 model enables the modeling of complex surfaces and adsorbate interactions, such as ZnO surfaces, thus opening the door for advanced ML-driven photocatalytic investigations. The entire project, including the code and the pre-trained models, is available open source through the FAIRChem initiative (https://fair-chem.github.io/, accessed on 1 August 2025).

5.3. Examples of Data-Driven Predictions in Photocatalysis Using ZnO-Based Materials

In a recent study by Dashti et al. (2024) [15], ML was applied to model and improve the prediction of pesticide photodegradation in water using ZnO-based photocatalysts. The authors applied and optimized four different ML models—the Multi-Layer Perceptron Artificial Neural Network (MLP-ANN), the Particle Swarm Optimization–Adaptive Neuro-Fuzzy Inference System, the Radial Basis Function, and the Coupled Simulated Annealing–Least Squares Support Vector Machine (CSA-LSSVM)—and trained them on a dataset extracted from literature sources. Input features included both process parameters (e.g., irradiation time, catalyst dosage, pH, initial pesticide concentration, light source, and dopant ratio) and structural descriptors (e.g., molecular weight and solubility of pesticides and dopants).

Among the models, the RBF network achieved the highest prediction accuracy, with a coefficient of determination (R²) of 0.978 and an average absolute relative deviation (AARD) of just 4.80%. For specific pesticides, such as 2-chlorophenol, triclopyr, and lambda-cyhalothrin, the CSA-LSSVM model performed better. Sensitivity analysis revealed that irradiation time had the most substantial positive influence on degradation efficiency, while initial pesticide concentration was the most negatively correlated parameter. The study demonstrates how ML tools can uncover nonlinear relationships and optimize photocatalytic processes more efficiently than traditional trial and error approaches, thereby contributing to the design of more effective ZnO-based nanocomposites for environmental remediation. The authors also emphasized the use of an online application for the extraction of data from graphs, WebPlotDigitizer [266].

In a recent study by Esmaeili et al. (2023), ML was integrated with experimental research to enhance the understanding and optimization of ZnO-based photocatalysts modified with CdS nanocrystallites for the visible-light-driven degradation of tetracycline [267]. Two robust ML models—based on artificial neural network (ANN) and gradient boosted regression tree (GBRT) methods—were developed to predict the degradation efficiency (% removal) based on key experimental parameters. These included initial tetracycline concentration, catalyst dosage, radiation time, ultrasonic wave presence, lamp power, Cd/Zn ratio, and ammonium persulfate concentration. The ANN and GBRT models achieved high prediction accuracies (R² = 0.99 and 0.98, respectively), closely matching the experimental results. Feature importance analysis revealed that the amount of catalyst and the Cd/Zn ratio were the most influential factors affecting photocatalytic efficiency, while initial concentration and ultrasonic treatment had lesser impacts. The successful application of ML in this context demonstrates its utility in modeling complex photocatalytic systems, guiding parameter optimization, and reducing experimental workload.

In the study by Lamouadene et al. (2025), ML was applied to explore and predict the energy band gap of doped ZnO systems [268]. The dataset consisted of experimental measurements from 107 samples of undoped and doped ZnO thin films. The input features included Urbach energy (indicating defect density), reactant concentration in mol/L, substrate temperature during deposition, light transmittance, and the type of deposition method used (ultrasonic spray, ultrasonic spray pyrolysis, pulsed laser deposition, or spray pyrolysis). The target variable was the experimentally measured energy band gap. This empirical dataset provided a basis for training ML models to predict the band gap based on material and process parameters. In particular, the authors applied Gaussian Process Regression (GPR), Support Vector Machine, and Random Forest models. Among the ML models evaluated for predicting the energy band gap, GPR demonstrated the best performance, achieving a correlation coefficient of 98.97%, RMSE of 0.0022, and MAE of 0.0020, indicating highly accurate predictions with minimal error. In comparison, SVM and RF yielded lower accuracies, with correlation coefficients of 83.70% and 76.40%, respectively, along with higher error metrics (SVM: RMSE = 0.0052, MAE = 0.0048; RF: RMSE = 0.0086, MAE = 0.0083). These results highlight the effectiveness of GPR for modeling nonlinear relationships in experimentally derived datasets for ZnO-based materials.

In the study by Chelghoum et al. (2024), ML was employed to predict the antibacterial activity of green-synthesized ZnO nanoparticles [257]. The authors developed predictive models using SVR and RF models, utilizing input features like particle size and shape, surface area, band gap, and synthesis conditions. The target variable was the antibacterial activity, quantified by the zone of inhibition against various bacterial strains. Among the models, SVR demonstrated superior predictive performance with an R² of 0.962, while the RF model achieved an R² of 0.911. These results highlight the potential of ML to accurately link nanoparticle characteristics to their biological activity, leading to a more efficient design of ZnO nanomaterials for biomedical applications.

In another recent study by Navarro-López et al. (2024) [269], ML was employed to predict bacterial survival in response to undoped and lanthanum-doped ZnO nanoparticles (NPs), offering a data-driven approach to optimize antimicrobial nanomaterials. Eight regression algorithms were evaluated, including Random Forest (RF), Extremely Randomized Trees (ERT), Gradient Boosting, K-Nearest Neighbors, and Support Vector Regression. Among them, the ERT model exhibited the highest predictive performance after hyperparameter tuning, achieving an R² of 0.95 and a mean absolute error (MAE) of 3.48%, demonstrating strong accuracy in modeling complex antibacterial outcomes. Feature importance analysis indicated that the two most important features were PA and EC. Although the authors did not introduce these abbreviations, it is most likely that they meant to indicate the presence of Pseudomonas aeruginosa and EC50 values. Interestingly, lanthanum content and nanoparticle size, though often considered key design variables, were found to have minimal predictive value.

6. Environmental Applications and Future Perspectives

In recent years, heterogeneous photocatalysis has been applied to several industrial processes (wastewater treatment, air purification, and fine chemical production). Among the various types of photocatalysts, ZnO-based materials have shown promising potential due to their excellent redox capabilities and environmental compatibility. However, the main disadvantage of ZnO-based photocatalysts is the difficulty of their separation from the medium and their regeneration. In order to improve this, the ZnO photocatalyst needs to be modified. Because conventional treatments have not shown high activity in the degradation and mineralization of hazardous organic substances, new alternatives are needed. This problem involves the introduction of new approaches and innovative methodologies to obtain hybrid composites that are able to meet the requirements of the industrial sector and environmental protection requirements. These composites should be simple to prepare, exhibit strong photocatalytic efficiency, be cost-effective, and be recycled and reused [270,271].

Pharmaceutical compounds are detected in environmental waters and municipal or hospital wastewaters due to their continual use in human and veterinary medicine. Direct emissions from domestic and industrial use, excretion via urine and feces, and disposal of pharmaceuticals in solid waste represent pathways through which pharmaceuticals enter the environment. The presence of pharmaceuticals in the aquatic environment poses a risk to both aquatic organisms and humans, as they are part of the food web. The impact pharmaceutical pollutants have on the ecosystem requires the development of alternative methods for their removal. The most promising alternative methods include AOPs [272]. Among these methods, photocatalysis with ZnO has attracted the attention of researchers in pharmaceutical wastewater treatment due to its high redox potential, chemical stability, and low cost. Under solar or visible light radiation, photogenerated e⁻–h⁺ can oxidize pharmaceutical pollutants and degrade them to CO₂, H₂O, and some inorganic ions [273].

The dye industry is characterized by the excessive discharge of toxic, untreated, and colored wastewater. Therefore, cost-effective photocatalysis processes are developed to degrade dye pollutants in the dye industry’s wastewater. The use of ZnO-based photocatalysts in the photocatalytic degradation of dyes showed higher efficiency compared to commercial ZnO photocatalysts [274]. Dye molecules could be easily degraded due to the high reactivity of the various active species, resulting in their high mineralization. ZnO has been successfully used to degrade different types of dyes, such as the light-sensitive bile pigment bilirubin, methyl orange, methylene blue, etc. Despite the outstanding properties of ZnO-based photocatalysts, it is also important to mention the disadvantages, such as reduced activity under visible light and insufficient recyclability. They remain critical challenges. Synthesis of photocatalysts that are active under visible light and possess structural robustness is critical to improve dye degradation efficiency [273,275].

Microplastics have attracted widespread attention because of their non-biodegradability and difficulty in handling, which represent a threat to the environment, humans, and animals. Advances in microplastic technologies include the application of photocatalytic processes and ZnO nanoparticles. Due to environmental pollution caused by microplastics with a size of ≤1 mm, ZnO nanorod arrays on nickel foam were synthesized. They showed excellent photocatalytic activity in the degradation of methylene blue and microplastics under UV irradiation [276]. Also, ZnO thin films showed efficient photocatalytic activity under visible light in the degradation of carbamazepine and microplastics. They have also been used in the degradation of carbamazepine and microplastics in the presence of solar light, the most green and renewable resource [277,278]. These findings imply that ZnO photocatalysts play an important role in microplastic remediation technologies. However, additional research is required to evaluate long-term degradation pathways, secondary pollution risks, and field-scale application.

Table 2. Degradation of different pollutants using ZnO-based materials.

No.	Material	Pollutant	Radiation Source	Degradation Efficiency	Reference
1	ZnO	Rhodamine B	Visible light	20%, 60 min	[279]
2	ZnO/carbon nanotubes	Rhodamine B	Sunlight	50%, 60 min	[280]
3	TiO2/ZnO nanofibers	Rhodamine B	Visible light	100%, 60 min	[281]
4	ZnO	Methylene blue	Visible light	86%, 180 min	[282]
5	ZnO/graphene composites	Methylene blue	Visible light	100%, 90 min	[283]
6	Ag/ZnO	Methylene blue	Sunlight	98%, 30 min	[284]
7	ZnO	Paracetamol	Visible light	54%, 240 min	[285]
8	Ag/ZnO	Paracetamol	Sunlight	80%, 240 min	[286]
9	ZnO/g-C3N4	Paracetamol	Visible light	90%, 60 min	[287]
10	ZnO	Amoxicillin	Visible light	38%, 90 min	[288]
11	Bi2WO6/nano-ZnO	Amoxicillin	Visible light	93%, 120 min	[289]
12	ZnO/TiO2	Amoxicillin	Visible light	94%, 210 min	[290]
13	ZnO	Polypropylene	Visible light	40%, 3600 min	[291]
14	ZnO nanorods	Polypropylene	Sunlight	100%, 11,760 min	[292]
15	ZnO	Polyethylene	Visible light	15%, 10,080 min	[293]
16	Fe-ZnO	Polyethylene	Sunlight	40%, 7200 min	[294]

summarizes the use of pure ZnO and modified ZnO photocatalysts in the degradation of pharmaceuticals, dyes, and microplastics. From the table, it can be seen that microplastic nanoparticles need much more time to degrade.

The chemical stability of ZnO-based photocatalysts depends on the form and application method of ZnO. Modification of ZnO nanoparticles with organic molecules or metal oxides improves their hydrolytic stability. The chemical stability of ZnO can be improved with three simple methods: protecting the surface by coating and encapsulating with polymers and chemicals for water repellent treatment; modifying ZnO nanoparticles by covering with surfactants or hybridization with other metal oxides [42]; and the formation of a ZnO–MOF shell structure. Many studies have shown that ZnO/SiO₂ and ZnS-Si/ZnO anchored structures possess strong chemical stability [295,296]. The thermal stability of the ZnO-based materials was studied using TGA (thermogravimetric analysis). GO-ZnO/SiO_2-3 was studied in the range of 25–800 °C in nitrogen. The GO-ZnO/SiO_2-3 showed an initial weight loss of 9.3% between 50 and 220 °C, which is ascribed to the degradation of organic species present in the material shell. At temperatures higher than 500 °C, GO degrades while ZnO remains stable up to 800 °C. These results indicate that ZnO photocatalysts are highly durable under thermal stress and that they can be reused across multiple catalytic cycles and under diverse environmental conditions [297,298].

The mechanical performance of photocatalysts is important in water remediation, pollution control, and solar energy conversion. Undesirable changes, such as cracking or chipping, could occur if the structural stability of photocatalysts is weak. This will reduce the performance of the photocatalyst. Even minor deterioration could increase the recycling costs of the photocatalyst. Because ZnO-based photocatalysts showed high photocatalytic activity, their practical application can be realized if their structural stability is improved [299]. Recent studies on ZnO-based photocatalysts can be classified as follows: (i) blending with other ceramic materials to improve strength, (ii) surface modifications to protect against evaporation of scalars, such as water molecules and organic compounds, on the catalyst’s surface, (iii) one-dimensional nanostructures to minimize surface tension, and (iv) the creation of hierarchical designs as core–satellite structures [300]. However, the relationship between long-term photocatalytic performance and mechanical stability remains insufficiently studied and should be explored in the future.

Catalyst design is the inverse problem of reaction prediction. Catalyst design methods include empirical approaches that transfer previous experience to the system in question and theoretical models that seek physically based solutions. Still, catalyst design remains challenging [301]. The inability of first-principles models to capture all relevant phenomena makes it challenging to gain timely and cost-effective access to new materials. In recent years, we have witnessed increasing interest in data-driven AI methods as a means of solving problems. The computational and labor costs of preparing accurate quantum–chemical or experimental data prevent the possibility of constraining the chemical space to generate catalysts with optimal efficiency [302]. Although these restraints limit the amount of data available for AI approaches, accumulated knowledge regarding these methods and their advantages is leading to numerous AI applications that improve the design of catalytic processes. These range from deriving general structure–performance and structure–property relationships to models that facilitate microkinetic simulations for mechanistic studies [303]. Predictive modeling, optimization algorithms, and simulation represent the main AI tools that play complementary roles. Quantitative structure–property relationships, catalyst performance based on molecular descriptors, represent a predictive model that enables a rapid assessment of potential photocatalyst candidates. Targeted design can be achieved through the optimization of algorithms that refine molecular properties. Simulation models can elucidate the mechanisms linking structure to behavior. Nowadays, AI tools are integrated to improve predictive accuracy and efficiency, with established methodologies to advance catalyst design [304].

The availability of accessible datasets limits the development of AI methodologies for photocatalyst design. Both the diversity and quality of the data are essential for the reliability of AI models [305]. Many have collected data from disparate sources, such as public repositories and journal articles, which lack sufficient quality, consistency, and trustworthiness to satisfy AI and ML standards. Consequently, these data are not suitable for developing generalizable models [306]. A typical approach is to generate consistent datasets by performing systematic analyses with uniform protocols or by conducting optimizations of specific features in high-throughput calculations. However, the models derived from these datasets often fail to generalize across the broader materials space or diverse phenomena required in emerging applications [307]. The data quality strongly affects the reliability of predictions from AI models, particularly in the design of materials. High-quality experimental data, including accurate structural information and process parameters, are essential for training dependable ML models [308]. Failing to meet these standards can result in inaccurate predictions that lack robustness. Beyond quality, the diversity of data also plays a critical role in model generalization. In some cases, when datasets exhibit a narrow distribution of properties, ML models could overfit and have problems generalizing effectively beyond the encountered range. In the era of AI, the design of AI-based materials still has to deal with many challenges linked to data quality and diversity [309,310].

The production capacity of ZnO is one million tons per year. Methods used for the synthesis of ZnO are thermal oxidation of Zn metal, direct precipitation from aqueous solutions containing Zn, vapor phase condensation, spray pyrolysis methods, hydrochemical methods, electrodeposition, and solid separation from natural metal concentrates, wastes, and metallurgical products [311]. The production of ZnO in different forms (powders, nanopowders, crystals, monocrystals, thin films, clusters, paracrystalline structures, homogenized mixtures with other zinc or non-zinc compounds, coatings, and colloidal suspensions) can be implemented and controlled by using ZnO precursors and suitable methods in combination with appropriate physical, chemical, structural, morphological, phase, and crystalline properties of the produced ZnO. The production routine of ZnO must be scaled up in order to produce sufficient quantities of ZnO for its applications and needs [312,313]. It is also necessary to develop scalable and optimized methodologies for the separation of nanomaterials formed during synthesis from the reaction medium. By using inexpensive precursors and green routes with short reaction times associated with scalable separation processes, high purity and high yields of ZnO could be obtained. Another major challenge for obtaining large volumes of ZnO-based photocatalysts for applications in kg or larger volumes is the ability to recover and reuse the initial ZnO photocatalyst dispersed in large solution volumes. It is challenging to recover hundreds of grams or kilograms of dispersed ZnO-based photocatalysts, purify them, and redisperse them in a new experimental trial and perform this process several consecutive times [314]. Therefore, the implementation of green synthesis routes and efficient separation technologies is essential for the transition from lab-scale to industrial-scale ZnO applications.

7. Conclusions

In this review, we comprehensively examined the latest advances in the development and optimization of ZnO-based composite photocatalysts for environmental applications. A particular emphasis was placed on strategies to overcome the intrinsic limitations of pure ZnO, such as limited visible-light absorption and rapid recombination of charge carriers. Material modifications, including metal and non-metal doping, heterojunction formation, and the incorporation of polymers, metal–organic frameworks, and carbon-based materials, were systematically discussed, with a focus on their synergistic effects and structure–property relationships that influence photocatalytic activity.

Furthermore, we addressed the growing role of computational modeling through atomistic calculations and ML in accelerating the design and optimization of ZnO photocatalysts. By integrating experimental data with ML algorithms, researchers have successfully predicted critical parameters, such as band gap energy and photocatalytic degradation efficiency, thereby reducing the need for time-consuming experimental trials. This review highlights the potential of integrating material innovation with data-driven approaches to inform the rational design of high-performance photocatalysts. Future research should focus on (i) expanding high-quality databases that include descriptors relevant to environmental performance under real-world conditions; (ii) improving the transferability of ML models by including complex pollutant matrices, reaction intermediates, and degradation by-products; (iii) developing interpretable ML frameworks that provide physical insights; and (iv) incorporating automated high-throughput experimentation to bridge the gap between theory and application.

In addition, hybrid workflows that combine density functional theory (DFT), molecular dynamics (MD), and ML can offer a multi-scale perspective to guide catalyst design with both electronic and morphological control. Promoting interdisciplinary collaborations between materials scientists, environmental engineers, and data scientists will be essential to realize the full potential of data-driven photocatalysis.