Transformative strategies in photocatalyst design: merging computational methods and deep learning

General principle of photocatalysis

Photocatalysis has emerged as a crucial technology with applications spanning environmental remediation, energy conversion, and sustainable chemical production^[28]. Its significance lies in its ability to harness solar energy to drive chemical reactions, contributing to pollution reduction, carbon neutrality, and green energy generation. One of the primary advantages of photocatalysis is its ability to operate in-situ without the need for external energy inputs, relying solely on solar energy^[29]. For example, in environmental remediation, photocatalysis has been extensively demonstrated under natural sunlight conditions^[30]. Furthermore, it supports simultaneous pollution reduction and carbon mitigation^[31]. This is evidenced by the photocatalytic reduction of CO₂ into useful fuels or chemicals, promoting both environmental sustainability and carbon neutrality^[32].

The fundamental mechanism^[33-35] of photocatalysis involves the excitation of electrons within a semiconductor upon absorption of photons with energy equal to or greater than its band gap. This excitation promotes electrons from the valence to the conduction band, leaving corresponding holes in the valence band. The photogenerated charge carriers (electrons and holes) then migrate to the surface, where they participate in redox reactions. Electrons predominantly participate in reduction reactions, where they are transferred to oxygen molecules (O₂) adsorbed on the photocatalyst surface, leading to the formation of superoxide radicals (•O₂^-). These superoxide radicals can subsequently react with protons to produce hydrogen peroxide (H₂O₂) or other reactive oxygen species (ROS), which are highly effective in degrading organic pollutants and reducing heavy metals. Conversely, the holes (h⁺) facilitate oxidation reactions by oxidizing water (H₂O) molecules adsorbed on the photocatalyst surface, generating hydroxyl radicals (•OH), which are highly reactive species capable of decomposing complex organic pollutants, splitting water or producing hydrogen peroxide, thus contributing to sustainable production. The overall efficiency of photocatalysis is primarily governed by critical factors such as charge carrier separation, charge mobility, and the efficient suppression of electron-hole recombination^[36]. Figure 1 presents an in-depth depiction of the pathways undertaken by photogenerated charge carriers, emphasizing the intricate redox reactions occurring at the photocatalyst surface, alongside recombination processes occurring both within the bulk and at the surface.

Figure 1. Fundamental mechanism of photocatalysis^[37]. Copyright 2022, Elsevier.

The core of photocatalysis lies in the photocatalysts. Traditional photocatalysts are typically bulk semiconductors^[38] that harness light to drive photocatalytic reactions, but they often suffer from low efficiency due to limited surface area and rapid recombination of charge carriers. Nanostructured photocatalysts^[39], on the other hand, offer significantly enhanced performance due to their increased surface area and improved charge carrier dynamics, which help reduce recombination and enhance photocatalytic activity. Heterojunction systems^[40,41] combine two or more semiconductors with different band structures, creating interfaces that facilitate more efficient charge separation and transfer, thus improving overall efficiency. Low-dimensional systems, such as quantum dots^[42], one-dimensional^[43], and two-dimensional (2D) structures^[44,45], exhibit unique quantum confinement effects that significantly alter their electronic properties. Quantum dots have tunable band gaps due to their nanoscale dimensions, enabling enhanced light absorption and charge separation^[46]. One-dimensional structures, including nanowires or nanotubes, provide high charge carrier mobility along their axial direction to improve performances^[47]. 2D systems feature high surface-to-volume ratios and exceptional electronic properties, offering more active sites for photocatalytic reactions^[48].

The performance of photocatalysts is critically determined by several key characteristics, each influencing different aspects of photocatalytic efficiency. (1) The microstructure, including grain size, specific surface area, and pore size, directly affects the availability of active sites and the overall surface reactivity; (2) Architecture factors, such as quantum dots, nanowires, and heterojunctions, influence light absorption and charge separation, with more advanced structural designs often improving charge transportation; (3) Composition, including stoichiometry, doping elements, and the presence of defects, plays a pivotal role in tuning electronic properties, thereby optimizing light absorption, charge carrier dynamics, and redox processes; (4) The electronic structure, particularly the band gap, density of states (DOS), and electron affinity, dictates the ability to absorb irradiation energy and generate charge carriers, which are essential for driving photocatalytic reactions; (5) Photoelectric properties, such as carrier mobility, carrier lifetime, and quantum yield, determine the effectiveness of converting absorbed light into chemical energy. Higher mobility and longer carrier lifetimes facilitate more efficient charge separation and reduce recombination. Table 1 lists the key characteristics of photocatalysts and their typical characterization techniques.

Common modification strategies for photocatalysts include doping^[49], cocatalyst loading^[50] and incorporating single-atom catalysts^[51,52]. Latest modification mechanisms, such as defect engineering, band structure engineering, and interfacial engineering, enable precise control at the atomic, electronic, and chemical bond levels, significantly enhancing photocatalytic performance^[53,54]. Recently, constructing heterojunctions^[55,56] remains a major focus of research, with types including Type I, Type II, Z-scheme, and S-scheme heterojunctions being extensively studied. In particular, S-scheme heterojunctions offer distinct advantages^[57,58], such as optimized band alignment and enhanced charge separation, which contribute to superior redox capabilities and reduced electron-hole recombination, making them highly promising for practical applications^[59]. For instance, in ribbon-based water purification, these modification strategies have been instrumental in enhancing photocatalytic degradation efficiency by improving charge carrier dynamics, optimizing active sites, and facilitating the breakdown of complex organic pollutants, thereby demonstrating significantly superior performance^[60].

Traditional procedure of photocatalyst design

Traditional photocatalyst design has predominantly relied on empirical methods, where photocatalysts are synthesized and evaluated through trial-and-error experimentation. The process typically begins with selecting semiconductors based on properties such as band gap and chemical stability, which indicate potential photocatalytic activity. Techniques such as sol-gel synthesis^[61], hydrothermal methods^[62], or chemical vapor deposition^[63,64] are commonly used to fabricate photocatalysts. After synthesis, thorough characterization is conducted using methods such as transmission electron microscopy (TEM), X-ray photoelectron spectroscopy (XPS), and ultraviolet-visible absorbance spectra (UV-Vis) spectroscopy to assess key features such as morphology, electronic structure, and light absorption properties. The photocatalysts are then tested under controlled laboratory conditions to determine their efficiency in processes such as pollutant degradation or water splitting.

Despite its long-standing use, this traditional approach presents several inherent limitations. The trial-and-error nature of the process is highly time-consuming and resource-intensive, requiring extensive experimentation without a guarantee of success. Additionally, the complexity of photocatalytic mechanisms, such as charge carrier dynamics, surface reactions and photon absorption, cannot be fully understood or optimized using experimental methods alone. As a result, many promising photocatalysts may go undiscovered, and the development of highly efficient systems is often impeded by the lack of predictive capability in traditional design approaches. These limitations have driven growing interest in computational techniques and data-driven methodologies to accelerate the discovery and optimization of photocatalysts^[65].

Progress of deep learning in photocatalyst design

Machine learning was first applied in materials science to predict properties and accelerate the discovery of new materials^[66,67]. Early models, including decision trees^[68] and support vector machines^[69], were employed to analyze datasets and detect patterns, thereby improving the efficiency of material screening and optimization processes. The transition from traditional machine learning to deep learning represented a paradigm shift in the field^[70], as deep learning, particularly through neural networks, allowed for the extraction of more complex, high-dimensional relationships from large datasets. This development significantly enhanced predictive accuracy and deepened our understanding of the intricate behaviors governing material properties. The integration of deep learning in photocatalyst design has experienced remarkable growth, as evidenced by the sharp increase in publications and citations from 2012 to 2024 [Figure 2A]. This upward trend underscores the expanding role of data-driven approaches in optimizing photocatalyst properties, highlighting the continued momentum of deep learning in transforming materials design. Table 2 summarizes notable research teams worldwide applying deep learning to design photocatalysts, highlighting diverse global efforts that underscore the expanding role of data-driven approaches in advancing photocatalysis.

Figure 2. General description of deep-learning-assisted photocatalyst design. (A) Trends of publications and citations in 2012-2024; (B) Clustering diagram of keywords; (C) Contribution from top 20 journals.

In Figure 2B, the keyword clustering analysis highlights three primary thematic areas in deep-learning-assisted photocatalyst design: methodology, materials, and applications. Prominent keywords such as “Learning”, “Machine”, and “Network” emphasize the critical role of machine learning techniques, particularly neural networks, in optimizing photocatalytic processes. The focus on “band gap” further suggests the integration of deep learning for theoretical predictions and mechanistic insights, supported by computational methods such as “DFT”. In terms of materials, clusters around “TiO₂” and “metal oxide” indicate widely studied photocatalysts. Keywords such as “rate” and “density” reflect efforts to optimize material properties, including electronic structure and surface functionalities. Meanwhile, keywords such as “degradation”, “water splitting” and “hydrogen” underscore the key applications of photocatalysis, such as environmental remediation through pollutant degradation and energy conversion via water splitting. The presence of terms such as “hydrogen” and “oxidation” highlights the critical role of photocatalysis in renewable energy generation and environmental sustainability. This categorization illustrates how deep learning fosters innovation across fundamental research and practical applications in photocatalysis.

High-impact journals have been instrumental in shaping the field, with several leading publications contributing significantly to the dissemination of cutting-edge research. As depicted in Figure 2C, Applied Catalysis B-Environment and Energy has emerged as a key platform. Chemical Reviews and ACS Applied Materials and Interfaces have also made substantial contributions, reinforcing the importance of this interdisciplinary approach. Additionally, journals such as Advanced Energy Materials and Journal of the American Chemical Society have been central to promoting advances in deep-learning-enhanced photocatalyst design, further solidifying the role of this technology in the future of materials discovery and development.

The photocatalyst design has seen substantial progress in materials, applications, mechanistic understanding, algorithms, and computility, as shown in Figure 3. Early photocatalysts, such as TiO₂ and ZnO, have evolved into more advanced systems, including nanomaterials, heterojunctions, quantum dots, and 1D/2D structures, greatly expanding their functionality. Initial applications focused on environmental remediation, specifically organic pollutant degradation and heavy metal reduction, but have since diversified to include CO₂ reduction, hydrogen production, sterilization, and nitrogen fixation, reflecting the growing versatility of photocatalysis in addressing global issues. Mechanistic insights have deepened, allowing precise control over electron-hole pair dynamics, band structures, and surface reactions, which has led to significant improvements in stability and efficiency. Concurrently, advancements in deep learning algorithms have enabled more accurate predictions of photocatalytic behavior, accelerating the discovery and optimization of photocatalytic systems. Additionally, enhanced computility, driven by breakthroughs in GPUs and parallel computing, has allowed researchers to simulate complex processes and manage large datasets, pushing the boundaries of what can be achieved in photocatalyst design. The interplay between deep learning and photocatalyst development is driving rapid innovation, where new machine learning models refine material discovery and mechanistic insights, while advances in photocatalysis provide rich data that fuel further algorithmic evolution. This dynamic feedback is accelerating progress toward the next generation of high-performance photocatalysts, with far-reaching implications for environmental and energy applications.

Figure 3. Progress of deep learning algorithm in photocatalyst design, application and mechanism together with computility development. DNN: Deep Neural Network; CNN: Convolutional Neural Network; GNN: Graph Neural Networks; SVR: support vector regression; RNN: Recurrent Neural Network; LSTM: Long Short-Term Memory; GRU: Gated recurrent unit; GAN: Generative Adversarial Networks; KAN: Kolmogorov-Arnold Network; KNN: K-nearest neighbors; GPR: Gaussian process regression; RFR: random forest regression; FNN: feed-forward neural network; ENIAC: electronic numerical integrator and computer; UNIVAC I: universal automatic computer I; ARM: advanced risc machines; GTX: Giga texel shader eXtreme; RTX: ray tracing texel eXtreme.

First-principles calculation

First-principles calculations are a cornerstone of modern computational materials design, offering insights into the fundamental properties of photocatalysts without relying on empirical parameters^[71,72]. Based on quantum mechanical principles, particularly DFT, first-principles methods allow for the accurate prediction of electronic structure, band gaps, and surface reactions, which are critical for understanding photocatalytic activity. By solving the Schrödinger equation for electrons, first-principles calculations describe the behavior of charge carriers, providing detailed information about electron-hole pair dynamics, defect states, and adsorption properties at the atomic level. DFT plays a pivotal role in these first-principles approaches, providing a practical and efficient method to approximate the many-body electron problem. Using DFT, properties such as charge density distribution, adsorption energies, and band structure can be determined with a level of detail that aids in deciphering the reactivity and efficiency of photocatalysts^[73]. The use of generalized gradient approximation (GGA)^[74,75] or more sophisticated hybrid functionals^[76] allows for refined predictions that align well with experimental observations, offering valuable insights that drive the rational design of advanced photocatalysts. Computational approaches to property optimization, such as cocatalyst addition and elemental doping, can be utilized to modify the electronic properties of photocatalysts, including increasing the band gap and enhancing charge separation. These calculated modifications are essential for reducing electron-hole recombination and ultimately improving photocatalytic efficiency, addressing complex design challenges effectively^[77].

As listed in Table 3, several widely used software packages are central to these calculations. VASP (Vienna Ab initio Simulation Package), one of the most commonly used tools, integrates projector augmented wave (PAW)/Perdew-Burke-Ernzerh (PBE) modules and hybrid functionals such as HSE06 for more accurate electronic structure calculations. Quantum ESPRESSO, known for its use of the PBE functional and DFT-D2 for dispersion corrections, is frequently employed for band gap and optical absorption predictions. Additionally, Materials Studio, with modules such as CASTEP, offers robust capabilities for calculating electronic properties, including band structure and light absorption coefficients. These software tools, combined with the growing computational power, enable researchers to predict the performance of potential photocatalysts with high precision, guiding experimental validation and accelerating the discovery process. These computational parameters, such as band gap, adsorption energy, and Gibbs free energy, are not only applicable to specific photocatalysts but also have broad applicability across various semiconductors for predicting photocatalytic activity. The universality of these first-principles-derived parameters across different systems provides a solid foundation for the theoretical design of photocatalysts. Through accurate calculations, these parameters help identify candidates with optimal electronic structures, thereby accelerating the development of efficient photocatalysts.

Table 3 also lists some classical works of photocatalyst designs based on first-principles calculation. One notable example exhibited in Figure 4A is the work by Shen et al., where DFT calculations played a pivotal role in designing a defective ZnTi-LDHs/Ti₃C₂O₂ photocatalyst aimed at improving the selectivity for C₂ organics in the photocatalytic reduction of CO₂^[78]. The DFT simulations revealed that the formation of a Schottky junction significantly enhanced charge separation and increased the concentration of photogenerated electrons on the photocatalyst surface, while oxygen vacancies facilitated CO₂ adsorption and activation by lowering the Gibbs free energy barrier for CO intermediate formation. This DFT-guided strategy not only optimized C₂ selectivity but also provided detailed insights into the reaction mechanisms, underscoring the crucial role of first-principles calculations in advancing the design of photocatalysts for high-value product formation. In another example shown in Figure 4B, Wang et al. utilized DFT to guide the molecular engineering of g-C₃N₄ through the integration of donor-acceptor units to address inherent structural defects^[79]. DFT calculations were employed to predict the optimal electron-donating unit, with benzaldehyde identified as the ideal candidate to enhance intramolecular charge transfer and narrow the band gap. This theoretical prediction was validated experimentally, where the modified g-C₃N₄ demonstrated a 3.73-fold improvement in photocatalytic tetracycline degradation compared to the unmodified version, indicating the critical role of DFT in optimizing photocatalysts for environmental applications. In summary, first-principles calculations, particularly those grounded in DFT, have proven indispensable in predicting and fine-tuning the electronic and structural properties of photocatalysts, thereby facilitating the rational design of highly efficient systems across a wide range of applications.

Figure 4. Typical works of photocatalyst design by first-principle calculations. (A) LDHs/Ti₃C₂O₂ designed by DFT calculation for photocatalytic reduction of CO₂ to C₂ organics^[78]. Copyright 2022, Elsevier; (B) DFT predirected design of donor-acceptor structured g-C₃N₄ for efficient photocatalytic tetracycline abatement^[79]. Copyright 2023, Wiley. DFT: Density functional theory; TDOS: total density of states; HOMO: highest occupied molecular orbital; LUMO: lowest unoccupied molecular orbital; PDOS: partial density of states; LDHs: layered double hydroxide.

HTS

HTS^[91,92] is a technique that rapidly evaluates large sets of materials through the integration of computational and experimental methods, facilitating the efficient identification of candidates with optimal properties. It is instrumental in materials science, allowing for the automation of the evaluation process, which significantly accelerates the discovery and optimization of novel materials. HTS allows for the swift screening of various compositions, surface modifications, and structural configurations, identifying those with enhanced photocatalytic performance. This method reduces the time and resources typically required for experimental validation and expands the search space for high-performance photocatalysts. The HTS process in photocatalyst design typically starts with the creation of a comprehensive library of candidate materials, generated through either computational predictions or experimental synthesis. These candidates are then evaluated in a systematic manner, employing automated computational simulations or high-throughput experimental platforms to assess critical photocatalytic properties, such as band gap, charge carrier dynamics, and surface reactivity. Based on the screening results, the most promising candidates are identified for in-depth experimental validation and further optimization, streamlining the discovery of high-performance photocatalysts.

HTS has been successfully applied in several key studies to accelerate the discovery and optimization of photocatalysts, demonstrating its effectiveness in identifying high-performance candidates [Figure 5]. In a study by Yang et al., HTS combined with machine learning was employed to explore ternary organic heterojunction photocatalysts for sacrificial hydrogen evolution^[93]. By conducting 736 experiments out of a possible 4,320 combinations, the most active systems achieved photocatalytic hydrogen production rates exceeding 500 mmol·g^-1·h^-1, showcasing the power of HTS in rapidly identifying high-performance photocatalysts. Sa et al. employed HTS to identify promising 2D Janus III-VI van der Waals heterostructures for solar energy applications^[94]. From a total of 19,926 heterostructures, 1,035 were selected based on stability, and further application-driven screening revealed 66 photocatalysts and 71 solar cell candidates, highlighting the efficiency of HTS in pinpointing top-performing systems for energy conversion. Singh et al. conducted a large-scale HTS of 68,860 potential photocathodes for CO₂ reduction, utilizing first-principles computations to assess synthesizability, corrosion resistance, visible-light absorption, and electronic structure compatibility with fuel synthesis^[95]. Out of this vast pool, only 52 systems met all criteria, with 43 newly discovered candidates emerging as promising for further exploration, underscoring the capacity of high-throughput methods to efficiently filter through vast datasets and identify high-performing photocatalysts. Thermodynamic stability is initially prioritized to eliminate impractical candidates and minimize computational burden, while band edge alignment is assessed last due to its computational intensity, applied only to candidates that meet all preceding criteria. Wang et al. employed a data-driven, large-scale screening approach to explore intrinsic direct-gap 2D systems for photocatalytic water splitting^[96]. Through a three-stage process of high-throughput DFT calculations, 16 candidates were identified as highly promising for efficient solar-to-hydrogen conversion, demonstrating the power of HTS in uncovering potential photocatalysts from vast, uncharted chemical spaces. HTS has emerged as a powerful tool for the rapid identification of high-performance photocatalysts by systematically exploring vast chemical spaces through a combination of computational modeling and experimental techniques. This approach not only expedites the discovery of photocatalysts with enhanced properties but also provides an efficient pathway for optimizing systems prior to in-depth experimental validation.

Figure 5. Photocatalyst designs by high-throughput screening. (A) Machine learning accelerated exploration of ternary organic heterojunction photocatalysts for sacrificial hydrogen evolution^[93]. Copyright 2023, American Chemical Society; (B) High-throughput computational screening of Janus 2D III-VI van der Waals heterostructures for solar energy applications^[94]. Copyright 2022, American Chemical Society; (C) Data-driven materials discovery of robust and synthesizable photocatalysts for CO₂ reduction^[95]. Copyright 2019, Springer; (D) Data-driven discovery of intrinsic direct-gap 2D materials as potential photocatalysts for efficient water splitting^[96]. Copyright 2024, American Chemical Society. PBE: Perdew-Burke-Ernzerh; ML: machine learning; HSE: Heyd-Scuseria-Ernzerhof.

Data collection and preprocess

Data forms the essential backbone of any deep learning model. In the photocatalyst design, data is typically derived from experimental measurements, published literature, and high-throughput computational simulations. These sources provide extensive information on material properties, performance metrics, and structural features, offering valuable insights into photocatalytic behavior. The process of data collection and preprocessing is critical for ensuring dataset quality and consistency. Data cleaning is conducted to address issues such as duplicate entries, missing values, and errors. To handle missing data, various imputation strategies were employed, tailored to the characteristics of the dataset and the specific attributes^[97]. A hybrid framework leveraging Markov Logic Networks (MLNs) was implemented to infer probable values by supplementing insufficient integrity constraints with learned instantiated rules, thus establishing a robust data cleaning mechanism^[98]. Additionally, an "active label cleaning" strategy^[99] was utilized to improve data quality, particularly in datasets with label noise. This method prioritized samples for re-annotation based on estimated label correctness and labeling difficulty, enabling a more targeted and efficient correction process compared to random selection, thereby optimizing data preparation and significantly enhancing the overall quality of the training dataset.

Standardization techniques such as min-max scaling and z-score normalization^[100] are employed to ensure uniformity across different data sources, particularly when integrating datasets with varying formats. For feature extraction, methods such as principal component analysis (PCA) and autoencoders^[101,102] are commonly used to reduce dimensionality and isolate the most relevant characteristics, transforming raw data into a structured format suitable for deep learning model inputs. Table 4 lists the popular databases available for photocatalyst design. These databases offer a wide array of data, ranging from crystal structures (ICSD, CSD, COD) to computational predictions and material properties (Materials Project, AFLOW, NREL). Some databases, including MatWeb and MatNavi, focus on experimental data, while others, such as the Open Quantum Materials Database (OQMD), provide computationally derived properties and high-throughput calculations, enabling researchers to explore new material candidates and optimize photocatalytic performance.

Feature engineering and descriptors

Feature engineering^[103] is a fundamental step in constructing deep learning models for the design of photocatalysts, where raw data must be transformed into structured and informative features that reflect the photocatalyst properties. This process directly affects the model capability to capture complex relationships within the data, ultimately determining its accuracy and predictive power. A crucial aspect of feature engineering is the use of descriptors, which quantitatively represent the physical and chemical characteristics of photocatalysts. Selecting the appropriate descriptors is essential for ensuring the model focuses on the most relevant features, significantly enhancing its ability to predict photocatalytic performance and optimize photocatalyst design^[104].

Common descriptors in photocatalyst design include electronic structure, geometric structure, and surface properties, as listed in Table 5. Electronic structure descriptors, such as band gap, valence band maximum (VBM), conduction band minimum (CBM), and DOS, are crucial for understanding light absorption and charge carrier dynamics. Geometric descriptors, including size, morphology, crystal planes, and lattice symmetry, provide insights into the spatial arrangement of atoms and its effect on photocatalytic activity. Surface property descriptors, including specific surface area, pore structure, and surface functional groups, influence reactivity and adsorption behavior. These descriptors are essential for building deep learning models that can accurately predict and optimize photocatalytic performance across various applications, such as water splitting, pollutant degradation, and hydrogen production. Furthermore, the descriptors encompass not only electronic structure parameters but also geometrical and surface characteristics, such as pore structure and surface functional groups. These features help establish nonlinear correlations between semiconductor properties and photocatalytic performance, facilitating a generalized performance prediction across various photocatalysts. Deep learning approaches leverage these diverse features to extract patterns from large datasets, aiding in identifying and optimizing the most promising candidate and enhancing the efficiency of photocatalyst design and application.

Model construction and training

The construction and training of deep learning models for photocatalyst design involve defining the model architecture, selecting relevant descriptors as input features, and optimizing the model parameters. Typical deep learning models with their features and applications are listed in Table 6. Traditional deep learning models, including deep neural networks (DNNs)^[115], are widely used for classification and regression in predicting photocatalytic properties. Convolutional neural networks (CNNs)^[116] analyze spatial data, aiding in surface and structural characterization. Recurrent neural networks (RNNs)^[117] and Long Short-Term Memory (LSTM)^[118] handle time-series data, crucial for studying reaction kinetics. State-of-the-art models such as transformers^[119] capture complex dependencies using self-attention mechanisms^[120]. Shapley additive explanations (SHAP)^[121] and Gradient-weighted class activation mapping (Grad-CAM)^[122] improve model interpretability by highlighting key features. Kolmogorov-Arnold networks (KANs)^[123] are adept at modeling nonlinear systems, revealing insights into complex photocatalytic processes.

Model validation and evaluation

Model validation and evaluation are essential steps in ensuring the robustness and effectiveness of deep learning models in photocatalyst design. These processes verify predictive accuracy and generalization ability, ensuring alignment between model predictions and experimental data, while mitigating issues such as overfitting and underfitting^[124]. Key evaluation metrics include mean squared error (MSE) and root MSE (RMSE), which quantify prediction errors, and R-squared (R²), which measures the proportion of variance explained by the model. Cross-validation and hold-out methods^[125] are commonly used to partition datasets into training and testing subsets, providing unbiased estimates of model performance.

Discovery of novel photocatalysts

Deep learning has revolutionized the discovery of novel photocatalysts by enabling the rapid screening of vast chemical spaces and predicting materials with superior photocatalytic properties. By leveraging data-driven models, deep learning accelerates the identification of candidates optimized for light absorption, charge separation, and photocatalytic efficiency, significantly reducing the time and resources required compared to traditional experimental approaches.

A recent study^[126] utilized machine learning to design perovskite oxide materials for photocatalytic water splitting, addressing the inefficiency of traditional trial-and-error methods in discovering new visible-light photocatalysts, as shown in Figure 7A. The study constructed structural-property models to predict hydrogen production rates and optimal bandgaps using algorithms such as gradient boosting regression (GBR), support vector regression (SVR), and backpropagation artificial neural networks (BPANN). The feature selection process began with an initial set of 24 features, comprising 18 atomic parameters and six experimental conditions. These features were then reduced to 20 through an initial filtering step, focusing on eliminating redundancies. Subsequently, advanced selection techniques such as Max Relevance Min Redundancy (mRMR) and embedded feature selection were applied, ultimately narrowing the set to 7-9 key features that significantly enhanced model accuracy while minimizing overfitting. The BPANN model achieved the highest prediction accuracy for hydrogen production, while GBR showed the best performance for bandgap prediction. Ultimately, 14 promising perovskite photocatalysts were identified from 30,000 candidates, and two online web servers were developed to share the prediction models, allowing public access for future research. Huang et al. applied machine learning techniques to address the complex task of selecting optimal cocatalysts for BiVO₄ photoanodes in photoelectrochemical systems, as exhibited in Figure 7B^[127]. The approach involved training multi-layer perceptron neural networks and tree-based ensemble models to predict high-performance cocatalyst-photoanode combinations. The random forest model achieved a classification accuracy of 96.30%, with cocatalyst type and preparation method identified as the most influential factors affecting photocatalytic performance. Additionally, the study utilized SHAP to derive heuristic rules for guiding the selection of promising cocatalyst/photoanode systems.

Figure 7. Machine learning aided discovery of novel photocatalysts. (A) Machine learning aided design of perovskite oxide materials for photocatalytic water splitting^[126]. Copyright 2021, Elsevier; (B) Machine learning strategy for designing high-performance photoanode catalysts^[127]. Copyright 2023, Royal Society of Chemistry.

There are also some successful cases in novel photocatalyst design accelerated by deep learning. Machine learning has been applied to the design of TiO₂-coated glass for improved NOx removal efficiency^[128]. Additionally, core-shell Au-silica nanoparticles^[129], lead-free Bi-based perovskites^[130], and g-C₃N₄-based single-atom photocatalysts^[131] have been identified through deep-learning approaches. Choudhary et al. developed the InterMat framework, which combines DFT and GNN to predict band offsets in semiconductor interfaces, enabling the rapid screening of over 1.4 trillion potential interfaces^[132]. This large-scale model, which is publicly available, offers an efficient tool for discovering novel photocatalysts. The application of deep learning in the discovery of novel photocatalysts has significantly accelerated the identification and design of high-performance photocatalytic systems, demonstrating its transformative impact across various applications.

Microstructure design of photocatalysts

The microstructure of photocatalysts is critical in determining their efficiency, as it directly affects light absorption, charge separation, and surface activity. Microstructures such as heterojunctions provide significant advantages by promoting more efficient charge separation and transfer at interfaces, which in turn enhances overall photocatalytic performance. Guevarra et al. employed a deep learning-based materials structure-property factorization (MSPF) to accelerate the design of photocatalyst structures^[133] [Figure 8]. This approach utilized Deep Reasoning Networks (DRNets) for phase mapping and matrix factorization for modeling key properties, revealing synergistic interactions between BiVO₄-like and Cu-based phases that significantly enhanced photocatalytic performance. The study highlights the role of deep learning in advancing structural innovation and optimizing complex photocatalytic systems.

Figure 8. Phase mapping in the Bi-Cu-V oxide system. (A) Photographs of 4 sputter-deposited composition libraries in Bi–Cu–V oxide system; (B) Representative XRD patterns collected on the composition libraries, where each XRD pattern is plotted at the location of the composition in the Bi–Cu–V composition graph; (C) The DRNets solution for one composition, showing that the XRD pattern of the Bi_0.11Cu_0.39V_0.50 oxide sample is composed of 42% BiVO₄ and 58% Cu₂V₂O₇; (D) The resulting phase map of the primary 14 of the 21 phases identified by DRNets, where each phase has a unique symbol whose point size indicates phase concentration in the composition graph^[133]. Copyright 2022, Springer Nature. XRD: X-ray diffraction; DRNets: deep reasoning networks.

Building on recent advances, further efforts have utilized deep learning models to explore complex structural modifications. Jiang et al. developed an interpretable CNN to predict active sites involved in TiO₂-mediated photocatalytic degradation of organic contaminants^[134]. The model employed EfficientNet to extract critical structural features and utilized Grad-CAM to highlight molecular regions most influential in determining degradation rates. Similarly, Zhou et al. applied graph CNNs (GCNNs) to establish a structure-property relationship for nanoporous cobalt zirconate, predicting ammonia yield from photocatalytic nitrogen fixation based on structural parameters such as pore volume and surface area^[135]. Therefore, the integration of deep learning in microstructural design has significantly advanced the optimization of photocatalyst structures, leading to enhanced performance across various systems, including heterostructures and perovskite oxides.

Property optimization of photocatalysts

Predicting key properties of photocatalysts by deep learning models has emerged as a critical research area, which offers high-precision predictions, greatly enhancing the optimization of photocatalytic performance. Bonke^[136] employed machine learning algorithms to optimize the multivariable performance of a self-assembled photocatalytic system for CO₂ reduction [Figure 9]. By defining a holistic performance metric that integrates multiple figures of merit, the model guided experimental optimization of yield, quantum yield, turnover number, frequency, and selectivity simultaneously. Additionally, Wang et al. demonstrated the efficacy of intelligent algorithms, particularly artificial neural networks (ANNs), in optimizing photocatalyst properties^[137]. These models enable accurate predictions of key photocatalytic performance metrics, significantly reducing the time and resources required for experimental iterations, thereby accelerating the design of efficient photocatalysts. This methodology provided a standardized approach for optimizing multimetric photocatalytic systems, offering deeper insights into performance bottlenecks and accelerating advancements in photocatalysis.

Figure 9. Multivariable multimetric optimization of self-assembled photocatalytic CO₂ reduction performance using machine learning algorithms. (A) Overview of the workflow; (B) Holistic improvement measured by objective function alongside improvement of both Yield_CO and TOF_CO; (C) Ascending Yield_CO and corresponding TOF_CO^[136]. Copyright 2024, American Chemical Society. TON: Turnovers for CO formation; TOF: turnover frequency; QY: quantum yield.

Beyond multivariable optimization, machine learning has been extensively applied to enhance the performance of diverse photocatalytic systems. For instance, random forest models have been used to assess the contribution of pretreatment methods on TiO₂ surfaces for CO₂ reduction^[138], while CatBoost models have optimized malachite green degradation efficiency using noble metal-doped BiFeO₃ (NM-BiFeO₃) composites^[139]. Additionally, decision tree models have predicted gas and liquid product types in photocatalytic CO₂ reduction over metal-organic frameworks^[140]. Gated recurrent unit (GRU) neural networks have forecasted hydrogen production in continuous water splitting systems^[141]. Moreover, GBR methods have modeled the degradation of persistent pollutants such as perfluorooctanoic acid (PFOA)^[142]. Through the integration of experimental data and theoretical models, deep learning has significantly enhanced the ability to predict and optimize key photocatalytic properties, such as light absorption, photocatalytic efficiency, and degradation performance, driving improvements across diverse applications in photocatalyst design.

Novel deep learning approaches of photocatalyst design

Beyond the common deep learning models, novel approaches continue to emerge, offering enhanced capabilities for photocatalyst design. These advanced methods improve adaptability and the ability to handle complex relationships in large datasets, driving significant progress in the field. Figure 10 reveals a hybrid method combining computational fluid dynamics (CFD), ANNs, and genetic algorithms (GA) for the analysis and optimization of micro-photocatalytic reactors aimed at NOx abatement^[143]. CFD simulations were used to investigate the effects of key variables, while ANN models were developed to predict NO conversion with high accuracy (R² = 0.9997). GA was then employed to optimize operating conditions, achieving full NO conversion (100%) under specific parameters, with residence time identified as the most influential factor. A multi-objective optimization further applied GA to maximize NO consumption while minimizing pressure drop, offering valuable insights into balancing performance metrics.

Figure 10. CFD, ANN and GA as a hybrid method for the analysis and optimization of micro-photocatalytic reactors for NOx abatement^[143]. Copyright 2021, Elsevier. CFD: Computational fluid dynamics; ANN: artificial neural network; GA: genetic algorithm; MSE: mean squared error.

In addition to hybrid CFD-ANN-GA approaches, recent advancements have introduced innovative methods to enhance the design and optimization of photocatalysts. Li et al. combined machine learning with high-throughput experimentation to identify organic molecules with high photocatalytic activity, successfully predicting molecular performance and significantly reducing experimental costs^[144]. Similarly, Parmar et al. applied a comparative approach using ANN and response surface methodology to optimize ciprofloxacin degradation, achieving high predictive accuracy and offering a novel strategy for pharmaceutical waste removal through computational modeling^[145]. The emergence of novel deep learning approaches has significantly enhanced photocatalyst discovery and optimization, providing innovative solutions that improve predictive accuracy and reduce experimental costs across various applications.

Application exploration for photocatalysts

Deep learning has significantly broadened the scope of photocatalyst applications by enabling precise predictions of system performance in complex environments. This advancement accelerates the discovery of new uses in areas such as environmental remediation, energy conversion, and medical applications. Malayeri et al. leveraged ANN and GA to optimize photocatalytic oxidation (PCO) reactors for air purification, thereby expanding the applicability of PCO technology in indoor environments^[146]. By accurately predicting volatile organic compound (VOC) and by-product concentrations, the ANN model facilitated the identification of optimal operating conditions, significantly enhancing the efficiency of harmful compound removal. This approach also minimized the formation of toxic by-products, addressing a critical challenge in the widespread adoption of PCO technology. The study illustrated how advanced machine learning techniques can broaden PCO application in the development of safer and more effective air purification systems [Figure 11].

Figure 11. Application of artificial neural network and genetic algorithm in optimization of photocatalytic oxidation reactor for air purifier design^[146]. Copyright 2023, Elsevier. ANN: Artificial neural network; GA: genetic algorithm; VOC: volatile organic compound.

Beyond air purification, deep learning has significantly expanded the applications of photocatalysts in fields such as water treatment^[147], energy conversion^[148], and medical applications^[149]. In water treatment, machine learning has optimized the photocatalytic degradation of pollutants such as dyes and pharmaceuticals^[150], enhancing process efficiency and introducing innovative treatment approaches. In the energy and medical sectors, deep learning has driven advancements in solar-to-hydrogen conversion^[151] and the design of antimicrobial surfaces^[152]. Therefore, it has greatly expanded the scope of photocatalyst applications, facilitating advancements in cutting-edge areas, driving innovative solutions to complex global challenges.

Mechanism insights for photocatalysis

Deep learning has become an essential tool in unraveling photocatalytic reaction mechanisms by analyzing reaction pathways and intermediates. These models offer profound insights into the fundamental processes of photocatalysis, enabling more precise experimental design and the optimization of photocatalytic systems. Kim et al. utilized a machine learning approach to unravel the complex mechanisms influencing the photocatalytic reaction rate constant (k) in semiconductor-based photocatalysts [Figure 12], specifically for dye removal applications^[153]. By employing a decision tree model and SHAP feature selection, the analysis identified 11 key input features that significantly impacted the reaction rate. Experimental conditions emerged as the most influential factor (59%), followed by atomic composition (39%), offering valuable insights into how process parameters and co-catalysts interact to affect photocatalytic performance. This study demonstrated the ability of machine learning models to enhance our understanding of reaction mechanisms by elucidating the interactions between multiple process variables.

Figure 12. Machine learning analysis to interpret the effect of the photocatalytic reaction rate constant (k) of semiconductor-based photocatalysts on dye removal^[153]. Copyright 2024, Elsevier. SVR: Support vector regression; SHAP: shapley additive explanations; LGBM: light gradient boosting machine; XGB: eXtreme gradient boosting.

Beyond reaction rate analysis, another work^[154] explored the excitonic effects in nearly 50 photocatalysts for CO₂ reduction, an often overlooked but crucial aspect of photocatalytic performance. By applying the Bethe-Salpeter formalism, this study identified six promising materials through optical property screening, offering new insights into the role of exciton binding energies in enhancing solar-energy harvesting applications. The works highlighted how deep-learning approaches were advancing the understanding of photocatalytic mechanisms, from reaction kinetics to excitonic effects, offering valuable insights for optimizing photocatalyst performance. Therefore, deep learning approaches in photocatalyst design have significantly advanced the field by enabling the discovery of novel photocatalysts, optimizing microstructures, and refining material properties for enhanced performance. These approaches have also expanded the range of photocatalyst applications, from environmental remediation to energy conversion and medical uses, while providing critical insights into the underlying mechanisms driving photocatalytic reactions. As deep learning models evolve, their integration with experimental data and computational methods continues to push the boundaries of photocatalyst design, paving the way for future breakthroughs in cutting-edge research areas.

Dataset availability and quality

One of the fundamental challenges in deep-learning-assisted photocatalyst design is the acquisition of high-quality, comprehensive datasets. Photocatalytic reactions involve complex interactions between factors such as composition, structure, environmental conditions, and experimental setups, making it difficult to compile standardized data across different studies. Existing datasets are often fragmented, with inconsistencies in reporting, missing information, and limited experimental reproducibility. The scarcity of large, labeled datasets also limits the training of robust deep learning models, reducing their ability to generalize across diverse photocatalytic systems. Addressing these challenges requires the aggregation of data from multiple sources and the development of uniform data collection protocols to ensure data quality and consistency.

To manage and organize the growing volume of photocatalytic data, effective data management strategies are essential. Implementing standardized data formats and ontologies can streamline the integration of diverse datasets from experimental and computational studies. Advanced data cleaning techniques, such as automated outlier detection and missing value imputation, are necessary to ensure the reliability of the data used for model training. The development of centralized databases or repositories for photocatalytic research will facilitate data sharing and collaboration within the scientific community. Cloud-based platforms and machine learning pipelines can further enhance data processing efficiency and scalability, ensuring continuous model improvement as more data becomes available.

Interpretable deep learning model

The “black box” nature of deep learning models poses a significant challenge in photocatalyst design, limiting the understanding of how model predictions are made. Improving the transparency of these models is essential to close the gap between predictions and physical experiments, enabling a clearer interpretation of how specific features influence photocatalytic performance. Approaches such as SHAP and local interpretable model-agnostic explanations (LIME) are proving effective in identifying the most critical input variables. The KAN model further enhances interpretability by providing a more quantifiable explanation of the relationships between inputs and outputs. Additionally, integrating physics-based principles through models such as PINN aligns the deep learning process more closely with known physical laws. These advances will facilitate a deeper understanding of photocatalytic mechanisms and promote the broader application of deep learning in photocatalyst design.

Scalability and computing resources

The increasing complexity and size of deep learning models, especially large-scale models, used in photocatalyst design require significant computational resources, presenting a challenge in managing vast datasets and running high-dimensional simulations. As large models such as transformers and foundation models become more prevalent, the demand for scalable computing infrastructures, such as high-performance computing clusters and cloud-based systems, has become essential to handle the intense computational workloads. Hardware advancements, particularly in the development of GPUs and TPUs, are driving improvements in training speed and model efficiency. Moreover, the development of more efficient architectures, along with advancements in large model fine-tuning, is reducing computational cost while maintaining high predictive accuracy. Techniques such as model pruning, quantization, and efficient distributed computing frameworks enable parallelization, allowing the processing of massive datasets across multiple machines. These large models open the door for unprecedented accuracy in predictions, but the hardware and algorithmic advancements will be critical to overcoming the computational bottlenecks associated with their scale. The integration of these advancements is key for making large-scale photocatalyst design feasible and scalable.

Reversal design of photocatalysts

Reversal design is a methodology that works backward from a desired property or performance target to identify the optimal structures or configurations, offering a powerful approach in materials science. In photocatalyst design, reversal design enables the identification of systems with optimized performance by working backward from specific targets, such as light absorption or photocatalytic efficiency, to tailor structural or compositional features. This methodology accelerates the discovery process by narrowing the search space to systems that meet predefined performance metrics. Deep learning enhances reversal design by using predictive models that efficiently map the relationship between structures and properties, allowing rapid iteration toward optimal solutions. It can generate new candidates through generative models and apply reinforcement learning to refine designs based on experimental feedback. With its ability to explore vast design spaces, deep learning-driven reversal design is transforming the way photocatalysts are discovered and developed.

Multiscale modeling and computation

Photocatalytic reactions inherently span multiple scales, from electron-level interactions to macroscopic phenomena, posing significant challenges for comprehensive modeling. The integration of models that describe different scales - such as atomic-scale electron dynamics, charge carrier transport, and bulk-phase mass transfer - into a unified multiscale framework remains a critical hurdle in photocatalyst design. Multiscale deep learning models provide an opportunity to bridge these gaps, enabling more accurate performance predictions by incorporating insights from various scales. This requires combining quantum simulations, molecular dynamics, and continuum models, which presents not only computational demands but also algorithmic complexity. Successful advances in multiscale modeling could yield deeper insights into how atomic and molecular processes influence macroscopic behavior, driving the development of more efficient photocatalysts. Additionally, ensuring the interpretability and practical applicability of these models is essential for their experimental validation. Addressing these challenges is a pivotal goal in advancing both photocatalyst design and materials informatics.

Future directions amidst challenges

Despite their substantial advantages in reducing experimental costs and time, deep learning and other computational methodologies present several inherent challenges. A primary challenge is the need for further enhancement of computationally designed photocatalysts. While integrated computational approaches can identify promising candidates, these photocatalysts frequently require additional modifications - such as surface engineering, doping, or structural optimization - to achieve optimal performance^[162]. Addressing these limitations often involves iterative cycles of computational predictions followed by experimental validation to refine photocatalyst properties^[163]. Another critical challenge involves scaling these technologies from laboratory settings to industrial applications. Photocatalysts that demonstrate exceptional performance on a small scale often encounter significant obstacles in maintaining efficiency and stability during scale-up. Overcoming these scale-up challenges requires advancements in reactor design, synthesis methodologies, and engineering practices to ensure economically feasible production processes. Moreover, the successful commercialization of photocatalyst technologies necessitates a multidisciplinary approach. Issues such as stability, selectivity, and overall efficiency cannot be fully addressed within a single discipline. Effective commercialization requires collaboration among chemists, materials scientists, chemical engineers, and industry professionals to bridge the gap between laboratory research and market-ready technologies, ultimately facilitating the transition to commercially viable photocatalytic solutions. In conclusion, although challenges remain, rapid progress in computational tools, coupled with targeted experimental efforts and interdisciplinary collaboration, suggests a promising future for photocatalyst design and application. With ongoing innovation, these technologies hold substantial potential for contributing to sustainable energy and environmental applications.