DFT and hybrid classical–quantum machine learning integration for photocatalyst discovery and hydrogen production

3.1 Solar–NIR radiation

The utilization of solar energy in the photocatalytic splitting of water has been pivotal in advancing clean energy technologies. Solar energy is abundant, cost-free, and inherently sustainable, making it the most attractive light source for large-scale hydrogen production. Traditional photocatalytic materials, such as titanium dioxide (TiO₂), were initially developed to absorb ultraviolet (UV) light, which comprises only about 4–5 % of the solar spectrum. However, as research advanced, it became evident that harnessing the visible and near-infrared (NIR) portions of the spectrum was critical for improving overall solar-to-hydrogen (STH) efficiency (Agrawal et al. 2024; Kumar et al. 2024; Patel et al. 2024; Priya et al. 2024). Modern solar simulators are of late considered a practical step to simulating electrochemical cells for water splitting in the laboratory; Figure 5 showcases a typical solar-simulator in the University College London for water splitting.

Figure 5:

Solar simulator-driven photoelectrochemical water splitting, illustrating the experimental setup used to evaluate photocatalyst efficiency under simulated sunlight. Adapted with permission from the University College London website (University College London 2024).

Recent innovations in material science (Ansari and Sillanpaa 2021; Cao et al. 2022) have enabled the development of semiconductors and doped materials capable of extending light absorption into the NIR range (700–2,500 nm). Upconversion nanoparticles (UCNPs), for instance, have emerged as effective mediators that convert NIR photons into higher-energy visible photons. By integrating UCNPs with conventional photocatalysts such as TiO₂ or WO₃, researchers have achieved significant improvements in water-splitting efficiency under NIR irradiation.

Moreover, tandem systems employing multi-junction photocatalytic materials have demonstrated improved photon utilization by sequentially absorbing photons across the UV, visible, and NIR regions. For example, bismuth vanadate (BiVO₄) coupled with a silicon-based photoelectrode has shown promise in maximizing STH conversion (Jin et al. 2022).

Beyond photon absorption, the photothermal effects (Qureshi et al. 2024) of NIR radiation have gained attention. Photothermal catalysis leverages localized heating effects induced by NIR photons to enhance reaction kinetics at the catalyst surface. Plasmonic nanomaterials, such as gold and silver nanoparticles, are often used to amplify these effects, leading to improved hydrogen evolution rates.

3.2 Mercury discharge lamps

Historically, mercury discharge lamps as displayed in Figure 6 played a pivotal role in the early stages of photocatalytic research. These lamps emit a broad spectrum of light, including UV, which was instrumental in validating the photocatalytic activity of materials like TiO₂.

Figure 6:

High-pressure mercury lamp marketed for its broad emission spectrum ranging from 200 nm to 800 nm, commonly used in photochemical experiments. Adapted with permission from the Shilpent website (Shilpent 2024).

Mercury discharge lamps offer high UV intensity (Broxtermann et al. 2017), making them suitable for studying UV-responsive photocatalysts. However, their inefficiency in energy consumption and environmental concerns related to mercury disposal have limited their modern-day applicability. Despite this, these lamps remain valuable tools as elucidated in Figure 6 for laboratory-scale experiments that require precise control over UV–vis light exposure (Horikoshi et al. 2019; Kubiak 2024; Meinhardova et al. 2023).

One notable application of mercury discharge lamps is their use in evaluating quantum efficiencies under controlled conditions. Researchers have used these lamps to identify reaction mechanisms by isolating the effects of UV-induced charge carriers. However, the lack of compatibility with visible or NIR-responsive photocatalysts has led to a gradual decline in their use, particularly as the field transitions to more sustainable and efficient light sources.

3.3 Light emitting diodes (LEDs)

The advent of light-emitting diodes (LEDs) (Bhattarai et al. 2024) has revolutionized photocatalysis by providing an energy-efficient, tunable, and environmentally friendly alternative to traditional light sources. LEDs offer narrow-band emission, enabling researchers to tailor light wavelengths to match the absorption spectra of specific photocatalysts.

LEDs consume significantly less power than mercury discharge lamps and have a longer operational lifespan. Additionally, their ability to emit light across the UV, visible, and NIR spectra makes them versatile for a wide range of photocatalytic applications. By utilizing LEDs with customizable wavelengths, researchers can optimize photocatalytic processes for various materials and reaction conditions. For example, blue and red LEDs are often paired with visible-light-responsive photocatalysts like BiVO₄ and cadmium sulfide (CdS), respectively.

Recent studies have demonstrated the integration of LED arrays in scalable photocatalytic reactors as demonstrated in Figure 7. Such systems combine the precision of wavelength control with the ability to uniformly illuminate large reaction surfaces. Additionally, the pulsed operation of LEDs has been explored as a method to enhance charge separation in photocatalysts. Pulsed LED systems intermittently provide light to the reaction system, allowing time for photogenerated charge carriers to participate in redox reactions before recombining.

Figure 7:

Schematic of the water splitting and storage process under white-LED irradiation, highlighting the photocatalytic flow enabled by modified g-C₃N₄ materials. Reproduced from (Tarighati Sareshkeh et al. 2023), scientific reports, 13, 15079 (2023), nature publishing group, under the terms of the creative commons attribution 4.0 international license (CC BY 4.0).

The evolution of light-driven photoreactions is closely tied to the development of photocatalytic materials optimized for specific light sources (Balarabe and Maity 2022; Khodadadian 2019). For instance, UV-responsive materials like TiO₂ were initially the primary focus, but advancements in doping, heterojunction formation, and plasmonic effects have expanded the applicability of materials to the visible and NIR regions. Furthermore, hybrid light sources that combine solar simulators with LEDs have been employed to mimic real-world conditions, enabling more accurate assessments of photocatalytic performance.

3.4 Emerging trends in light sources

Emerging light sources such as perovskite-based emitters and laser-induced systems represent innovative directions in photocatalysis, as summarized in Table 2. Perovskite materials (Lim et al. 2019; Zhu et al. 2019), renowned for their tunable bandgaps and high quantum yields, are increasingly explored for hybrid configurations where they serve as both absorbers and emitters.

Their spectral flexibility enables precise matching with semiconductor photocatalysts. Meanwhile, wireless light sources such as laser-based systems are being employed to achieve spatially resolved excitation, allowing for selective activation of catalytic sites. Although still in developmental stages, these techniques offer promising opportunities for precision-controlled and patterned photocatalytic reactions.

4.1 Emerging binary semiconductor materials

Binary semiconductor systems are gaining traction due to their synergistic properties. These materials combine the strengths of individual components, such as enhanced light absorption, improved charge separation, and stability (Meyer et al. 2012).

The combination of cadmium sulfide (CdS) and zinc sulfide (ZnS) in core–shell structures helps reduce photocorrosion and improve charge separation as demonstrated by Vamvasakis et al. (2023) in Figure 8. This system is particularly effective under visible light, achieving high hydrogen evolution rates (Jiang et al. 2016). As a metal-free semiconductor, graphitic carbon nitride (g-C₃N₄) offers a unique approach to photocatalysis, as in the case of (Wang et al. 2024a). Its bandgap of around 2.7 eV allows it to absorb visible light, and its compatibility with various co-catalysts makes it versatile. Coupling g-C₃N₄ with noble metals like platinum or ruthenium significantly enhances its photocatalytic activity (Vasilchenko et al. 2022). Perovskite materials, with their adjustable bandgaps and excellent light absorption properties, are emerging as promising candidates for high-efficiency photocatalysis. However, stability issues continue to be a significant challenge.

Figure 8:

Demonstration of CdS/ZnS nanocomposites as a binary semiconductor system operating under visible light for enhanced photocatalytic hydrogen generation. Reproduced from (Vamvasakis et al. 2023), nanomaterials, 13(17), 2426 (2023), MDPI, under the terms of the creative commons attribution 4.0 international license (CC BY 4.0).

Developing semiconductor materials for photocatalysis still faces several challenges, despite progress. Achieving the right balance between light absorption and charge carrier movement in the bandgap remains a critical issue. Many semiconductors, especially those sensitive to visible light, are prone to photodamage or deterioration during reactions. The rapid recombination of the generated electrons and holes reduces overall efficiency, requiring innovative approaches to separate charges. Scaling up the production of high-performing semiconductors while maintaining consistency and cost-effectiveness is a significant obstacle (Villa et al. 2021).

Research is now concentrating on combining improved materials and methods to solve these problems. Using machine learning in material design allows for predictive modeling and optimization, which is speeding up the discovery of new semiconductor materials. The use of quantum dots as light-absorbing agents shows potential for expanding light absorption into the near-infrared region. Combining organic and inorganic materials in hybrid systems provides a way to achieve both high efficiency and stability in photocatalysts (Ma et al. 2023).

4.2 Hybrid quantum dots and upconversion materials

Hybrid quantum dots (QDs) and upconversion materials (Chen et al. 2022; Gui et al. 2024) represent cutting-edge advancements in the field of photocatalytic semiconductors. These materials are designed to overcome the limitations of conventional semiconductors, such as restricted light absorption and inefficient charge carrier dynamics. By combining quantum dots with upconversion nanoparticles, researchers have created hybrid systems capable of extending light absorption into the near-infrared (NIR) region and enhancing photocatalytic efficiency.

4.3 Material rate kinetics, thermodynamic thresholds, and light absorption

The performance of photocatalytic systems hinges not only on material properties but also on their intrinsic kinetics and thermodynamic constraints. A critical benchmark in hydrogen evolution reactions (HER) is the overpotential (η), the excess potential required beyond the standard redox potential to drive water splitting. For effective H₂ generation, the conduction band minimum (CBM) of the photocatalyst must lie more negative than the H⁺/H₂ redox potential (0 V vs. NHE), while the valence band maximum (VBM) should be more positive than the O₂/H₂O redox potential (+1.23 V vs. NHE).

Wang et al. (2023b) highlighted that the rate of hydrogen evolution can be described by the Arrhenius-like relation R ∝ e − Δ G ‡ / k T , where ΔG^‡ is the activation free energy, modulated by surface defect states and co-catalyst binding. Their kinetic simulations show that TiO₂ and BiVO₄ systems with optimized co-catalyst loadings exhibit rate constants in the range of 10⁴–10⁵ mol m⁻² s⁻¹ under visible light.

Furthermore, light absorption coefficients (α) govern the depth of photon penetration and hence the effective carrier generation. Li et al. (2025b) employed time-resolved spectroscopy to correlate high α values (>10⁵ cm⁻¹) in narrow bandgap semiconductors with superior internal quantum efficiency (IQE), especially in NIR-driven systems. The interplay of α, carrier diffusion length, and surface recombination velocity (S) ultimately controls the external quantum efficiency (EQE).

To guide material selection, Figure 9 illustrates a simplified schematic of the band positions, overpotentials, and light penetration profiles in representative TiO₂, ZnO, BiVO₄ and other systems shown in Table 3, highlighting their HER thermodynamics and optical constraints.

Figure 9:

Schematic of band alignment, overpotential thresholds, and representative light absorption depths for selected photocatalysts.

5.1 Role of co-catalysts

Co-catalysts play a pivotal role in enhancing the efficiency of photocatalytic water splitting by facilitating charge separation and catalyzing surface reactions (Tian et al. 2023). Noble metals like platinum (Pt) are highly effective hydrogen evolution co-Catalysts (HER) co-catalysts due to their low overpotential and high conductivity. However, cost and scarcity have driven the exploration of alternatives such as transition metal phosphides (Ni₂P) and Sulfides (MoS₂). Oxygen evolution co-catalysts (OER) are often the rate-limiting step in water splitting. Co-catalysts such as IrO₂ and RuO₂ are commonly used but are expensive. Emerging options include cobalt-based catalysts (Co₃O₄) and Perovskite Oxides such as LaNiO₃ (Qi et al. 2023). illustrates in Figure 10a how a two-step photoexcitation under a co-catalyst occurs with an aqueous redox mediator enhances water splitting processes, whereas (H. Zhang et al. 2024) demonstrates energy diagram for PEC water splitting reaction in Figure 10b using an n-type semiconductor photoanode and a counter electrode immersed in an electrolyte.

Figure 10:

Redox-enhancing designs for PEC-based water splitting using co-catalyst strategies (a) and carrier dynamics (b). (a) Illustration of a Z-scheme energy diagram with co-catalyst positioning to enhance solar water splitting efficiency. Reproduced from (Qi et al. 2022), nature communications, 13(1), 484 (2022), nature publishing group, under the terms of the creative commons attribution 4.0 international license (CC BY 4.0). (b) Photoelectrochemical (PEC) water splitting involves three main steps: (i) Light absorption to excite electrons, (ii) charge transport through the semiconductor, and (iii) surface redox reactions via charge injection. Adapted with permission from (H. Zhang et al. 2024), sustainable energy and fuels (2024), Royal Society of chemistry. (© 2024 Royal Society of chemistry. licensed for reuse in reviews in chemical engineering).

5.2 Heterojunctions in charge separation

Heterojunctions (Wang et al. 2024a) are interfaces between different semiconductors, designed to separate charges by creating an internal electric field. One common configuration is the Type-II Heterojunction, which aligns the conduction and valence bands of two semiconductors to enhance charge separation, such as TiO₂ coupled with CdS. Another type is the Z-Scheme System, which mimics natural photosynthesis by using two semiconductors with complementary band structures, where electrons and holes from the two photocatalysts recombine, leaving highly energetic charge carriers for redox reactions, as seen in systems combining BiVO₄ and g-C₃N₄, which exhibit superior charge separation and redox capabilities.

Despite significant advancements, several challenges remain in achieving efficient and scalable water splitting. These include charge recombination, which reduces efficiency, and can be addressed by using co-catalysts to capture charge carriers and designing nanostructures to shorten charge carrier pathways. Catalyst stability is another challenge, as photocatalysts often degrade due to photocorrosion or oxidation, but can be addressed with protective coatings and stable co-catalysts. Additionally, both the hydrogen evolution reaction (HER) and oxygen evolution reaction (OER) require high overpotentials, increasing energy consumption, so developing catalysts with lower overpotentials is crucial. Finally, optimizing surface properties to enhance water molecule adsorption and reaction rates is a vital area of focus.

5.3 Light absorption and charge dynamics

Photocatalytic water splitting begins with photon absorption and exciton generation. Semiconductors with band gaps in the ∼1.8–3.0 eV range (BiVO₄, g-C₃N₄, CdS) efficiently harvest visible light and have band edge positions straddling the H⁺/H₂ and O₂/H₂O redox potentials shown in Figure 11a Upon band-gap excitation, electrons in the conduction band (CB) and holes in the valence band (VB) must be rapidly separated and transported to reactive sites, or else recombination will squander the absorbed energy. Material nanostructuring and heterojunction design are key to promoting charge separation: for instance, Type-II semiconductor junctions (TiO₂/CdS) and Z-scheme architectures (BiVO₄/g-C₃N₄) create internal fields or stepwise excitation pathways that funnel electrons and holes to different materials. Efficient charge dynamics are evidenced by longer charge-carrier lifetimes and suppressed recombination rates, often achieved by incorporating co-catalysts or surface defect states that trap carriers and mediate surface reactions. In summary, an optimal photocatalyst balances strong light absorption with effective charge separation, ensuring that photogenerated electrons and holes reach the surface with sufficient energy to drive the hydrogen evolution reaction (HER) and oxygen evolution reaction (OER).

Figure 11:

Energy band alignment and reaction pathway schematics for photocatalytic water splitting. (a) Schematic representation of energy band alignment in a photocatalyst system. The conduction band (CB) and valence band (VB) positions are aligned relative to the hydrogen and oxygen redox potentials. Co-catalysts and heterojunctions facilitate directional charge separation, improving photocatalytic efficiency by reducing recombination and enhancing electron–hole migration across interfaces. (b) Schematic free energy diagram illustrating the hydrogen evolution reaction (HER) and oxygen evolution reaction (OER) pathways on a catalyst surface. The energy barriers for intermediate steps (such as H^*, OH^*, OOH^*) are depicted, highlighting thermodynamic requirements for efficient water splitting in photocatalytic and electrocatalytic systems.

5.4 Surface thermodynamics and slab modeling

Once charge carriers reach the surface, water splitting proceeds via adsorbate-driven reaction steps whose energetics can be elucidated by density functional theory (DFT) slab models. Key intermediates include adsorbed hydrogen (H^*) for HER and adsorbed oxygen species (O^*, OH^*) for OER, as depicted in Figure 11b. The ideal catalyst surface binds intermediates neither too weakly nor too strongly (Sabatier principle), which for HER translates to a hydrogen adsorption free energy ΔGH^* ≈ 0 eV as the optimum. DFT calculations on Ni₃Mo alloys, for example, reveal how surface orientation and composition influence these thermodynamics.

In alkaline HER (where the initial Volmer step involves water dissociation), the Ni₃Mo(101) facet exhibits a low water dissociation barrier (∼0.50 eV) and a near-optimal hydrogen binding energy (ΔGH^* ≈ −0.18 eV, close to that of Pt), outperforming other Ni₃Mo facets in predicted HER activity (Pham et al. 2021). These slab-model insights indicate that the Ni₃Mo(101) surface provides an ideal dual functionality: Mo sites facilitate H₂O activation while Ni sites optimally bind the resulting H^*.

More generally, ab initio thermodynamic analyses map out free-energy profiles for each elementary step (H₂O dissociation, H^* adsorption/desorption, OH^* formation for OER), allowing identification of the rate-limiting steps and energetics. Such calculations have shown that OER typically faces higher intrinsic barriers than HER due to multi-electron transfer requirements, rationalizing why OER co-catalysts (IrO₂, Co–Pi) are often needed. By constructing surface Pourbaix diagrams and calculating activation energies on different facets or defect sites, researchers can predict surface reactivity trends and guide catalyst optimization prior to experimentation.

5.5 Surface-specific modeling of reaction energetics

Building on the insights from slab modeling, surface-specific approaches increasingly address the limitations of uniform facet assumptions.

While this study focuses on hybrid DFT and QML methods for photocatalyst discovery, we acknowledge that site-dependent reaction energetics, particularly those arising from surface terminations, defects, and vacancies, play a critical role in realistic water-splitting catalysis.

In high-throughput screening, this challenge is addressed via two complementary approaches: (i) surface-specific DFT datasets, such as the Open Catalyst Project (OC20/OC22), which explicitly model adsorption energies on diverse facet geometries and defect sites, and (ii) transfer learning techniques, where models pre-trained on bulk crystal or slab datasets are fine-tuned using adsorption energy data. Graph neural networks (GNNs), such as DimeNet++ and GemNet, have been adapted to capture local coordination environments, enabling the prediction of adsorption energies with site sensitivity.

Furthermore, recent developments in QML include active site attention mechanisms that allow models to focus on catalytically relevant atomic sites. Integrating such models into the screening pipeline ensures surface specificity without needing prohibitively expensive full DFT treatment for every site.

These models not only enhance the physical realism of catalyst screening but also serve as foundational tools for the DFT–ML workflows discussed next.

5.6 DFT–ML workflows for adsorption and energetics

Increasingly, high-throughput ab initio calculations are being coupled with machine learning (ML) to accelerate catalyst discovery as represented in Figure 12a. In this approach, DFT computations provide a database of structures, adsorption energies, and reaction barriers, which train ML models to predict energetics for unexplored materials. For instance, Chun et al. (2021) integrated a DFT-derived dataset with ML to screen ternary alloy electrocatalysts beyond the reach of brute-force DFT alone.

Figure 12:

Illustration of machine learning strategies (top) and theoretical-experimental alloy design workflows (bottom) in catalysis. (a) Schematic diagram of ternary alloy configuration search, theoretical predictions, and experimental validation for the oxygen reduction reaction (ORR). Reproduced from (Chun et al. 2021), chem catalysis 1(4), 855–869 (2021), Elsevier, under the terms of the creative commons attribution 4.0 international license (CC BY 4.0). (b) An illustrative overview of computational paradigms for machine learning, including data-driven modeling and feature extraction strategies in materials science. Reproduced from (Li et al. 2022b), npj computational materials, 8, 127 (2022), under the terms of the creative commons attribution 4.0 international license (CC BY 4.0).

Their data-driven search identified an optimal PtFeCu composition, which was experimentally validated to exhibit ∼3-fold higher ORR mass activity than Pt/C (0.67 A mg⁻¹ vs. 0.22 A mg⁻¹) with superior durability. Although focused on the oxygen reduction reaction, this study demonstrates the power of ML-guided exploration of large compositional spaces relevant to water-splitting co-catalysts and electrodes. In the context of HER, Abraham et al. (2023) developed a multi-step ML workflow to predict hydrogen adsorption free energies on 2D MXene catalysts. They calculated ΔGH^* for 1,125 candidate MXenes via DFT and trained a gradient boosting regressor to rapidly predict ΔGH^* for thousands of others, achieving a low mean absolute error of ∼ 0.36 eV. This data-driven model pinpointed trends in composition: H binding on MXenes was most favorable when H adsorbed atop an outer metal site on O-terminated MXenes containing Nb, Mo, or Cr, with several predicted candidates showing ΔGH^* near 0 eV (even surpassing Pt’s activity).

Likewise, single-atom catalyst (SAC) screening has benefited from DFT–ML synergies. Jyothirmai et al. (2023) e explored a series of metal and non-metal dopants on a g-C₃N₄ support, using DFT to evaluate H adsorption and then training ML models to generalize across the periodic table. Notably, they identified B, Mn, and Co single-atom dopants on g-C₃N₄ as top performers for HER, and among various algorithms tested, support vector regression yielded the best predictive accuracy in that study.

These examples highlight how DFT–ML workflows can rapidly screen vast design spaces (material compositions, surface facets, and ligands), deliver quantitative energetics (adsorption free energies within a few tenths of an eV of DFT values), and discover novel catalyst candidates with desirable thermodynamics. Such workflows typically involve feature engineering (to capture chemical descriptors like electronegativities, d-band centers), model training/validation (often requiring hyperparameter optimization and cross-validation for reliability), and uncertainty quantification to flag predictions for confirmatory DFT or experimental tests.

5.7 Integrated mechanistic perspectives

A comprehensive mechanism of photocatalytic water splitting emerges by integrating the above aspects, light absorption, charge carrier dynamics, surface reaction thermodynamics, and data-driven catalyst optimization, into a unified framework. In practical systems, these factors are interdependent: the semiconductor’s band alignment dictates the driving force for surface reactions, while the catalyst’s surface kinetics (a low ΔGH^* or OER overpotential) dictates the required charge carrier energy and density.

Advanced studies increasingly emphasize this interplay. For example, co-catalyst nanoparticles (Pt, NiO_x) are chosen not only for favorable HER/OER energetics but also for their ability to siphon electrons or holes efficiently from the light absorber, bridging the gap between charge dynamics and surface catalysis. Likewise, facet engineering in photocatalysts (exposing high-activity crystal faces) must go hand-in-hand with enhancing photon absorption in those regions (via light trapping or plasmonic enhancement) to truly leverage the facet’s catalytic potential.

From a modeling perspective, multi-scale simulations are being developed that couple photogenerated carrier transport models with surface reaction kinetics, providing an integrated picture of how a given material’s optoelectronic properties and surface chemistry collectively determine overall hydrogen output. This holistic view is crucial for guiding next-generation designs: it suggests, for instance, that simply having an ideal band gap is insufficient if surface reaction barriers remain high, and vice versa. Researchers are therefore pursuing synergistic improvements, tuning electronic structure (through doping and, phase junctions) and catalytic interface (through atomically dispersed active sites and, strain engineering) concurrently.

Future high-impact directions include dual-function materials that simultaneously optimize light harvesting and catalysis (plasmonic photocatalysts that generate “hot” carriers to lower reaction barriers) and closed-loop discovery platforms where experimental feedback refines DFT–ML models to account for real-world factors like kinetics and stability. In summary, an integrated mechanistic understanding, spanning photon to molecule, is being established as the foundation for breakthroughs in photocatalytic hydrogen production.

5.8 Advanced mechanisms and emerging trends

Plasmonic nanoparticles, such as gold (Au) and silver (Ag) (Lee and El-Sayed 2006; Link et al. 1999), enhance light absorption through localized surface plasmon resonance (LSPR) (Sarina et al. 2013). This generates hot electrons that participate in redox reactions. Plasmonic photocatalysis is particularly promising for NIR-responsive systems (Lee et al. 2024). Catalysts capable of performing both HER and OER on the same surface are being developed to simplify reaction systems and reduce costs. For instance, MoS₂ modified with Ni(OH)₂ has shown dual functionality (Subramanian et al. 2018).

Photothermal catalysis utilizes localized heating effects to accelerate reaction kinetics. This approach is often combined with plasmonic systems to enhance efficiency. To address current issues and make the technology commercially viable, research should concentrate on incorporating AI and machine learning to use predictive algorithms for finding ideal catalyst designs and reaction conditions, exploring abundant materials and scalable production methods to develop cost-effective catalysts, and designing multi-photon systems that can enhance light absorption.

These emerging strategies are reflected in recent experimental advances that demonstrate the practical outcomes of optimized photocatalytic systems. Table 4 provides a comparative summary of recent experimental efforts in binary photocatalytic systems for hydrogen evolution, highlighting variations in H₂ yield, quantum efficiency (QE), and light source. These results underscore the role of material composition, heterojunction design, and co-catalyst integration in enhancing water-splitting performance.

For instance, PtCu/TiO₂ and COF–Pt architectures show superior yield under visible light, while Al:SrTiO₃-based systems excel under UV. This tabulated evidence reinforces the mechanistic discussions in Sections 5.1–5.3 by contextualizing how band structure engineering, charge separation strategies, and light harvesting directly translate to experimental hydrogen production efficiency.

6.1 Classical machine learning in photocatalyst design

Classical (non-quantum) ML algorithms have become integral to accelerating photocatalyst discovery. Regression models predict continuous material properties critical to photocatalysis (Abraham et al. 2024; Arabaci et al. 2025; Wang et al. 2025). For example, linear and polynomial regression can provide baseline fits for how a dopant concentration or particle size affects the hydrogen evolution rate. More powerfully, non-linear regressors like Support Vector Regression (SVR) and Gaussian Process Regression (GPR) capture complex relationships such as how co-catalyst addition influences the band gap or surface charge transfer efficiency (Kumar and Singh 2021).

GPR, in particular, is valued for providing uncertainty estimates along with predictions, flagging which predicted photocatalyst performance values are less certain and may need experimental validation. Recent advances in classical machine learning (ML) have significantly enhanced photocatalytic materials discovery and optimization. Regression models predict key continuous properties such as bandgaps, reaction rates, and absorption coefficients. Linear regression provides baseline insights, while Support Vector Regression (SVR) is effective for capturing non-linear dopant-bandgap relationships (Awad and Khanna 2015). Gaussian Process Regression (GPR) enables uncertainty-aware predictions for unexplored material spaces (Deringer et al. 2021). Classification methods such as Random Forest (RF) (Breiman 2001; Rigatti 2017) and K-Nearest Neighbors (KNN) (Al-Dahidi et al. 2024; Javed et al. 2024; Peterson 2009) support activity screening, while Decision Trees aid interpretability (Kowsari et al. 2019).

Classification algorithms help screen materials by categorizing them (“active” vs “inactive” photocatalysts). Random Forest (RF) classifiers have been used to mine feature importance from large materials datasets, for instance, identifying which elemental or structural descriptors most strongly distinguish highly active photocatalysts. Decision trees and ensemble methods (Gradient Boosted Trees, Extreme Random Trees) produce human-interpretable decision rules, useful for quick heuristic screening. Clustering methods like k-Nearest Neighbors (KNN) can group similar materials and have been applied to discover clusters of semiconductors with comparable optoelectronic properties (Barik and Woods 2024; Lu et al. 2024). In a recent work, a suite of classifiers (KNN, RF, SVM, neural networks) was trained on molecular descriptors of photocatalysts and achieved >87 % accuracy in distinguishing high versus low H₂-generation performers. Such models act as a filtration step to rapidly eliminate unpromising candidates from experimental consideration.

Deep learning methods are increasingly adopted as dataset sizes grow. Feed-forward neural networks (FNNs) and deeper multilayer perceptrons have been trained to predict photocatalytic efficiencies from rich feature sets (DFT-computed attributes, synthesis conditions), often outperforming simpler models in capture of non-linear effects. Convolutional Neural Networks (CNNs), originally developed for image recognition, have been repurposed to analyze material imaging data (SEM or TEM images of catalysts) and even diffraction or spectral patterns, automatically extracting microstructural features correlated with activity. Recurrent Neural Networks (RNNs) have seen use in analyzing time-series data such as real-time reaction kinetics, helping to model photocatalytic reaction progress under varying illumination or reactant feed conditions. Furthermore, autoencoders (unsupervised deep models) have been applied to reduce the dimensionality of materials data, for example, encoding a large set of candidate material compositions or structures into a smaller set of latent variables, which can then be explored to find novel candidates with similar latent features to known good photocatalysts.

Deep learning models expand ML capabilities. Feedforward Neural Networks (FNNs) predict electronic and photocatalytic properties (Das et al. 2024; Glorot and Bengio 2010; Yu and Cao 2025 ), and Recurrent Neural Networks (RNNs) analyze time-dependent reaction kinetics (Medsker and Jain 2001; Mienye et al. 2024). Autoencoders help extract latent structure-property relations, while Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) (Kao et al. 2023; Mihandoost et al. 2024; Pakornchote et al. 2024; Yoon et al. 2024) drive inverse design. GANs have predicted novel materials with optimized bandgaps from high-throughput DFT datasets (Hu et al. 2022), and VAEs have suggested stable configurations for defect-tolerant structures.

Generative models are pushing the frontier of materials discovery by creating new hypothetical materials with targeted properties. Generative Adversarial Networks (GANs) have been trained on databases of semiconductor structures and their band gaps so that the generator network can propose new crystal structures expected to have optimal band gaps or stability for water splitting (Li et al. 2021). In one case, a GAN learned from thousands of DFT-evaluated compounds and was able to output candidate semiconductors with predicted band gaps in the 2.0–2.5 eV range (ideal for visible-light absorption) that were not present in the training set. Variational Autoencoders (VAEs) similarly have been used to explore chemical compositional space: by sampling the latent space, researchers generated novel alloy compositions predicted to be stable and photo-active, guiding experimental synthesis efforts. A recent study by Hu et al. (2022) combined high-throughput DFT with a GAN to successfully propose oxide photocatalysts with improved visible-light absorption (Kumar and Singh 2021).These generative models, while still nascent in application, hold promise for inversely designing photocatalysts by specifying desired performance metrics and letting the model “imagine” candidates to meet them.

Reinforcement learning (RL) techniques are being explored to optimize photocatalytic processes and synthesis in a closed-loop manner. In RL, an agent algorithm iteratively adjusts variables (actions) in response to feedback (rewards). For instance, researchers have implemented Q-learning and policy-gradient algorithms to dynamically tune reaction conditions (light intensity, pH, catalyst loading) to maximize hydrogen output. In one demonstration, an RL agent controlling a simulated photocatalytic reactor learned to maximize the hydrogen evolution rate by favoring a lower pH and higher light flux, balancing the trade-offs between reaction kinetics and catalyst stability.

Robotic experimentation platforms have also started using RL to guide the synthesis of photocatalysts: the agent proposes synthesis parameters (calcination temperature, dopant concentration), the experiment is performed, and the measured activity is fed back to the agent to inform the next suggestion. This closed-loop approach was shown to discover a better-performing co-catalyst deposition method for TiO₂ in fewer iterations than a human-guided trial-and-error search (Haghshenas et al. 2024; Li et al. 2025a). While still emerging, RL offers a path toward self-driving labs for photocatalyst optimization. Reinforcement Learning (RL) is emerging for dynamic photocatalytic process optimization. RL agents adapt light intensity, catalyst concentration, and temperature for optimal H₂ output. Q-Learning has been applied to hydrogen yield maximization (Watkins and Dayan 1992; Zhou et al. 2024 ), while Deep Deterministic Policy Gradient (DDPG) can tune continuous variables in multi-parameter systems (Xie et al. 2024).

Case studies and model interpretability: A growing emphasis in ML-assisted materials research is on interpreting the models to extract human-understandable insights. Black-box models can sometimes identify correlations without offering reasoning, so tools like SHapley Additive exPlanations (SHAP) and feature importance analysis are applied to decipher what the models have learned. For example, Kumar and Singh (2021) developed an interpretable ML model to screen 3,099 candidate 2D photocatalysts (metals with different ligand environments). By using SHAP values, they found that descriptors capturing differences in chemical hardness and electronegativity between the metal and ligands were among the most influential for predicting stability.

This feature attribution aligned with chemical intuition: highly stable compounds followed the Hard–Soft Acid–Base principle, where hard metal cations bond strongly with hard anions and vice versa. The SHAP analysis in their study highlighted, for instance, that a larger difference in electronegativity ( Δ χ ‾ ) between metal and ligand tends to lower the formation energy (indicating a stable, exothermic compound). Such insights guided them to identify specific materials like monolayer HfSe₂ and ZrSe₂ as top candidates, which were predicted to have solar-to-hydrogen (STH) conversion efficiencies near their theoretical limits.

These predictions provide concrete targets for experimental synthesis and testing under real solar illumination. Indeed, bridging computation to experiment is a key validation step: an AI-driven screening is only truly successful when a predicted material is fabricated and demonstrates the anticipated performance in actual water-splitting tests. While many ML predictions await experimental verification, a few successes are reported; for example, Fe-doped graphitic carbon nitride (Fe/g-C₃N₄) was recently optimized via an ML model and then synthesized, achieving improved H₂ evolution under sunlight consistent with the model’s forecast. Showcasing such end-to-end validations (prediction → fabrication → testing) will be crucial to build confidence in AI-designed photocatalysts.

Despite their promise, classical ML models face several challenges in photocatalysis applications. One major issue is the “small data” problem, high-quality experimental data are often limited due to the cost and time of measurements. Traditional ML algorithms typically excel with big data; with small datasets they risk overfitting. Researchers are tackling this by incorporating domain knowledge and advanced strategies: transfer learning (pre-training models on large related datasets like general inorganic compounds, then fine-tuning on specific photocatalyst data) and active learning (iteratively picking the most informative new experiments to perform, based on model uncertainty) have shown success in effectively using limited data. Another challenge is the integration of physical constraints.

Purely data-driven models might violate physical laws (predicting an impossible band structure); to prevent this, new physics-informed ML approaches embed constraints or known equations (like mass balance, thermodynamic limits) into the learning process. Model interpretability is also crucial: catalyst scientists are understandably hesitant to trust a black-box prediction without understanding the rationale. Interpretable models like decision trees or simpler linear models with physically meaningful descriptors are sometimes favored, or post-hoc explanation tools (like SHAP and LIME) are applied to complex models as illustrated above.

Finally, extrapolative predictions (predicting far outside the training domain) remain risky; an ML model trained on, say, oxide semiconductors may not reliably generalize to a novel metal-organic photocatalyst. This necessitates caution and often a human expert in the loop to assess whether a model’s prediction for an unprecedented material is sensible. In summary, classical ML has become a powerful complement to theoretical and experimental techniques in photocatalysis, especially when coupled with rigorous data handling and interpretability to ensure that the insights derived are physically meaningful and practically actionable.

The integration of experimental insights with machine learning (ML) architectures offers a powerful framework for materials discovery. Table 4, originally presented in the water splitting section, complements this discussion by providing empirical benchmarks, such as H₂ yield and quantum efficiency, that serve as target variables for ML training. These metrics can be used in supervised regression and classification models to guide photocatalyst selection and optimization.

Additionally, Table 5, provides a detailed comparison between theoretically predicted and experimentally measured values for critical descriptors, band gaps, overpotentials, and H₂ evolution rates, highlighting the discrepancies and uncertainties inherent in simulations. These gaps underscore the importance of coupling DFT predictions with experimental validation, which can be further bridged by ML models trained on such hybrid datasets. When used alongside the DFT–AI strategies outlined in Table 6, this tripartite integration (experiment–theory–AI) forms a closed-loop system for accelerating photocatalyst discovery. It enables not only predictive modeling but also uncertainty quantification and explainable recommendations, advancing interpretable, data-driven materials science.

6.2 Quantum machine learning (QML) in photocatalysis

Quantum machine learning combines principles of quantum computing with ML algorithms, offering potential speed-ups and new capabilities for materials design (Biamonte et al. 2017; Cong et al. 2019; Huang et al. 2021; Jayan and Babu 2024; Kao et al. 2023; Steane 1998; Vedavyasa and Kumar 2024). In the context of photocatalytic water splitting, QML is an exciting emerging area aimed at improving the accuracy of simulations and the efficiency of exploring candidate materials. The appeal of QML lies in the unique features of quantum computing: qubits can exist in superposition and become entangled, enabling certain computations to scale more efficiently than on classical bits. For materials problems, this means a quantum machine learning model might handle the quantum-mechanical complexity of catalyst behavior (such as electron correlation or tunneling effects) more naturally than a classical model.

One direction of QML research is using quantum computers to accelerate quantum chemistry calculations. Quantum algorithms (like the Variational Quantum Eigensolver) can potentially find ground-state energies or reaction barriers faster than classical quantum chemistry, which could allow rapid evaluation of photocatalyst properties within an ML loop Cerezo et al. (2021). For example, a quantum neural network or quantum kernel model could be trained to predict a material’s band gap or exciton binding energy by implicitly processing quantum states of that material.

Quantum Support Vector Machines (QSVMs) use quantum-enhanced feature spaces to perform classification of materials: initial demonstrations have shown that QSVMs can categorize crystal structures or material phases with accuracy comparable to classical SVMs while using fewer training samples. In one benchmark study, a QSVM was applied to classify transition-metal compounds and achieved ∼99 % test accuracy, essentially matching the 100 % accuracy of a classical ML model on the same task. This indicates that current QML models are at least competitive with classical approaches on small datasets, though not yet vastly superior in raw performance. The theoretical advantage of QML is better seen in certain contrived problems where quantum algorithms show exponential speedup in learning efficiency.

For instance, Huang et al. (2022) proved that a quantum learner could determine properties of a physical system with exponentially fewer experiments than a classical learner in specific scenarios. While such quantum advantage has not been explicitly realized for photocatalysis datasets yet, these proofs-of-concept energize the field by suggesting that as quantum hardware improves, we might tackle currently intractable design spaces (exploring an astronomically large chemical space of catalyst compositions) more feasibly with QML.

Several types of QML models are under investigation for materials design Nandy et al. (2022); Rebentrost et al. (2014); Li et al. (2024). Quantum variational circuits can serve as quantum neural networks: they consist of parameterized quantum gate sequences whose parameters are optimized (via a classical optimizer) to minimize a cost function. These have been used to fit potential energy surfaces and could learn the relationship between atomic structure and photocatalytic activity by directly encoding quantum states of reactants and catalysts. Quantum generative models, like quantum GANs or quantum variational autoencoders, have been theorized as well; a quantum GAN might generate candidate molecular structures in superposition, potentially discovering new catalysts with desired properties using fewer parameters than a classical GAN. Early studies outside photocatalysis have shown that a quantum GAN can model molecular data distributions with high fidelity, hinting that it might recommend novel photocatalyst structures more efficiently.

An area where QML could be particularly impactful is dealing with the quantum nature of excited states and charge dynamics in photocatalysis. Processes like exciton formation, charge separation, and surface reactions are fundamentally quantum-mechanical. Classical ML models often rely on descriptors (band gaps, effective masses) computed via DFT, which might miss subtle many-body effects. QML models, by contrast, can incorporate quantum simulation results more directly or even run quantum subroutines to evaluate properties on the fly. For example, one could envision a QML-driven active learning loop where a quantum computer evaluates the excited-state energy of a proposed catalyst (via a quantum subroutine akin to time-dependent DFT), and the ML model uses that information to decide the next candidate to evaluate. In principle, such a setup could navigate the search space of, say, defect-engineered photocatalysts more efficiently by quantum-accelerating the property predictions inside the loop.

It is important to note that currently QML is largely in the research phase, and clear demonstrations of its superiority on real photocatalysis problems are yet to be seen (Cong et al. 2019). In practice, today’s quantum hardware (NISQ devices) is limited by noise and small numbers of qubits, which restricts the size of materials that can be simulated and the depth of circuits that can be run reliably. Many reported QML successes are on simplified or synthetic tasks (like distinguishing very small molecules or idealized datasets). For instance, one study classified 350 materials into two structural classes using QSVM and classical SVM, both achieved ∼100 % on the simplified task after careful feature selection. Such results show QML can reach classical-level accuracy, but they do not yet prove an advantage.

Researchers often find that without error correction, quantum models may even underperform well-tuned classical models on noisy, complex data. Therefore, a realistic near-term view is that hybrid classical-quantum approaches will be most useful: one might use classical ML for data preprocessing and feature engineering and a quantum model for a specific bottleneck (like computing a quantum kernel or a molecular energy) where it excels.

As quantum devices improve, we anticipate more head-to-head comparisons on practical datasets (using quantum kernels for predicting metal oxide band alignments) to rigorously test whether QML offers speed or accuracy gains. Even if outright performance gains are modest initially, QML provides valuable experience in incorporating quantum data into our ML workflows; for example, using quantum-calculated descriptors (from quantum subroutines) as features in classical models has already enhanced accuracy for small molecular datasets.

QML represents a forward-looking synergy between two cutting-edge fields. For photocatalysis, the ultimate vision is a quantum-infused discovery platform: imagine an algorithm that could evaluate a photocatalyst’s electronic structure on a quantum computer, feed that into an ML model that predicts the catalyst’s performance and then use quantum optimization to suggest how to tweak the catalyst (via doping or structural modifications) to improve efficiency. Achieving this will require overcoming current hardware constraints and developing new QML algorithms tailored to material science problems. The progress so far, though mostly theoretical or demonstrated on model systems, is encouraging. As summarized in Table 7, QML offers theoretical advantages in data efficiency, quantum feature representation, and generative modeling capabilities, even though classical ML remains dominant in current practical applications. Continued advances in quantum computing could make QML a viable component of photocatalyst design pipelines in the coming decade.

High-throughput material screening is computationally expensive when using classical DFT calculations. QML can reduce the computational overhead by learning quantum behavior patterns from small datasets and extrapolating them to predict properties of unexplored materials, as exploited by Ajagekar and You (2023) in Figure 13.

Figure 13:

The energy-based model utilizes samples from a quantum annealer to map (a) molecular structures to their properties via a (b) GraphConv neural network. (c) The system estimates free energy landscapes and iteratively optimizes candidate molecules with targeted properties using quantum annealing. Reproduced from (Ajagekar and You 2023), npj computational materials, 9, 143 (2023), under the terms of the creative commons attribution 4.0 international license (CC BY 4.0).

6.3 Integration of DFT, AIMD, and ML for photocatalysis

A powerful trend in recent research is the integration of first-principles simulations (like DFT and AIMD) with AI/ML methods to create a virtuous cycle of material optimization. DFT has long been the workhorse for computing properties such as band structures, surface energies, and reaction barriers for photocatalysts. However, DFT calculations can be too slow to screen thousands of candidates or simulate large interface models. ML can act as a surrogate to DFT, learning from a limited set of DFT results and then predicting properties for uncomputed systems almost instantly.

This DFT→ML surrogate approach was used by Huo et al. (2024) to predict band gaps for over 1,200 semiconductor compositions: an SVR model trained on high-throughput DFT results achieved a root-mean-square error of ∼0.36 eV, rapidly pinpointing new compositions in the desired 2–2.5 eV band gap range. Similarly, ML models have been trained to predict formation energies and stability of materials, allowing quick screening of which hypothetical compounds are likely to be synthesizable and robust. An example is the work by Elbaz and Caspary Toroker (2024) on spinel oxides, where a combination of kernel ridge regression and random forest predicted formation energies to within 0.02 eV of DFT values, flagging the most stable compounds for experimental focus.

Beyond static properties, ab initio molecular dynamics (AIMD) simulations provide insight into dynamic processes such as surface reactions, carrier transport, and catalyst stability at finite temperatures. AIMD can capture phenomena like water molecule adsorption and splitting on a catalyst surface or thermal disorder effects on a catalyst’s crystal structure (Cheng and Sprik 2010; Shao et al. 2019). However, AIMD is notoriously expensive, typically limited to <1,000 atoms and <100 ps with DFT accuracy. Here, ML comes to the rescue via machine-learning force fields (MLFF) or interatomic potentials.

By training on a database of DFT-calculated forces and energies (often from many short AIMD trajectories or random structure perturbations), MLFFs (such as neural networks or Gaussian approximation potentials) can emulate DFT accuracy at a tiny fraction of the cost. This enables much longer and larger-scale simulations, for instance, allowing a nanosecond-scale simulation of a photocatalyst nanoparticle interacting with water, which would be infeasible with direct AIMD. Recent works have started to develop MLFFs for photocatalytic materials; for example, a universal graph neural network potential was trained on a diverse set of metal oxides and shown to predict surface interaction energies within ∼0.1 eV of DFT (Ping et al. 2015).

Such potentials make it possible to simulate carrier diffusion lengths and surface recombination probabilities by propagating atomic motions over longer times, thereby connecting directly to properties like charge carrier mobility and lifetime that are critical for photocatalyst efficiency (Guo et al. 2018; Spiegelman et al. 2020). One study reported a constrained AIMD simulation of the water oxidation mechanism on a TiO₂ surface (with an organic dye sensitizer), which provided a detailed step-by-step picture of oxygen evolution; the authors noted that extending this to broader time/length scales would require ML-accelerated dynamics due to the prohibitive cost of conventional AIMD (Lee et al. 2007).

Time-dependent DFT (TD-DFT) and many-body techniques (GW/BSE) are also being integrated with ML to tackle excited-state properties like optical absorption spectra and exciton behavior. Photocatalytic performance is governed not only by band gaps but also by exciton binding energies (the energy holding electron-hole pairs together) and charge-carrier effective masses (which influence mobility) (Samanta et al. 2022). Traditional DFT (with semi-local functionals) often misestimates these, so higher-level calculations or experimental data are needed. In a recent large-scale study of organic photocatalysts, TD-DFT was used to compute excitonic energy levels and ionization potentials for hundreds of molecules, which then served as input features for an ML model predicting hydrogen evolution rates (Bonin et al. 2014). The authors specifically calculated the exciton binding energy (difference between electronic and optical band gaps) for each molecule, noting that many had large Eeb that could hinder free charge generation.

By including descriptors like exciton binding energy, singlet–triplet splitting, and dielectric constant, their ML model could recognize which molecules had inherently favorable excited-state dynamics for charge separation. Impressively, the model identified a set of structural motifs (electron-withdrawing groups to reduce exciton binding) that correlated with higher H₂ evolution, guiding chemists to focus on those motifs in new designs. This demonstrates the power of marrying advanced electronic-structure calculations with ML: the former contributes physically rich, relevant features, and the latter finds patterns within those features to predict outcomes.

AIMD + ML methods also help evaluate catalyst stability under realistic conditions, which is crucial for practical deployment. For example, AIMD simulations of a photocatalyst in water can reveal if the material’s surface will leach ions or undergo phase transformations. By analyzing many snapshots from AIMD (or using ML to classify trajectories), one can predict the catalyst’s long-term stability or identify which surface facets are prone to deactivation. In one report, AIMD simulations of a TiO₂–water interface revealed significant discrepancies in predicted band edge positions when using a standard DFT functional, underscoring the need for careful choice of methods or empirical corrections.

Efforts are underway to use reinforcement learning on AIMD trajectories to discover reaction pathways; for instance, an RL agent can bias the simulation to escape shallow metastable states and find the critical transition state for water splitting on a surface, effectively accelerating rare-event sampling. These hybrid physics-AI approaches constitute a form of “self-driving MD”, where ML guides the simulation to efficiently explore important events (surface bond breaking, defect formation) without brute-force time integration.

Looking ahead, the integration of DFT, AIMD, and ML is expected to become even tighter. We foresee autonomous workflows where: (i) An ML model is trained on an initial batch of DFT data (small set of materials or reactions). (ii) The model predicts a host of new candidates (materials or reaction conditions) likely to perform well. (iii) Those candidates are fed into high-fidelity DFT or AIMD calculations for validation. (iv) The new results are added to the training pool to refine the ML model (active learning loop). Such closed-loop systems can iteratively converge on optimal solutions with far fewer total computations than a brute-force search. Early versions of this paradigm have been demonstrated, for example, to identify promising photocatalytic CO₂ reduction materials by cycling between ML predictions and DFT verification. Another future direction is incorporating multi-modal data: experimental characterization (X-ray spectra, microscopy images) could be fused with DFT data in training an ML model, providing a more holistic view of what makes a good photocatalyst. Graph Neural Networks (GNNs), which naturally represent atoms and bonds as graph inputs, are particularly promising for unifying such data and have shown success in learning material properties across chemical and structural diversity.

In conclusion, recent advances illustrate that the synergy of computational physics and machine learning is a powerful accelerator for photocatalyst development. By handling the data deluge from high-throughput computations and experiments, ML extracts actionable knowledge (trends, key descriptors, candidate suggestions) that can be fed back into theory and synthesis. At the same time, anchoring ML models in physical reality through DFT/AIMD data and interpretability techniques ensures that these models do more than just fit the data; they contribute to understanding why certain materials work well.

As the community continues to address challenges like data quality, model generalization, and quantum computing integration, we move closer to a paradigm where discovering a new water-splitting catalyst is less about serendipity and more about design, driven by algorithms that tirelessly learn from every piece of data available. This approach is paving the way toward photocatalysts with the efficiency and stability needed for sustainable hydrogen production, discovered on accelerated timelines unimaginable just a decade ago.

7.1 Challenges in DFT–artificial intelligence integration

DFT calculations and experimental data often suffer from inconsistencies, making it difficult to train reliable AI models. Different research groups use varying exchange-correlation functionals, basis sets, and pseudopotentials, leading to non-uniform datasets. Inconsistencies in measurement techniques and environmental conditions can further complicate AI model training. Establish standardized protocols for DFT calculations and experimental validations. Initiatives like the Materials Project and Open Catalyst Project are paving the way by creating large, standardized databases.

While DFT is computationally efficient compared to other quantum mechanical methods, it remains resource-intensive for large-scale or high-throughput studies. Integrating AI models with DFT exacerbates this issue when training on vast datasets or performing iterative calculations. Use surrogate models like Gaussian Process Regression (GPR) or Neural Networks (NNs) to approximate DFT outputs for rapid predictions. Develop hybrid frameworks that combine low-fidelity models with high-fidelity DFT calculations to balance accuracy and efficiency.

AI models, particularly deep learning architectures, often function as black boxes, making it challenging to interpret their predictions and validate results against physical principles. Implement Explainable AI (XAI) techniques, such as SHapley Additive exPlanations (SHAP) or Local Interpretable Model-Agnostic Explanations (LIME), to identify prime features driving predictions. Use physics-informed neural networks (PINNs) to incorporate known physical laws directly into AI models.

AI models trained on specific material systems often struggle to generalize to new compositions or structures, limiting their utility in discovering novel photocatalysts. Employ transfer learning, where pre-trained AI models are fine-tuned for specific material classes with limited data. Use unsupervised learning methods to cluster similar materials and identify transferable patterns.

7.2 Future perspectives in DFT–artificial intelligence integration

Integrating quantum mechanics into AI models represents a significant frontier in improving the accuracy and scalability of DFT-AI systems. Quantum Machine Learning (QML) leverages quantum computing to accelerate the training and inference of AI models, particularly for DFT-like simulations, by processing complex quantum states more efficiently than classical methods. Quantum Neural Networks (QNNs) combine the strengths of quantum systems and AI algorithms, offering powerful tools to solve high-dimensional problems in material science, such as the prediction of electronic structures and reaction dynamics. These quantum-inspired approaches promise to reduce computational costs and unlock novel pathways for designing advanced photocatalysts.

Automated AI-driven workflows are also poised to revolutionize material discovery by enabling inverse design and active learning frameworks. Inverse design uses AI to propose material structures or compositions with desired properties, which are then validated through DFT calculations. Active learning further enhances efficiency by iteratively selecting the most informative data points, optimizing computational and experimental resources. Multi-fidelity modeling provides additional scalability by combining low-cost approximations, such as semi-empirical methods, with high-accuracy DFT calculations. This hybrid approach allows researchers to generate large training datasets for AI models while maintaining precision, bridging the gap between speed and computational intensity.

Finally, expanding data repositories and integrating multi-modal data are essential for advancing DFT-AI integration. Initiatives like the Materials Project, JARVIS-DFT, and Open Catalyst Project provide standardized datasets, but future repositories must also incorporate multi-modal data, such as electronic structures, optical properties, and experimental synthesis conditions. Graph Neural Networks (GNNs) offer a promising solution for representing atomic structures as graphs, enabling efficient learning of material-property relationships. Multi-task learning can further enhance material screening by simultaneously predicting multiple properties, such as bandgap, stability, and defect tolerance. These advancements will facilitate a more comprehensive understanding of photocatalyst performance, accelerating innovation in sustainable energy solutions.