The efficiency of plastic waste-to-hydrogen conversion depends on several interrelated parameters, including reaction temperature, catalyst composition, feedstock characteristics, residence time, and process atmosphere. Appropriate catalyst selection can reduce operating temperatures and improve product selectivity in polymer conversion systems^[66]. Process optimization studies show that increasing reforming temperature from 500 to 1,000 °C can enhance hydrogen yield by up to 83.53%, although improvements beyond ~700 °C become marginal. Elevated pressures generally reduce hydrogen and CO yields, while an optimal steam-to-plastic ratio (~2) improves conversion efficiency^[67].

Despite these advances, several challenges limit industrial deployment. Thermochemical processes often generate tar and heavy hydrocarbons that cause equipment fouling and catalyst deactivation, particularly for oxygen-rich plastics such as PET^[68]. High reaction temperatures (> 730 °C) increase energy demand and reduce economic feasibility^[68]. In addition, heterogeneous mixed-plastic feedstocks containing additives or contaminants (e.g., chlorine, sulfur, metals) complicate process control, which may lead to harmful emissions and catalyst poisoning^[69]. Economic constraints, including high capital costs and limited waste-processing infrastructure, further restrict large-scale adoption, while environmental compliance requirements add operational complexity^[70].

The choice of catalysts is intimately linked to the specific conversion process: thermocatalysts play a central role in high-temperature processes such as pyrolysis, gasification, and steam reforming. Metal-based and zeolite-supported catalysts (e.g., Ni/Al₂O₃, Fe/SiO₂, Ni-Fe/γ-Al₂O₃) are optimized for efficient polymer breakdown and hydrogen evolution^[75,76]. In contrast, electrocatalysts drive electrochemical upcycling at lower temperatures, enabling selective hydrogen production from plastic-derived intermediates (e.g., ethylene glycol from PET) using materials like Ni-CoP or SACs like Ru^[12,77,78]. Photocatalysts utilize solar energy for ambient-temperature hydrogen production and plastic degradation, with high-entropy oxides and two-dimensional transition metal carbides/nitrides (Mxene)-based composites showing prospects for simultaneous waste valorization and clean fuel generation^[24,79]. A schematic summary of application-specific catalysts across different conversion pathways is shown in Figure 5A.

Catalyst composition fundamentally governs the reaction pathway and determines the resulting product distribution. Monometallic catalysts such as Ni, Fe, and Co, often supported on oxides like SiO₂ or Al₂O₃, are widely used for thermochemical processes^[80]. The support not only disperses the active phase but also influences acid-base balance, redox properties, and carbon deposition resistance. For instance, Fe/SiO₂ and Ni/SiO₂ catalysts with larger particle sizes have been reported to enhance hydrogen yields and improve the quality of carbon nanotubes (CNTs) during gasification of polypropylene^[75]. Bimetallic and multimetallic systems (e.g., Ni-Fe, Co-Fe, Ni-Co) further exploit synergistic effects, leading to higher hydrogen yields and superior carbon nanomaterial production, especially when supported on MgO or zeolites^[81]. The synergistic effects between altered electronic configurations and intermetallic interfaces promote hydrogen release while inhibiting carbon deposition. At the mechanistic level, transition metals such as Ni, Fe, and Co provide active sites for C-C and C-H bond activation through adsorption-dehydrogenation pathways^[82,83]. The generated hydrocarbon radicals and intermediates further undergo β-scission and steam reforming reactions to release H₂, while oxygen vacancies or basic supports suppress coke formation. Composition of catalysts is also illustrated in Figure 5B.

The dimensionality of catalysts ranging from 0D nanoparticles to 3D hierarchical structures directly influences active site exposure, electron transport, and resistance to deactivation. Zero-dimensional catalysts (nanoparticles, quantum dots, single-atom sites) maximize surface area and active site density, which is advantageous for rapid hydrogen evolution and high-value carbon formation in thermocatalytic and microwave-assisted processes^[75,78]. Two-dimensional materials (nanosheets, layered zeolites, MXenes) offer large, accessible surfaces and improved metal dispersion, enhancing selectivity and charge transfer in hydrocracking, steam reforming, and photocatalysis^[24,84,85]. Three-dimensional architectures (mesoporous supports, composite ferrites, biochar) provide hierarchical porosity and robust frameworks, supporting mass transfer and catalyst stability in high-throughput processes^[76,80]. Hybrid and multidimensional catalysts combine these features, leveraging synergistic effects for optimal performance in complex reaction environments. Figure 5C shows types of catalysts based on dimensional architecture.

Electrocatalysis operates under mild conditions, typically at lower temperatures and ambient pressures, thereby reducing energy input and minimizing byproduct formation. In contrast, thermocatalytic processes generally require elevated temperatures (400-900 °C) and often yield mixed, less-selective products^[97]. Electrocatalysts achieve high Faradaic efficiencies (up to 96%) and enable precise control over reaction pathways, with the potential to enhance hydrogen yield and product purity compared with thermocatalytic and photocatalytic systems^[98]. Moreover, electrocatalytic systems can be powered by renewable electricity, further enhancing their environmental sustainability compared with fossil-fuel-dependent thermal processes. In contrast, photocatalysis, despite being solar-driven, often exhibits low quantum efficiencies and slow reaction kinetics^[12].

The rational design of electrocatalysts (e.g., dual-atom, nanostructured, heteroatom-doped) enables the tuning of activity and stability, thereby enhancing overall performance and scalability^[99]. At a fundamental level, electrocatalysts reduce the overpotential required to drive electrochemical reactions, improving energy efficiency. For instance, in water splitting, efficient catalysts for the HER and oxygen evolution reaction (OER) enable operation closer to thermodynamic potentials, thereby lowering electricity consumption^[100]. In addition, electrocatalysts accelerate reaction kinetics by facilitating interfacial electron transfer at the electrode-electrolyte interface.

Nanostructure engineering further enhances catalytic performance. For example, WS₂ nanosheets with edge-rich structures exhibit significantly improved hydrogen evolution activity at low overpotentials^[101], primarily due to the high density of exposed active sites and defects that promote charge transfer and catalytic turnover^[102]. Beyond activity, electrocatalysts also offer superior selectivity by directing reaction pathways toward desired products while suppressing competing reactions. In CO₂ reduction, for instance, well-designed catalysts can achieve CO selectivity of up to ~95% while maintaining stable current densities^[94]. Despite these advantages, several challenges remain for the practical implementation of electrocatalytic plastic upcycling. Catalyst stability and durability under prolonged operation, together with the cost of advanced catalyst architectures, continue to limit large-scale deployment. In addition, maintaining high selectivity in complex, real-world feed systems—particularly those involving heterogeneous plastic-derived intermediates—remains nontrivial. Importantly, the underlying reaction mechanisms and structure-activity relationships are not yet fully understood, which hinders the rational design and optimization of electrocatalysts. Addressing these challenges will be essential for translating laboratory-scale demonstrations into practical applications.

This section outlines the mechanistic pathway, key reactions, and intermediates involved in electrocatalysis, while highlighting how this approach surpasses other catalytic routes in selectivity, efficiency, and operational conditions. Figure 7 illustrates the reaction mechanism of electrocatalysis for the conversion of plastic to hydrogen.

1. Depolymerization and pretreatment

Plastic waste such as PET is first depolymerized via alkaline hydrolysis into monomers including ethylene glycol (EG) and terephthalic acid. This step is essential for enabling subsequent electrochemical transformations, as the monomers are more readily oxidized than the parent polymer^[103,104]. Equation 1 shows the example reaction derived from the original study of^[105].

(1) $$ \begin{equation} \begin{aligned} PET + 2\mathrm{NaOH} \longrightarrow \mathrm{Na_{2}TPA}~(\mathrm{sodium~terephthalate}) + EG \end{aligned} \end{equation} $$

2. Anodic oxidation of monomers

At the anode, EG undergoes electrooxidation on advanced catalysts (e.g., Au/Ni(OH)₂, dual-doped NiS nanosheets). The process proceeds through several key steps:

• Adsorption: EG adsorbs onto the catalyst surface, facilitated by adjacent hydroxyl groups^[106,107].

• Dehydrogenation and C-C bond cleavage: Sequential oxidation leads to the formation of intermediates such as glycolaldehyde, glyoxal, and ultimately glycolic acid or formate^[108]. Representative reactions are shown in Equations 2 and 3^[109].

• In situ surface evolution. The catalyst undergoes dynamic structural reconstruction which enhances catalytic activity and selectivity^[12].

(2) $$ \begin{equation} \begin{aligned} \mathrm{HOCH_{2}CH_{2}OH} + \mathrm{2OH^{-}} \longrightarrow \mathrm{HOCH_{2}COO^{-}} + \mathrm{2H_{2}O} + \mathrm{2e^{-}}~(\mathrm{to~glycolic~acid}) \end{aligned} \end{equation} $$

(3) $$ \begin{equation} \begin{aligned} \mathrm{HOCH_{2}CH_{2}OH} + \mathrm{4OH^{-}} \longrightarrow \mathrm{2HCOO^{-}} + \mathrm{4H_{2}O + 2e^{-} (to~formate)}\end{aligned} \end{equation} $$

Through surface engineering (such as Au/Ni(OH)₂ and dual-doped NiS), catalysts provide high surface area, abundant active sites, and favorable electronic structures for efficient EG adsorption and oxidation^[108]. Meanwhile, in situ reconstruction during operation leads to the formation of oxyhydroxide or vacancy-rich phases that facilitate C-C bond cleavage and improve product selectivity^[110].

Accurate identification of catalytic active sites is crucial for developing efficient catalysts^[128]. Since Taylor introduced the concept of active sites in 1925^[129], researchers have been dedicated to identifying and characterizing the active sites of catalysts to better understand catalytic reaction mechanisms^[130]. These sites represent specific regions on the catalyst surface where reactions occur and thus directly determine catalytic activity and efficiency. Recently, ML has emerged as a powerful approach for identifying and characterizing catalytic active sites with improved accuracy^[120]. In parallel, C, N, and non-noble metals have gained increasing attention as sustainable alternatives to traditional noble metal catalysts^[131]. ML offers distinct advantages in addressing the complexity of catalyst design and synthesis. It enables the theoretical design of catalyst composition and structure, optimization of synthesis conditions, and automated high-throughput screening^[132]. As a result, ML-driven approaches allow more precise prediction of catalytic performance and significantly accelerate the rational design of advanced electrocatalysts for hydrogen production, thereby streamlining research workflows.

In this domain, recent research^[133] innovatively combined ML with in-situ X-ray absorption spectroscopy to identify the active sites of TMNCs. The results revealed that isolated Ni sites act as the active species for CO₂RR, exhibiting dynamic and heterogeneous behavior that adapts to reaction conditions. While alloy catalysts offer diverse surface-active sites due to their complex composition and atomic arrangement, this structural diversity also complicates the precise identification and optimization of active sites. To address this challenge, ML-assisted simulations have been increasingly applied. For example, neural network potential-based molecular dynamics simulations were used to investigate oxide-derived Cu catalysts for CO₂RR, identifying three square-shaped active sites that are critical for C-C coupling reactions^[134]. Similarly, large-scale analysis of 40,000 active sites on amorphous alloy surfaces enabled the determination of an optimal atomic composition (Pd:Cu:P:Ni = 0.51:0.33:0.09:0.07), with Pd d-electrons identified as the key contributors to catalytic performance. In addition, they employed the SOAP-ML model to reveal the underlying mechanism of catalyst durability, linking it to Ni dealloying behavior. The close consistency with experimental observations demonstrates the accuracy and reliability of the model and highlights its potential for guiding the design of advanced amorphous alloy catalysts^[135].

In addition to catalyst structure prediction, ML has been increasingly applied to predict product distributions and hydrogen yields in thermochemical plastic conversion systems. ML applications in plastic thermochemical conversion focus on predicting aggregate outputs such as oil and gas yields, activation energies, or monoaromatic production rather than explicitly resolving elementary catalytic reaction networks^[136,137]. Nevertheless, these approaches effectively capture the complex relationships among feedstock properties, catalyst characteristics, and operating conditions, and can be extended to hydrogen evolution in catalytic plastic pyrolysis. Tree-based ensemble models, particularly RFs, are widely used due to their ability to handle nonlinear systems with limited mechanistic information. For example, RF models have been used to predict hydrogen yield and process performance indices in plastic-waste gasification using variables such as temperature, pressure, and gasifying-agent flow rate, achieving prediction accuracies above R² > 0.99 and revealing strong nonlinear effects of temperature and steam-to-plastic ratios on hydrogen production efficiency^[37]. Similar RF-based approaches have been applied to pyrolysis systems, where models predict liquid yields and hydrogen content in pyrolysis oils using feedstock composition and operating conditions as inputs^[138], while feature-importance and partial-dependence analyses reveal how temperature and feed characteristics influence gas yields and hydrogen-containing species^[139]. Although these studies were originally developed for biomass systems, the same RF-based interpretation frameworks are increasingly applied to plastic pyrolysis and co-pyrolysis systems^[140]. Beyond RF, gradient-boosted models such as XGBoost have been employed in catalyst-driven plastic conversion. These models are trained on multi-source datasets to predict oil yields and gasoline-range hydrocarbons from plastic composition and zeolite catalyst properties, enabling inverse catalyst design and optimization of operating conditions^[136]. Ensemble approaches have also been used to predict monoaromatic-rich oil yields, where variables such as temperature, catalyst-to-feed ratio, plastic fraction, and zeolite Si/Al ratio are identified as dominant factors^[137]. In addition, boosted regression tree models have been applied to catalytic HDPE pyrolysis to predict activation energy as a function of conversion and heating rate, effectively learning reduced representations of degradation kinetics^[141]. Across these studies, ML models typically combine feed descriptors (polymer fractions, ultimate analysis, or biomass presence), catalyst descriptors (surface area, Si/Al ratio, catalyst loading), and operating parameters (temperature, residence time, pressure, and heating rate) to predict outputs such as oil yield, gas composition, or hydrogen fraction^[142]. Tree-ensemble methods are particularly attractive because they can approximate high-dimensional catalytic reaction networks without explicitly defining individual reaction pathways while simultaneously providing interpretable insights through feature-importance, partial-dependence, and SHAP analyses^[143]. Hydrogen-focused ML studies in related thermo-catalytic systems further demonstrate the applicability of these approaches. For example, tree-based ensemble models have been used to predict hydrogen fractions in syngas during biomass-plastic co-gasification based on particle size, blend ratio, and temperature^[144], while ensemble models combining RF, LightGBM, and bagging regressors have successfully predicted photocatalytic hydrogen evolution rates from catalyst loading, pH, and oxidant concentration^[145]. Although these studies demonstrate the strong capability of RF models for predicting hydrogen-related outputs in thermochemical systems, most existing applications focus on overall product yields or syngas composition rather than explicitly modeling hydrogen evolution within detailed catalytic plastic pyrolysis reaction networks.

Another important challenge in plastic waste conversion is the heterogeneity of real feedstocks, which typically contain mixtures of polymers, along with additives and contaminants. Such variability can strongly influence hydrogen yield, product distribution, and catalyst stability. Recent studies suggest that machine learning models can address this challenge by incorporating physics-informed feed descriptors, such as elemental composition, polymer mass fractions, and contaminant indicators, rather than relying solely on nominal plastic-type labels. For example, ML models using ultimate analysis descriptors have shown strong predictive capability for pyrolysis product yields across mixed plastic feeds^[142], while hydrogen production studies have used feed composition variables together with operating conditions to predict syngas composition from mixed plastics^[146-148]. Furthermore, catalyst stability can also be incorporated into ML frameworks by including catalyst descriptors (metal type, support properties, acidity, surface area) and by using deactivation indicators such as coke formation or activity loss as prediction targets. Experimental studies demonstrate that different polymers generate distinct carbon deposition behaviour on the same catalyst, highlighting the importance of feed composition in catalyst lifetime prediction^[149]. To build reliable ML models for such heterogeneous systems, multi-fidelity datasets combining experimental measurements with simulation-generated data from process models (Aspen-based pyrolysis or gasification simulations) can be employed to explore wide feed-composition and operating-condition spaces^{[147,150,151]}. Such frameworks can support multi-objective optimization of catalyst design and operating conditions for hydrogen yield, catalyst stability, and environmental performance under realistic mixed-plastic feed conditions.

The selection of appropriate input features (descriptors) plays a decisive role in improving the predictive accuracy of ML models and uncovering the key factors that influence catalytic activity and selectivity^[152]. Effective descriptor engineering and feature selection are essential for high-performance ML models in data-driven catalyst discovery. Recent studies employ physically interpretable descriptors, including adsorption energy distributions and d-band center properties. For instance, the ARSC framework decouples atomic, interaction, synergistic, and coordination effects via a PFESS-based feature selection strategy, enabling efficient catalyst discovery while avoiding over 50,000 DFT calculations^[121]. The descriptors show that a good description of catalyst performance can be achieved by quantifying the intrinsic properties of the catalyst and atoms. These descriptors serve as inputs under various machine learning methods, among which GBR has proven effective in predicting and classifying SAC performance. The ML strategies presented in this study can significantly reduce computation time while providing valuable insights for high-throughput screening of SACs^[153]. Integrating domain-specific feature engineering with robust selection techniques can significantly speed up catalyst development while preserving physical relevance and reducing computational overhead. For plastic-to-hydrogen systems, it is important to identify which catalyst and process descriptors best predict hydrogen evolution from plastic-derived feedstocks. It is equally important to understand the limitations and failure modes of ML models to ensure robust catalyst design and process optimization.

Recent ML-driven catalysis studies consistently identify electronic/structural, textural, and feedstock/process descriptors as dominant predictors of hydrogen yield and selectivity in plastic waste reforming and pyrolysis processes^[154]. Electronic and structural descriptors are particularly influential because they directly govern adsorption strength, reaction energetics, and catalytic behavior^[155,156]. In particular, the adsorption energies of key intermediates (e.g., H*, CO*, and OH*) are widely regarded as central descriptors in both HER and related electrochemical and thermochemical reactions^[157]. However, the d-band center and related mean d-band descriptors strongly correlate with adsorption energies and catalytic activity^[158], often following volcano-type relationships in metals and alloys. Additional features such as charge-transfer energy, e.g- electron occupancy, and metal-oxygen bond length further refine this understanding through their effects on adsorption-energy trends. Oxygen-vacancy concentration and catalyst redox state regulate adsorbate binding, spillover, and activation barriers. Meanwhile, metal dispersion and particle size control the density and nature of active sites, thereby influencing both catalytic performance and deactivation^[159,160]. Among compositional variables, transition metal loading (e.g., Ni, Fe, Co) and bimetallic synergy effects emerge as highly influential descriptors governing C-C bond cleavage efficiency and hydrogen evolution pathways. In addition, textural descriptors such as surface area, pore volume, pore hierarchy, and support acidity/basicity influence active-site accessibility and mass transport, thereby affecting catalyst performance^[161]. Process-level descriptors, including temperature, heating rate, residence time, and steam-to-carbon ratio, often exhibit the highest SHAP feature importance values in interpretable ML frameworks, indicating their dominant role in reaction kinetics and thermodynamic equilibrium^[162]. Furthermore, plastic feedstock composition, particularly polymer chain length, degree of branching, and heteroatom content, strongly influences hydrogen production efficiency, highlighting the need for chemically meaningful polymer descriptors in ML datasets^[163]. Interpretable ML methods such as SHapley Additive exPlanations (SHAP) and partial dependence analysis have revealed that oversimplified descriptors, such as elemental weight percentages alone, fail to capture the complex interplay between catalyst structure and activity. Instead, physically grounded descriptors, including metal-support interaction strength, oxygen vacancy concentration, reducibility index, and acidity distribution, provide superior predictive power and mechanistic interpretability^[26]. These findings highlight the importance of incorporating structure-sensitive and physics-informed descriptors to avoid spurious correlations and enhance model interpretability in catalyst design. To further contextualize the datasets and validation strategies used in ML-driven catalyst design, Table 2 summarizes representative studies, highlighting their dataset sources, machine learning models, validation approaches, and key contributions toward sustainable hydrogen production from plastic waste.

In recent years, the integration of ML and DFT has shown tremendous potential in catalyst research^[121]. DFT allows researchers to explore materials at the atomic scale by revealing detailed insights into their electronic structures, thereby facilitating the discovery of novel material properties and underlying reaction pathways^[169]. It can accurately predict the key catalyst properties including active sites, electronic structure, surface adsorption energies, reaction pathways, etc., providing a solid theoretical foundation for understanding catalyst processes^[170]. The synergy between ML and DFT offers a powerful approach to reducing both experimental and computational costs, particularly for screening complex catalytic systems and large material datasets. This integration not only deepens our understanding of surface interactions and catalytic mechanisms but also accelerates the identification and development of novel catalysts^[120]. A recent research publication^[171] provides an extensive study that detailed the use of ML models for the reconstruction of catalysts for ORR. This study can also be translated to plastic-to-hydrogen workflow. A framework for an integrated catalyst discovery workflow combining ML, DFT, synthesis, characterization, and testing for hydrogen and formate production from plastic-derived intermediates can be developed using this study as a good supporting baseline. Further innovation is exemplified in a study, which presents a holistic framework combining GNNs, reinforcement learning, physics-informed neural networks (PINNs), and generative models (VAEs) to accelerate optimization of nanomaterials for photocatalysis and hydrogen generation, achieving 15%-20% yield improvements and a 40% reduction in experimentation cycles^[172]. Recent advancements demonstrate the transformative role of AI in accelerating catalyst discovery and optimizing hydrogen production. Vipin and Padhan^[33] presented a comprehensive study of ML applications in electrocatalyst design for HER, OER, HOR, and ORR, using models like RFs, SVR, and XGBoost to bridge DFT-derived descriptors with catalytic performance, enabling high-throughput screening beyond traditional trial-and-error methods. Another major leap involves the use of GNNs for catalyst screening. A recent review article outlines how GNN-based frameworks can model catalyst-adsorbate interactions, enable accurate predictions of active sites and reaction mechanisms, and capture nuances in atomic structure and symmetry^[173]. Recent studies highlight the transformative impact of combining ML with DFT tools such as VASP and Materials Studio to accelerate catalyst design and mechanistic modeling^[37,174]. For instance, one study investigated a high-entropy, graphene-supported single-atom catalyst (FeCoNiRu-HESAC) using DFT to compute overpotentials for HER, OER, and CO₂RR. ML was then applied to identify the most predictive electronic descriptors of catalytic performance, leading to the rational design of catalyst compositions with theoretically substantiated stability and reactivity^[174]. In another study, it is demonstrated how DFT-calculated Gibbs free energy of hydrogen adsorption (ΔG_H*) on transition-metaldoped transitionmetal dichalcogenides (TM@MX₂) could be efficiently predicted using ML models such as support vector regression (SVR), achieving an R² value of 0.92. Here, DFT via VASP generated training data, enabled rapid screening of promising HER catalysts without exhaustive QM simulation^[37]. Wang et al.^[175] utilised ML to predict the adsorption free energies of reaction intermediates OH*, O*, and OOH* on transition metal AB alloy catalysts. They identified the adsorption sites using Pymatgen software and calculated intermediate free energies at different sites with DFT, and trained ML models that were subsequently validated against DFT results. The study demonstrated that ML could accurately predict the active sites and reaction intermediate free energy for OER/ORR reactions with much higher efficiency than DFT calculations (> 150,000 times), significantly accelerating the design of bifunctional oxygen electrocatalysts. These workflows combining DFT-level accuracy for training data generation and ML for rapid prediction, significantly reduce the experimental and computational burden^[176]. They also enabled high-throughput candidate screening, and scalable exploration of complex catalyst spaces, making them particularly useful in the development of catalysts tailored for plastic waste-derived hydrogen generation.

Recent studies suggest that machine learning can also be integrated with life cycle assessment (LCA) frameworks to support sustainability-oriented catalyst design. In ML-LCA workflows, supervised learning models such as RFs, ANNs, and gradient boosting can be trained to predict environmental impact indicators derived from life-cycle inventory datasets, including greenhouse gas emissions, energy consumption, and material usage^[177-179]. Such approaches have already been applied in bioenergy and chemical process systems, where ML models predict life-cycle inventory parameters and environmental indicators directly from process variables and feedstock properties, significantly reducing the computational effort associated with traditional LCA calculations^[178,180].

Extending this concept to catalyst discovery, ML models could be trained to simultaneously predict catalytic performance and sustainability-related metrics associated with catalyst synthesis and deployment. Input features may include catalyst composition, synthesis temperature, solvent use, and processing steps, while output targets could include cradle-to-gate carbon footprint, cumulative energy demand, or process mass intensity derived from LCA studies^[181-183]. Integrating ML with LCA therefore provides a pathway toward multi-objective catalyst optimization, enabling the identification of catalyst systems that simultaneously maximize hydrogen production performance while minimizing environmental impacts.