From Experimental Optimization to AI Empowerment: Advanced Strategies of Oxygen Evolution Reaction for Large‐Scale PEMWE

ABSTRACT

Proton exchange membrane water electrolysis (PEMWE) has emerged as a core technology for green hydrogen production owing to its fast dynamic response and efficient compatibility with intermittent renewable energy sources. It is now transitioning from megawatt‐ to gigawatt‐scale applications, yet still faces many practical challenges. The oxygen evolution reaction at the anode is a key breakthrough point for enhancing the overall performance of PEMWE. In recent years, significant progress has been made in this field via traditional experimental optimization, while artificial intelligence has also become a major driver of technological paradigm shifts. This paper first systematically reviews the core challenges of PEMWE's large‐scale application. Then, it delves into the optimization strategies for the anode reaction of PEMWE from three dimensions, establishing a strategy system that covers traditional experimental optimization to artificial intelligence (AI) assisted driving. Finally, future development directions toward scaled‐up application are prospected. This paper aims to bridge traditional experimental optimization methods and AI‐enabled approaches, accelerating the large‐scale application of PEMWE and promoting the high‐quality development of the green hydrogen industry.

This review focuses on the anode oxygen evolution reaction of proton exchange membrane water electrolyzers (PEMWE), systematically summarizing the three major challenges that currently restrict the large‐scale application of PEMWE. It also proposes a strategy system that combines traditional experimental optimization with artificial intelligence empowerment, highlighting the latest strategies for enhancing the performance of PEMWE. This review offers new solutions for the industrial application of PEMWE.

1. Introduction

Under the background of the global energy structure's transition toward clean and low‐carbon, hydrogen energy has become an important support carrier for replacing traditional fossil fuels and promoting the decarbonization process of the energy system [1, 2, 3]. Among the mainstream hydrogen production pathways, water electrolysis, featuring prominent merits including low carbon footprint and superior compatibility with renewable energy, has become a core supporting technology for the green and sustainable development of the hydrogen energy industry [4, 5, 6, 7]. Water electrolysis technologies mainly include alkaline water electrolysis (AWE), anion exchange membrane water electrolysis (AEMWE), and proton exchange membrane water electrolysis (PEMWE) [8]. Although the first two types of electrolyzers have their own technical characteristics, AWE is inherently limited by low electrolytic efficiency, while AEMWE suffers from insufficient operational stability, making them difficult to meet the requirements of large‐scale commercial applications [9, 10]. In contrast, PEMWE, with its low internal resistance, rapid dynamic response, high hydrogen purity, and excellent energy conversion efficiency, has demonstrated unique competitiveness in the field of hydrogen production [11, 12, 13]. This technology can be efficiently coupled with intermittent renewable energy sources such as wind and solar power [14]. It enables second‐level power change response and maintains stable operation under low‐load conditions. In addition, PEMWE adopts a compact stack design, effectively simplifying the construction process of distributed hydrogen production facilities [15]. These advantages fully demonstrate the huge potential of PEMWE in large‐scale commercial applications, making it a research hotspot in the current green hydrogen production technology field.

At present, PEMWE is in a critical development stage of transitioning from megawatt‐level demonstration applications to gigawatt‐level large‐scale mass production [16]. The International Energy Agency (IEA) projects that the annual installed capacity of the PEMWE will surge to 43 GW by 2030. Although it has significant technical advantages, there are still many bottlenecks restricting its industrial‐scale application process. Compared with the cathode hydrogen evolution reaction (HER), the anode oxygen evolution reaction (OER) is an inherently slow four‐electron transfer process, which not only requires a high overpotential but is also limited by slow reaction kinetics [17, 18, 19, 20]. Therefore, focusing on the optimization research of the OER reaction has become a key breakthrough for improving the overall performance of PEMWE. The kinetic limitations of OER impose strict requirements on the performance of catalysts, urgently requiring the catalysts that have both high catalytic activity and long‐term stability in industrial‐level acidic environments [21, 22]. However, the commercial catalysts currently used in industrial applications are mostly based on rare and expensive precious metals such as iridium and ruthenium, which leads to high costs for the industrial application [23, 24]. Although researchers have attempted to control costs by reducing the loading of precious metal catalysts, this has led to the rapid performance degradation of the catalysts under the harsh OER conditions of strong acidity and high potential. In addition, at industrial‐level current densities, a large number of bubbles are generated in a short time during the electrolysis process [25]. The accumulated bubbles will seriously hinder the mass transfer process within the electrolyzer, significantly reducing the energy efficiency of the entire electrolysis system in large‐scale applications [26]. For a long time, the optimization research of the comprehensive performance of PEMWE has mainly relied on repeated experiments to explore the optimal process in a vast sample space. Although this method is accurate and scientific, it has the disadvantages of long research cycles and high experimental uncertainties, and is highly dependent on the prior experience of researchers. With the rapid development of artificial intelligence (AI), its application in fields such as material design and performance prediction has shown irreplaceable advantages. It can efficiently handle high‐dimensional and nonlinear complex data, significantly shorten the research cycle, and reduce research costs, providing a new technical path for breaking through the bottlenecks of PEMWE [26, 27]. At present, most of the research and review literature on PEMWE mainly focuses on innovative strategies to enhance the performance of PEMWE based on traditional experimental methods, with insufficient attention paid to the core challenges of large‐scale application of PEMWE and the enabling role of AI technology in this field. There is a lack of systematic review and summary. Therefore, it is essential to integrate traditional experimental approaches with emerging AI technologies and establish a complementary research framework. This will help address gaps in review research within this field, providing up‐to‐date theoretical support and practical guidance for the industrialization of PEMWE.

This review focuses on the OER reaction of PEMWE and systematically analyzes the three core practical challenges faced in the large‐scale application of PEMWE. On this basis, this paper constructs a full‐dimensional technical strategy system from traditional experimental optimization to AI‐assisted driving. It reviews the latest research progress in improving the comprehensive performance of PEMWE from three dimensions: “Precious‐metal optimization & non‐precious‐metal exploration”, “Mass transfer enhancement” and “Artificial intelligence empowerment”. Finally, this paper points out that although many new strategies for improving the performance of PEMWE have been proposed, there is still room for further optimization and improvement in its large‐scale application. It further clarifies the four key research directions in the future. This review aims to comprehensively summarize the performance optimization strategies for the large‐scale application of PEMWE. It seeks to drive the shift of PEMWE research paradigm from traditional experiment‐driven approaches to artificial intelligence‐assisted design. By combining the advantages of both methods, it promotes the industrial‐scale deployment of PEMWE and the efficient, sustainable development of the green hydrogen industry (Figure 1).

FIGURE 1.

The current challenges of OER reaction for PEMWE and the framework for promoting its large‐scale application.

2. Challenges of Industrial‐Scale PEMWE Application

At present, although PEMWE has entered the initial stage of commercialization and the phase of large‐scale deployment, it has not yet achieved substantial and comprehensive popularization in all fields. In the process of promoting large‐scale application, the core challenges mainly stem from the three pivotal scientific issues: cost constraints, catalyst deactivation, and limitations in the mass transfer process. Therefore, a thorough analysis of the intrinsic mechanisms of these challenges is of crucial strategic significance for precisely formulating the path of technological breakthroughs and accelerating the advancement of electrolyzers toward larger‐scale industrial applications.

2.1. Cost Constraints

One obstacle to establishing gigawatt‐scale PEMWE is its cost. The U.S. Department of Energy has explicitly set a hydrogen target of achieving $1 per kilogram of clean hydrogen by 2030 [29]. Commercial PEMWE currently widely employ platinum‐group metal catalysts such as iridium and ruthenium, which are scarce in reserves, persistently expensive, and subject to severe price fluctuations (Figure 2a) [30]. In commercial electrolyzers, the loading of Ir is typically between 2.0 and 3.0 mg_Ir cm⁻² [31, 32, 33, 34]. The precious metal catalysts used in the OER reaction usually account for about 25% of the total cost of the membrane electrode assembly (MEA) in commercial PEMWE [35, 36]. In addition, the harsh acidic and high‐potential operating environment of the OER reaction further increases the manufacturing, operation, and maintenance costs of the equipment. The performance optimization of traditional PEMWE mainly relies on experiments. Researchers need to conduct extensive explorations in a multi‐dimensional parameter space of vast elemental compositions and ratios. Such traditional experimental methods have problems of long cycles, low efficiency, and high costs, and sometimes it is difficult to precisely synthesize target materials [37, 38, 39]. In some specific experimental scenarios, the operating conditions are extremely harsh, and obtaining data through traditional experimental methods consumes a lot of time and economic costs. Moreover, traditional experiments only rely on point‐by‐point testing and empirical judgment, and are unable to efficiently process massive experimental data, making it difficult to improve the overall research efficiency. Therefore, the high comprehensive costs, including economic and time costs, have created a certain competitive gap in terms of economics between PEMWE and other water electrolysis technologies, hindering its large‐scale application development.

FIGURE 2.

(a) Global supply and market pricing of platinum group metals. Reproduced with permission [28]. Copyright 2017, Elsevier. (b) Schematic diagram of active site dissolution of catalysts. (c) Mechanism diagram of the AEM and LOM reaction pathways. (d) Scheme of catalyst surface passivation. (e) Mass transport illustration in membrane electrode assembly, including water, gas and protons.

2.2. Catalyst Deactivation

Catalysts are crucial for PEMWE. The strong acid and high‐potential environment during the OER process can easily lead to their deactivation [40]. In‐depth analysis of the deactivation mechanism can precisely identify the root cause of performance decline and provide theoretical support for the design of high‐stability catalytic systems. The deactivation of anode catalysts can be summarized as active site dissolution, lattice oxygen participation, and surface passivation.

Catalyst dissolution depends on its intrinsic elemental characteristics (Figure 2b). IrO₂ has a certain stability in acidic media due to the strong Ir‐O bond [41]. However, when the potential exceeds 1.6 V (relative to RHE), the active sites are oxidized to the high‐valent Ir⁶⁺ state and dissolve [42]. RuO₂ is prone to over‐oxidation to soluble RuO₄ in strong acid and high‐potential conditions [43]. When the potential exceeds 1.4 V (relative to RHE), its dissolution rate increases exponentially [44]. Transition metal‐based catalysts such as Co₃O₄ have insufficient stability in acidic conditions. Protonation of lattice oxygen weakens the metal‐oxygen bond, leading to the dissolution of cobalt active sites [45, 46, 47, 48]. Additionally, impurities such as Cl⁻ in the electrolyte can accelerate catalyst dissolution. Only 10 ppm of Cl⁻ can be oxidized to Cl₂ or HClO, increasing the dissolution rate of Ir by 5 times and causing the catalyst layer to delaminate [49].

The long‐term stability of the catalyst is closely related to its reaction path. At present, the mainstream reaction pathways mainly include the adsorbate evolution mechanism (AEM) and the lattice oxygen‐mediated mechanism (LOM) (Figure 2c). In the AEM pathway, the evolution of oxygen atoms relies entirely on the conversion of adsorbed intermediates (*O, *OH, *OOH), and all O─O bond formation strictly originates from liquid‐phase water molecules. However, this pathway is limited by thermodynamic energy barriers. Due to the upward shift of the oxygen atom's p‐band center, the LOM path is less restricted by thermodynamic energy barriers [50, 51]. However, lattice oxygen directly participates in the formation of O─O bonds and generates oxygen vacancies. Although external oxygen molecules can dynamically compensate for these oxygen vacancies, the compensation rate is usually slower than the defect generation rate, leading to irreversible metal site leaching on the catalyst surface and subsequent crystal structure collapse and deactivation [52, 53, 54, 55].

Surface passivation is a common interfacial phenomenon in the harsh operating conditions of PEMWE anodes (Figure 2d). It refers to the formation of catalytically inert oxide/hydroxide passivation layers on the catalyst surface driven by the oxidation potential. The generated passivation layer occupies active sites and forms a physical barrier, severely hindering the adsorption of water molecules, proton transport, and electron transfer processes [56, 57]. This significantly slows down the reaction kinetics, resulting in a decrease in the overall efficiency of the electrolyzer. Especially under large‐scale and long‐term operation conditions, the passivation phenomenon will continue to intensify. As the reaction proceeds, the local passivation zones can evolve into a global irreversible phase transformation, ultimately leading to the collapse of the catalyst structure and complete loss of performance.

2.3. Mass Transfer Hindrance

The gas‐liquid two‐phase flow within the PEMWE poses significant challenges to mass transfer (Figure 2e). The evolution of bubbles involves nucleation, growth, and detachment processes [26, 58]. At the high current densities required for large‐scale applications, the bubble generation rate far exceeds the removal rate, easily clogging the electrolyte channels and covering the active sites of the electrodes. The pore structure and wettability of the porous transport layer (PTL) and the catalytic layer (CL) in the MEA directly determine the mass transfer efficiency [59]. If the flow channels are designed improperly or the wettability is poorly controlled, it will aggravate bubble accumulation, reduce mass transfer and cause local overheating within the electrolyzer. Mass transfer resistance can even generate mechanical stress, leading to the deterioration of the catalyst structure [60, 61, 62]. In addition, ions such as Ca²⁺ and Mg²⁺ in the feed water can combine with impurities such as phosphate, easily forming scale and blocking the channels [18]. In large‐scale application scenarios, the uneven distribution of water and gas within the large‐area electrodes will further amplify the mass transfer limitations. The voltage loss caused by mass transfer resistance directly increases the electrolysis energy consumption, significantly reducing the energy efficiency and economic viability of large‐scale PEMWE applications.

3. Strategies for Promoting Scalable PEMWE

To tackle the challenges hindering the large‐scale application of the PEMWE, it is necessary to break through the limitations of laboratory research and advance toward real industrial deployment. This paper proposes a strategic pathway from classical experimental optimization to AI‐enabled approaches, aiming to improve the comprehensive performance of PEMWE from three core dimensions. To systematically summarize the enhancement effects of different strategies, Table 1 lists the key data related to the industrial application adaptability of PEMWE under various optimization strategies. Based on Table 1, the following text provides a detailed explanation and analysis of the mechanism and implementation effect of each category of optimization strategies.

TABLE 1.

Summary of PEMWE performance parameters under various strategies

Methods	Precious metal content	Tafel slope	Stability performance	Degradation rate	Refs.
Ru‐Co2MnO4.5	20 µg cm−2	29 mV dec−1	Over 100 h (1 A cm−2)	0.7mV h−1	[68]
MD‐RuZrCoCrCeO2	0.37 mg cm−2	46.5 mV dec−1	600 h (0.2 A cm−2)	21.67 µV h−1	[73]
HEA@Ir‐MEO	0.4 mg cm−2	56.2 mV dec−1	500 h (1 A cm−2, 65°C)	29.06 µV h−1	[78]
Nd‐CoOx	/	82 mV dec−1	Over 100 h (0.2 A cm−2)	/	[84]
Ir/IrOx‐SO3H	/	45.1 mV dec−1	1000 h (1 A cm−2, 80°C)	/	[97]
CoLaCa‐SWMI	/	66.74mV dec−1	200 h (1 A cm−2)	/	[101]
LCM	/	/	1200 h (2 A cm−2 with the start‐stop interval of 400 h)	/	[105]
MS Ir‐PTE	0.12 mg cm−2	69 mV dec−1	2000 h (1.8 A cm−2, 80°C)	11.66 µV h−1	[14]
Patterned wettability control	/	/	/	/	[121]
PTL with a gradient pore structure	48.4 µg cm−2	61.8 mV dec−1	210 h (0.5 A cm−2, 60°C)	88 µV h−1	[131]

3.1. Precious‐Metal Optimization & Non‐Precious‐Metal Exploration

In the process of large‐scale application of PEMWE, the high potential and strong acidic extreme environment of OER poses a severe challenge to the long‐term comprehensive performance of catalysts [63, 64, 65]. Currently, the commonly used catalysts are mainly classified into noble metal catalysts and non‐noble metal catalysts. For noble metal catalysts such as Ir‐based and Ru‐based, it is necessary to further enhance their catalytic activity and structural stability under the premise of minimizing the loading of noble metals [24]. Given the high cost and scarcity of noble metals, the development of non‐noble metal OER catalysts has also become an urgent need in the research field. However, under the harsh OER reaction conditions, the vast majority of 3d transition metal oxide catalysts are prone to severe structural degradation, which restricts their practical application [66, 67]. To achieve a balance of low cost, high activity and high stability, researchers have developed a variety of strategies to optimize the performance of noble metal catalysts and explore new non‐noble metal catalysts. These strategies can be roughly divided into four categories: reaction path modulation, structural design, interface engineering, and construction of a self‐healing catalyst.

3.1.1. Modulation of Reaction Pathways

The catalytic activity and stability of catalysts are significantly influenced by the reaction pathways. Both the AEM and the LOM are difficult to achieve a balance between activity and stability, failing to meet the industrial demands for large‐scale application of PEMWE. Therefore, researchers have been dedicated to developing new reaction pathways to overcome these bottlenecks and have successfully discovered two novel reaction paths: oxide pathway mechanism (OPM) and vicinal deprotonation mechanism (VDM). These new pathways fundamentally break through the inherent limitations of traditional reaction pathways by reconstructing the reaction process, optimizing the role of intermediates, and enhancing the stability of the lattice structure.

He et al. [68] successfully synthesized a ruthenium single‐atom anchored on cobalt‐manganese oxide (denoted as Ru‐Co₂MnO_4.5). This catalyst adopts the OPM reaction path and can effectively maintain catalytic activity and stability even at an extremely low loading of 20 µg·cm⁻² of precious metal. Figure 3a shows a schematic diagram of the OPM reaction principle. The OPM pathway does not rely on the traditional *OOH intermediate product but directly couples oxygen radicals (*O) on the bimetallic active sites to form O₂, thereby breaking through the limitations of the traditional AEM mechanism, accelerating the reaction rate, and reducing catalyst damage [69, 70, 71]. The researchers conducted density functional theory (DFT) calculations to study the mechanism regulation and the intrinsic reasons for the excellent performance of Ru‐Co₂MnO_4.5. As shown in Figure 3b, the formation energy of oxygen vacancies at two positions around the Co atom in Ru‐Co₂MnO_4.5 is significantly higher than that in Co₂MnO_4.5. This indicates that the introduction of Ru single atoms can enhance the stability of lattice oxygen and prevent its dissolution. They also found that the loading of Ru significantly shortens the Co─O bond length (Figure 3c). This confirms that the introduction of Ru can generate local compressive strain, effectively inhibiting the collapse of the catalyst's crystal structure [72]. As shown in Figure 3d, unlike the AEM pathway, the rate‐determining step of OPM is *O + *OH → *O + *O + H⁺ + e⁻ [11]. At the Ru‐Co position, the energy barrier is only 1.99 eV. At the Ru‐Mn position, the energy barrier is 2.11 eV. Therefore, Ru‐Co₂MnO_4.5 preferentially follows the OPM path at the Ru─Co dual active sites [24]. Additionally, they tested the stability of this catalyst in a PEM device, and the results showed that the device could operate stably for 100 h with a degradation rate of only 0.7 mV h⁻¹ (Figure 3e).

FIGURE 3.

(a) Schematic representation of the OPM mechanisms for Ru‐Co₂MnO_4.5. (Green: Ru, blue: Co, red: O, white: H). (b) Oxygen vacancy formation energies at distinct sites in Ru‐Co₂MnO_4.5 and Co₂MnO_4.5. The inset displays the structural model for oxygen vacancy in Ru‐Co₂MnO_4.5. (c) Co−O bond lengths around the Ru site in Ru‐Co₂MnO_4.5 and Co₂MnO_4.5. The insets display the local structures before and after Ru single‐atom doping. (d) Gibbs free energy profiles for Ru−Co₂MnO_4.5 at Ru−Co sites under both AEM and OPM reaction pathways. (e) Chronopotentiometry curve of the PEMWE employing Ru−Co₂MnO_4.5 as the anode and 60% Pt/C as the cathode at 80°C. Reproduced with permission [68]. Copyright 2025, American Chemical Society. (f) Schematic illustrations of the LOM and VDM mechanisms. (g) The peak area ratio of ³⁴O₂/³²O₂ determined by mass spectrometry for MD‐RuZrCoCrCeO₂ and C‐RuO₂. (h) Vacancy formation energies associated with surface lattice oxygen and subsurface oxygen species in MD‐RuZrCoCrCeO₂ and RuO₂. (i) Chronopotentiometry curves of C‐RuO₂ and MD‐RuZrCoCrCeO₂ at 200 mA·cm_geo ⁻² in the PEM electrolyzer. Reproduced with permission [73]. Copyright 2026, John Wiley and Sons Ltd.

In addition to the OPM pathway, Peng et al. [73] further proposed the VDM reaction path. They successfully prepared a multi‐element rutile phase solid solution MD‐RuZrCoCrCeO₂ by uniformly doping and modifying MOFs precursors. Traditional Ru‐based catalysts tend to follow the LOM pathway in acidic media, where lattice oxygen directly participates in the reaction process, leading to easy deterioration of the catalyst structure and poor stability [74]. In contrast, the MD‐RuZrCoCrCeO₂ catalyst mainly follows the VDM reaction path. The oxygen intermediate adsorbed on the Ru active sites can transfer protons to adjacent lattice oxygen sites, achieving a rapid deprotonation process and effectively reducing the reaction energy barrier (Figure 3f) [75]. The researchers used ¹⁸O isotope labeling combined with in‐situ differential electrochemical mass spectrometry (DEMS) technology to complete five consecutive cyclic voltammetry (CV) tests. The results in Figure 3g show that the ³⁴O₂/³²O₂ ratio of MD‐RuZrCoCrCeO₂ is approximately half that of C‐RuO₂, directly confirming that the degree of inhibition of the LOM reaction path in MD‐RuZrCoCrCeO₂ is 50%. This significantly enhances the stability of the catalyst under acidic conditions. To further verify the stable anchoring ability of lattice oxygen in MD‐RuZrCoCrCeO₂, the researchers calculated and analyzed the formation energy of oxygen vacancies in surface lattice oxygen and subsurface oxygen in MD‐RuZrCoCrCeO₂ and RuO₂. As shown in Figure 3h, the formation energy for the removal of surface lattice oxygen in MD‐RuZrCoCrCeO₂ is 2.71 eV, much higher than 2.29 eV in RuO₂. This indicates that the doping of Zr and Co elements can enhance the stability of lattice oxygen within the crystal structure, providing stable proton transfer sites for the efficient operation of the VDM reaction path [76, 77]. The theoretical calculation results show that the activation energy of the rate‐determining step of the VDM reaction path is 1.57 eV, which is lower than that of the traditional reaction path. The long‐term stability test results of the PEMWE in Figure 3i shows that the electrolyzer loaded with MD‐RuZrCoCrCeO₂ catalyst can operate stably for 600 h, with a voltage decay rate of only 21.67 µV·h⁻¹, fully demonstrating its good potential for large‐scale application. After evaluation, the hydrogen production cost of the PEMWE equipped with MD‐RuZrCoCrCeO₂ is US$0.89 kg⁻¹, which is lower than the standard set by the US DOE.

The innovative OPM and VDM reaction mechanisms have broken through the inherent activity limit of AEM and the stability defect of LOM. Their rapid reaction kinetics and high stability make them more suitable for the dynamic operating conditions of fluctuating power sources such as wind and photovoltaic power. However, they still have some shortcomings. Given that PEMWE devices generally adopt a stacked assembly structure, the interface contact, electron transfer efficiency, and long‐term operational stability of OPM and VDM type catalysts in real stacked systems have not been fully verified. Therefore, future research can be conducted in this direction to promote the application of these new catalytic mechanisms from laboratory basic research to actual large‐scale application.

3.1.2. Structural Design

The structure of the catalyst is the core framework that determines its intrinsic catalytic activity and structural stability, directly influencing the hydrogen production efficiency and service life of PEMWE in large‐scale applications. Reasonable structural design can precisely control the microstructural and electronic properties of the catalyst, thereby effectively inhibiting the dissolution of metal components, reducing lattice distortion, and alleviating structural stress caused by high working potential and bubble impact. In recent years, researchers have proposed two new structural strategies, namely the core‐shell structure and the dynamic “breathing” structure, which mitigate metal corrosion and buffer operational stress damage, thus providing feasible approaches for improving the catalytic performance of PEMWE.

Cheng et al. [78] successfully synthesized HEA@Ir‐MEO nanoparticle catalysts using a two‐step wet chemical reduction method (Figure 4a). The core feature of this catalyst is its core‐shell structure, consisting of an IrRuNiMo medium‐entropy oxide shell (Ir‐MEO) and an IrRuCoNiMo high‐entropy alloy (HEA) core. First, NaBH₄ was used as a reducing agent to prepare HEA nanoparticles in an ethylene glycol system. Subsequently, the spontaneous reduction and deposition of Ir⁴⁺ were utilized to construct a medium‐entropy oxide shell on the surface of the HEA particles. Eventually, a core‐shell structured HEA@Ir‐MEO nanoparticle catalyst was formed. Due to the differences in atomic radius and electronegativity among Ru, Co, Ni, and Mo elements in the core, it can induce ultra‐small crystal size and lattice contraction effects, thereby providing a stable substrate support for the outer Ir‐MEO shell [79, 80]. The high‐angle annular dark‐field scanning transmission electron microscopy (HAADF‐STEM) pseudo‐color image in Figure 4b visually confirmed the successful construction of the Ir‐rich shell on the surface of the HEA core [81]. Quantitative analysis revealed that the shell thickness was approximately three atomic layers. As shown in Figure 4c, the metal leaching level of HEA@Ir‐MEO was significantly lower than that of pure HEA, indicating that the Ir shell can effectively inhibit the leaching of metal components and enhance the structural stability of the catalyst [82, 83]. Additionally, the stability number (S number) of HEA@Ir‐MEO was higher than that of pure HEA, further verifying its structural stability advantage in the OER reaction. The researchers further conducted application tests of this catalyst in actual PEMWE. The results reveal that the electrolyzer exhibits a degradation rate of merely 29.6 µV·h⁻¹ after 500 h of continuous electrolysis, strongly confirming the excellent long‐term stability of the catalyst at industrial current densities (Figure 4d). As shown in Figure 4e, when the Ir loading of HEA@Ir‐MEO was only 0.4 mg·cm⁻², the corresponding electrolysis voltages at current densities of 2.0 and 3.0 A·cm⁻² were only 1.70 and 1.85 V, respectively.

FIGURE 4.

(a) Schematic of the prepared HEA and HEA@Ir‐MEO. (b) Fake color image of AC‐HAADF‐STEM. (c) ICP‑MS determination of dissolved metal ions in HEA and HEA@Ir‑MEO following activation process and chronopotentiometry testing. Inset: corresponding S‑number recorded during 24‑h electrolysis. (d) Chronopotent‐iometric curve of the PEM electrolyzer using HEA@Ir‐MEO catalyst, measured at 1.0 A cm⁻² under 65°C with Nafion 115 membrane. (e) Polarization curves of the PEM electrolyzer recorded at 80°C using Nafion 115 membrane. Reproduced with permission [78]. Copyright 2024, Wiley‐Blackwell. (f) Structural models of Nd‐CoO_x in the OER process. (g) Changes in Co─O and Nd─O bond length and the Raman shift of A_1g peak as a function of applied voltage. (h) Schematics depicting the structural modifications of CoO_x and Nd‐CoO_x before and after OER. (i) Long‐term chronopotentiometry evaluation of Nd‐CoO_x at 200 mA cm⁻² in the PEMWE device over 100 h. Reproduced with permission [84]. Copyright 2025, Springer Nature.

Although non‐noble metal spinel‐type Co₃O₄ catalysts exhibit comparable OER activity to noble metal catalysts, their practical application is still limited by issues such as easy dissolution and insufficient conductivity. Zhong et al. [84] developed a Nd‐CoO_x catalyst. The doping of Nd elements can trigger a dynamic “breathing mode” structure in the catalyst, endowing it with good structural flexibility and effectively inhibiting structural collapse under external high potentials. The researchers successfully prepared the Nd‐CoO_x catalyst using a two‐step method combining microwave‐assisted heating (MAH) and solution combustion synthesis (SCS). Figure 4f shows the “breathing mode” behavior characteristics of Nd‐CoO_x after applying an electric potential. During the reaction, the Co─O and Nd─O bonds can be compressed and stretched, thereby balancing the atomic‐scale structural stress generated within the catalyst during the OER reaction. It was found that the introduction of Nd atoms first compressed the Co─O bonds to a length shorter than the initial length of the Co─O bonds in pure Co₃O₄, and then exerted a stretching effect on the surrounding Co─O bonds [21, 22, 85]. As shown in Figure 4g, the bond lengths of Co─O and Nd─O in the Nd‐CoO_x catalyst could fully recover to their initial states before and after the OER reaction, while the bond lengths of both types in pure CoO_x catalysts could not recover to their initial levels after the reaction. This indicates that this dynamic “breathing” structure is controllable and reversible, thereby maintaining its structural stability and avoiding the decline in catalytic performance due to irreversible structural evolution and resulting structural stress [86]. To deeply explore the mechanism of action of this catalyst, researchers adopted the DFT computational method to construct a theoretical calculation model of the Nd‐CoO_x catalyst. The results indicated that for pure CoO_x, the presence of vacancies could enhance its OER catalytic activity, but it was also prone to further oxidation of the Co active sites, thereby causing continuous contraction of the catalyst lattice [86, 87, 88]. In contrast, the structural “breathing mode” triggered by the introduction of Nd atoms, with its elastic structural changes, could effectively resist structural damage under high potential (Figure 4h). As shown by the test results in Figure 4i, the PEMWE electrolyzer based on this catalyst could stably operate for over 100 h at a current density of 0.2 A·cm⁻², with the working voltage maintained at around 1.69 V, demonstrating excellent potential for practical application. Meanwhile, the Tafel slope of Nd‐CoO_x is relatively small, at only 82 mV dec⁻¹, indicating that Nd‐CoO_x has favorable OER kinetics.

The above two strategies can effectively solve the problem of catalyst structure collapse under harsh reaction conditions through precise structural design. However, in the large‐scale application of PEMWE, this type of core‐shell structure catalyst highly depends on the synergistic effect of five elements. During the large‐scale synthesis process, element segregation is prone to occur, making it difficult to achieve the uniform atomic distribution at the laboratory level, which in turn leads to significant performance differences between batches of the catalyst. The dynamic “breathing” structure is expected to be extended to other lanthanide element systems, providing new ideas for the development of non‐noble metal catalysts. However, the current stability tests of this structure have not yet reached the industrial application standard in terms of current density. Subsequent stability tests at higher current densities are needed to achieve a more comprehensive assessment of its performance.

3.1.3. Interface Engineering

The OER process relies on the microscopic behavior at the catalyst‐electrolyte interface. The proton transfer efficiency, hydration mode, adsorption and transformation characteristics of intermediates at the interface directly determine the speed of the catalytic kinetics [45, 89, 90, 91]. Interface stability is a key factor affecting the long‐term lifespan of the catalyst and preventing the leaching of active components [92, 93]. Interface engineering is indispensable for improving the performance of PEMWE and is a crucial approach to enhancing the catalytic efficiency and stability of PEMWE in large‐scale applications. Recently, two new interface engineering strategies have been proposed, namely the sulfonic groups (‐SO₃H) grafting strategy and the lanthanum doping strategy.

Previous studies have confirmed that sulfonate adsorbed on the catalyst surface can promote the proton transfer process and accelerate the OER reaction [94, 95]. However, the interaction between sulfonate and the catalyst is weak, and it will seriously leach during the long‐term operation of PEMWE [96]. Therefore, Zhang et al. [97] designed and developed the Ir/IrO_x‐SO₃H catalyst based on the interface engineering strategy by covalently bonding the ‐SO₃H to IrO_x. This strategy can achieve precise and stable grafting of ‐SO₃H groups through fewer steps at a lower operating temperature. The reaction mechanism is as follows: the hydroxyl groups on the surface of Ir/IrO_x undergo a nucleophilic reaction with 1,3‐propanesultone, and the sulfonic acid group forms a covalent bond with the surface of Ir/IrO_x through propoxy linkers (−OCH₂CH₂CH₂−) (Figure 5a). The OER reaction mechanism on the catalyst surface is shown in Figure 5b: (1) H₂O and ‐SO₃ ⁻ interact through hydrogen bonds, promoting the formation of the *OH intermediate; (2) ‐SO₃ ⁻ accelerates the deprotonation of *OH through hydrogen bond interaction, generating the *O intermediate; (3) *O is hydrated to form the *OOH intermediate; (4) ‐SO₃ ⁻ further promotes the deprotonation of *OOH, ultimately releasing O₂. The characterization results before and after the reaction show that the characteristic absorption peak of ‐SO₃H does not disappear or shift, indicating that the group is relatively stable. The Tafel slope of Ir/IrO_x‐SO₃H is relatively small, at 45.1 mV dec⁻¹. In the 1000 h continuous operation test of the PEM electrolyzer, the Ir/IrO_x‐SO₃H catalyst only showed a slight voltage increase of 17 mV (Figure 5c), demonstrating excellent stability. In addition, this preparation strategy can achieve efficient large‐scale synthesis of tens of grams of Ir/IrO_x‐SO₃H catalysts, meeting the batch production requirements of membrane electrode assemblies (MEAs). Researchers have fabricated a large‐scale MEA of 54 cm × 45 cm based on this catalyst (Figure 5d), fully demonstrating the potential for large‐scale application of this interface modification strategy.

FIGURE 5.

(a) Synthesis diagram of grafting ‐SO₃H catalyst. (b) Schematic diagram depicting the OER mechanism over the Ir/IrO_x‐SO₃H surface. (c) Chronopotentiometry curves of PEMWE assembled with Ir/IrO_x‐SO₃H and Ir/IrO_x anodes were obtained at 1.0 A cm⁻² and 80°C. (d) Photograph of the large‐scale Ir/IrO_x‐SO₃H MEA with dimensions of 54 cm × 45 cm. Reproduced with permission [97]. Copyright 2025, Springer Nature. (e) Schematic diagram of H₂O molecules on pure Co₃O₄ and lanthanum‐modified Co₃O₄ surfaces. (f) The interfacial water structure evolution of CoLaCa‐SWMI and Ca‐Co₃O₄ catalysts within the potential range of 1.2–1.8 V. (g) Polarization curves of CoLaCa‐SWMI and Co₃O₄ measured in pure water at 80°C. (h) Chronopotentiometry test performed on CoLaCa‐SWMI with 25 cm² active area at 1 A cm⁻² in PEMWE. Reproduced with permission [101]. Copyright 2026, Springer Nature.

Considering that the dissolution of non‐noble metal oxides is based on bond breaking and ion solvation, a regulated interface water environment is expected to reduce the risk of metal site dissolution caused by excessive polarization of water molecules on the catalyst surface [98, 99, 100]. Peng et al. [101] designed and prepared the CoLaCa‐SWMI catalyst, achieving a synergistic improvement in the activity and stability of non‐noble metal catalysts in PEM electrolyzers by regulating the microenvironment effect at the water‐catalyst interface at the atomic scale. They introduced lanthanum and calcium into Co₃O₄ to prepare CoLaCa‐SWMI. The introduction of Ca can create highly active coordination‐unsaturated cobalt sites, significantly enhancing the intrinsic activity of the catalyst. The doping of La can moderately reduce the number of hydrogen bonds on the surface of Co₃O₄, effectively inhibiting the dissolution and loss of cobalt sites. Figure 5e visually explains the intrinsic mechanism of La doping in reducing the metal dissolution rate from the perspective of electronic structure. The 3d orbitals of Co are in an unfilled state and can form a large number of hydrogen bonds with water molecules [102]. However, the 4f orbitals of lanthanum oxide are shielded by the fully filled 5s²5p⁶ electron layer and are difficult to form bonds with water molecules, thus creating a hydrogen bond defect surface [103, 104]. On the surface with hydrogen bond defects, the binding energy of the metal‐oxygen bond is significantly enhanced, leading to an increase in the dissociation energy of the cobalt active sites in CoLaCa‐SWMI. In the CoLaCa‐SWMI catalyst, the proportion of O‐H groups on the catalyst surface at 3600 cm⁻¹ is 20%, much higher than that of the Ca‐Co₃O₄ sample, indicating that this interface engineering strategy does not form excessive hydrogen bonds during the OER process (Figure 5f). The researchers also constructed a PEMWE system. As shown in Figure 5g, this system can achieve current densities of 1 and 2 A·cm⁻² at cell voltages of 1.78 and 1.98 V, respectively, demonstrating excellent electrochemical performance and meeting industrial application requirements. During the test, the degradation rate of PEMWE within the range of 0 to 200 h was only 0.06 µV·h⁻¹. To promote its large‐scale application, the researchers assembled a PEMWE system using a large‐sized anode with an active area of 25 cm² for testing. The results showed that the electrode containing CoLaCa‐SWMI could stably operate for 200 h at a current density of 1 A·cm⁻², with no significant decline in catalytic performance (Figure 5h).

Both of these studies are based on the industrial application requirements of the PEMWE for hydrogen production. They have expanded and verified the interface‐engineered electrocatalysts from small‐scale laboratory test systems to large electrodes. This design fully highlights the practical application value and industrialization potential of their catalytic strategy. However, the actual operating conditions in industrial scenarios are more complex, such as frequent start‐stop of the electrolyzer and temperature fluctuations. Therefore, future research can further simulate more complex environments to expand its catalytic performance and stability under extreme conditions.

3.1.4. Self‐Healing Catalysts

To address the bottleneck of insufficient stability of non‐noble metal catalysts in large‐scale applications, catalysts with dynamic self‐healing functions offer innovative solutions to this problem. Unlike the traditional catalyst design concept centered on passive anti‐degradation, self‐healing catalysts can utilize the ion dissolution‐redeposition behavior under electrocatalytic conditions to construct a dynamic and reversible active site regeneration system. Their unique self‐healing function can actively suppress the excessive dissolution of the active components of the catalyst, thereby extending the catalyst's service life under actual operating conditions.

Recently, Hou et al. [105] successfully prepared La‐doped cobalt manganese oxide (LCM) catalysts with dynamic self‐healing properties. The LCM catalysts were synthesized using an aqueous nitrate solution calcination method at a calcination temperature of 250°C. Figure 6a shows a schematic diagram of the dynamic self‐healing mechanism of the LCM catalyst. Its core advantage lies in the stable main skeleton structure, and the dissolved active sites can be self‐repaired through the redeposition process. They investigated the structural and morphological evolution after continuous operation for 24 h at a current density of 2 A cm⁻² in a PEMWE. Co ions dissolved and entered the electrolyte, while Mn elements preferentially underwent oxidation reactions to act as a charge buffer, effectively inhibiting the excessive dissolution of Co ions [106, 107]. They observed that the LCM catalyst still retained the characteristic A_1g vibration peak of the spinel structure after the reaction, confirming its stable spinel skeleton structure (Figure 6b). Additionally, a non‐crystalline layer with a thickness of approximately 1 nm formed on the catalyst surface (Figure 6c). This non‐crystalline layer was generated by the surface reconstruction process, providing abundant active sites for the dissolution‐redeposition process. As shown in Figure 6d, after PEMWE operation, the oxygen vacancy/lattice oxygen (O_V/O_L) ratio of the LCM catalyst increased from 1.1 to 2.0, indicating a significant increase in the oxygen vacancy concentration. Abundant oxygen vacancies can regulate the electronic structure of the catalyst and introduce additional energy levels, providing favorable conditions for the redeposition of metal ions [66, 87, 108, 109]. The theoretical calculation results in Figure 6e,f indicate that under high potential conditions, Co elements will deposit on the catalyst surface in the form of CoO₂, which is more prone to redeposition compared to the ionic state Co²⁺. The additional O atoms in CoO₂ can fill the oxygen vacancies in the LCM catalyst, repairing the defective surface and further enhancing its structural stability. In contrast, the re‐deposition energy barrier of the cobalt‐manganese oxides (CM) catalyst is higher than that of the LCM catalyst, confirming that the La‐induced oxygen vacancies are the key to promoting efficient self‐healing of the LCM catalyst under high potential. PEMWE performance tests demonstrated that the anode loaded with LCM catalyst exhibited excellent stability. With a start‐stop interval of 400 h, after continuous operation for 1200 h, the electrolyzer energy consumption increased by only 8.45% while maintaining the initial hydrogen production rate (0.837 L H₂ cm⁻²·h⁻¹) (Figure 6g).

FIGURE 6.

(a) Schematic diagram depicting the dynamic stability mechanism of LCM catalysts. (b) Raman spectra of LCM before and after 24 h at 2 A cm⁻² in PEMWE. (c) False‐colored bright field STEM image of LCM after 24 h at 2 A cm⁻² in PEMWE. (d) High resolutions of O 1s XPS spectra of LCM after 24 h at 2 A cm⁻². (e) Phase diagrams illustrating the re‐deposition behavior of Co‐containing species on LCM and CM. (f) Diagram of the re‐deposition mechanism at lower and higher potentials. (g) Galvanostatic tests of the LCM‐catalyzed PEMWE operated at 2 A cm⁻² at 80°C with Nafion 115 membrane (LCM anode catalyst loading 0.5 mg cm⁻²). Reproduced with permission [105]. Copyright 2025, Royal Society of Chemistry.

The LCM catalyst has broken away from the reliance on the scarce noble metal Ir and innovatively proposed a dynamic self‐healing strategy suitable for the PEMWE system. However, the current preparation process of LCM catalysts is still limited to small‐scale synthesis in the laboratory. In the future, efforts should be made to develop scalable preparation techniques to promote the practical industrial application of this dynamic self‐healing strategy.

3.2. Mass Transfer Enhancement

The ultimate goal of large‐scale application of PEMWE is to achieve long‐term stable hydrogen production under low cell voltage and high current density. Under high current density conditions, the anode of the electrolyzer must ensure the rapid transport of water molecules to the active sites of the CL and the efficient desorption and discharge of oxygen bubbles [110]. If the mass transfer process is limited, it can easily lead to the active sites being blocked by bubbles, a decline in electrolysis efficiency, and even local dehydration and failure of the proton exchange membrane [111, 112, 113]. By enhancing the mass transfer process, excellent electrolysis performance can be maintained while reducing the loading of precious metals in the catalyst [114]. In addition, efficient mass transfer within large‐scale stacks is crucial for the stable operation of industrial‐scale hydrogen production systems. To address the issue of mass transfer enhancement, various strategies have been developed, including the construction of nanochannels, the patterned regulation of surface wettability, and the design of gradient pore structures, which support the development of PEMWE toward large‐scale application.

3.2.1. Nanochannels

The precise size design of nanochannels can prevent the random diffusion of gases and provide dedicated transport paths for reactants and products. Additionally, by leveraging the capillary action of the channels, stable material transport in PEMWE can be maintained under harsh conditions such as low feed and low humidity. Peng et al. [14] proposed the construction of a nanochannel structure beneath the PTL islands to provide an additional path for the interfacial transport of oxygen and water, effectively solving the key problem of low interfacial mass transfer efficiency.

The research team found that after actual operation of PEMWE, the anode CL and PTL interface presented two distinct morphological features: the area in direct contact with the PTL showed indentation deformation, while the non‐contact area remained intact and flat. The researchers defined the high‐contact area as the PTL island. A set of irregular voids will form within the indented regions, which essentially arise from the accumulation of local gas pressure as oxygen produced by the OER cannot be rapidly vented out, squeezing the catalyst layer. Figure 7a shows a schematic diagram of the interfacial transport kinetics of the anode PTL/CL‐PEM interface. The research first used femtosecond (FS) laser ablation technology to etch the PTL surface and prepare a nanochannel structure (Figure 7b). The width of the nanochannels was controlled within the range of 50 to 100 nm, resulting in each PTL island being densely packed with nanochannels (Figure 7c) [115]. To further increase the number of etched channels, the research team also used fiber laser ablation to prepare microscale patterns and combined it with femtosecond laser etching to construct nanochannels, naming the modified electrode MS Ir‐PTE [116, 117]. As shown in Figure 7d, compared with Ir‐PTE and FS Ir‐PTE, the double‐layer capacitance and mass activity of MS Ir‐PTE were significantly improved. Such multi‐scale composite laser etching drastically enlarges the electrochemically active interfacial area, alleviates the stacking and agglomeration of catalyst particles, and boosts catalyst utilization efficiency. The ohmic overpotential of the nanochannel electrode decreased, confirming that this structure can effectively improve the hydration state of the proton exchange membrane (Figure 7e) [118]. These results indicate that the nanochannels beneath the PTL islands can significantly promote the interfacial transport of water and oxygen, thereby increasing the electrode utilization. The researchers further conducted vapor‐fed water electrolysis experiments under different inlet relative humidity (RHs) conditions. Under vapor‐fed conditions, the electrolysis performance of PEMWE is more sensitive to the water transport path [119]. As shown in Figure 7f, when the inlet relative humidity is as low as 70%, the performance of the MS Ir‐PTE modified PEMWE is basically the same as that at 100% relative humidity, with only a slight deviation when the current density is higher than 600 mA·cm⁻². The nanochannels form continuous water delivery pathways, compensating for insufficient water supply under low‐humidity vapor feed and mitigating mass transport limiting losses at high current densities. In addition, MS Ir‐PTE exhibits excellent long‐term operational stability in PEMWE. As shown in Figure 7g, under a constant current of 1.8 A·cm⁻², at 80°C and atmospheric pressure, the PEMWE can operate stably for 2000 h, with an average degradation rate of only 11.66 µV·h⁻¹.

FIGURE 7.

(a) A schematic of the transport behaviors of liquid water reactant, oxygen gas, electrons, and protons across the anode PTL/CL‐PEM interfaces. (b) Schematic diagram illustrating the PTE/PEM interfaces established by the nanochannel electrode. (c) SEM image revealing the surface morphology of the nanochannel structure on a PTL island. (d) Double‐layer capacitance (dark blue) and mass activity at 1.45 V (yellow) evaluated for each electrode. (e) Comparison of ohmic overpotential for the three PTEs. The error bars correspond to the standard deviation from two independent tests. (f) Polarization curves recorded at different RHs for MS Ir‐PTE under vapor‐fed water electrolysis. (g) Stability test of MS Ir‐PTE at 1.8 A·cm⁻² under 80°C and ambient pressure. Reproduced with permission [14]. Copyright 2024, Elsevier.

The nanochannel strategy can meet the demand for continuous and stable operation of large‐scale PEMWE for thousands of hours. Nevertheless, traditional PEMWE electrodes mostly adopt mature preparation techniques such as ink coating and ultrasonic spraying, while nano‐channel electrodes rely on laser etching technology, which has poor compatibility with existing electrode mass production lines. If promoting its industrial application, the existing production lines need to be restructured, which will incur certain equipment modification and technology adaptation costs. However, from a long‐term development perspective, this strategy still has good feasibility and promotion value in high‐performance and long‐life PEMWE devices.

3.2.2. Patterned Wettability Control

During the operation of PEMWE, the wettability of PTL directly determines the gas‐liquid two‐phase flow pattern and gas saturation inside it. Superhydrophilic PTLs tend to cause bubbles to accumulate for a long time before detaching, while completely hydrophobic PTLs result in insufficient water transport during the OER process [60, 61, 120]. To overcome the inherent mass transfer bottleneck of PTLs with a single wettability, Liao et al. [121] established a three‐dimensional, two‐phase, dual‐scale pore network model (PNM). By designing the wettability of PTLs in a patterned manner, they achieved spatial separation and directional transport of the gas and liquid phases.

Figure 8a shows a schematic diagram of the established PNM model. The researchers extracted the key geometric parameters of the pore regions of PTL and CL, modeled each individual pore separately, and connected them through throat structures to ultimately construct a dual‐scale PNM model with pore‐throat characteristics [122, 123]. A quasi‐static capillary invasion percolation algorithm is adopted in this model to simulate the transport behavior of oxygen inside pore channels, while the dissolved oxygen generation process in the catalyst layer is coupled to reveal the two‐phase transport laws at the anode pore scale. By comparing the calculated polarization curves and gas saturation data with experimental measurements, the PNM model was verified to have a small calculation error, confirming its rationality and effectiveness [124, 125]. Based on this validated model, the researchers further explored the interaction laws between the PTL pores/surfaces and liquid water. The study found that, compared with “uniform wettability” and “random mixed wettability”, constructing ordered hydrophilic/hydrophobic regions within the PTL with a fixed “width” and “hydrophilic/hydrophobic ratio” can effectively control the gas distribution state within the PTL, achieving gas‐liquid separation (Figure 8b) [126]. As shown in Figure 8c, when the pattern width is large, the gas saturation shows a “double‐peak” distribution characteristic, while when the pattern width is small, it only shows a “single‐peak” distribution. This phenomenon indicates that narrow‐width patterned designs can force gas to occupy small pore spaces, preventing large pores from being blocked by liquid water, thereby ensuring efficient liquid water transport through large pores and achieving gas‐liquid separation. Similarly, as shown in Figure 8d, when the spatial proportion of hydrophilic and hydrophobic regions within the PTL is large, a similar gas‐liquid separation effect can be obtained. Model calculation results show that when the pattern width is 25 µm, the system gas saturation is the lowest (18%), the number of gas pores is the largest, and the average pore diameter is the smallest (Figure 8e) [127]. This indicates that at this width, the gas dispersion is excellent and does not affect the water transport efficiency of large pores. As shown in Figure 8f, when the hydrophilic/hydrophobic ratio is 3:1, gas can be dispersed in more small pores, the overall gas saturation is lower, and the mass transfer efficiency is optimal. The 3:1 hydrophilic‐hydrophobic ratio balances the continuity of water transport channels and the quantity of gas exhaust channels. It ensures sufficient water supply to the catalyst layer while rapidly discharging oxygen generated by reactions, and mitigates the mass transfer overpotential induced by bubbles covering catalytic active sites. Therefore, when the PTL adopts a patterned wettability design with a hydrophilic/hydrophobic ratio of 3:1 and a pattern width of 25 µm, the optimal gas‐liquid separation effect can be achieved.

FIGURE 8.

(a) Schematic illustration of dual‐scale PNM establishment based on reconstructed PTL and CL. (b) Schematic of the PEM electrolyzer anode PTL with patterned wettability. (c) Gas pore size distribution for patterned structures with different widths. (d) Gas pore size distribution for patterned structures with various patterned wettability ratios. (e) Gas‐phase saturation, gas pore number and average gas pore diameter at different patterned widths. (f) Gas‐phase saturation, gas pore number and average gas pore diameter at various patterned wettability ratios. Reproduced with permission [121]. Copyright 2025, American Chemical Society.

The patterned wettability strategy theoretically solves the key problem of gas and liquid phases mutually occupying channels and blocking the transport path, achieving a synergistic improvement in the two‐phase transport efficiency. Currently, this strategy is still at the PNM numerical simulation stage, and the related theoretical predictions have not been experimentally verified in actual PEMWE devices. In the future, multi‐physical field performance tests and characterizations of PEMWE can be conducted based on the existing model. By comparing theoretical results with experimental data, the PNM model and key parameters of patterned wettability can be continuously optimized to promote the practical application of this strategy.

3.2.3. Gradient Pore Structure

In the research on mass transfer optimization in the PTL/CL interface, a stepwise gradient distribution of pore size and porosity of the porous transport layer along a specific spatial direction can better match the mass transfer processes, such as water transport, gas release, and charge conduction within the PEMWE system [128, 129, 130]. Based on this design concept, Zhou et al. [131] prepared a microporous layer (MPL) using cerium dioxide (CeO₂) nanoparticles and microcrystalline cellulose (MCC) as raw materials and coated it on the PTL surface, successfully constructing a composite porous transport layer with a gradient pore structure.

The MPL they prepared can effectively fill the large‐sized pores on the PTL surface, optimize the pore size distribution, and form a continuous conductive network at the PTL/CL interface. The more uniform surface morphology after regulation can significantly suppress the accumulation of oxygen at the interface and promote the efficient transport of water molecules to the catalytic active sites (Figure 9a). This is because a smooth interface weakens the attachment and nucleation sites for oxygen bubbles, and the tiny pores generate stronger capillary driving force to continuously and rapidly evacuate oxygen produced by electrolysis, thus alleviating concentration polarization caused by gas clogging under high current densities. The reduction in surface roughness leads to a uniform pressure distribution on the catalyst array, which can maximize the integrity of the PTL structure. The SEM characterization results in Figure 9b show that the PTL surface after coating with MPL presents good flatness and the pore size is significantly reduced. As shown in Figure 9c, when the pore size is less than 80 µm, the penetration of the PTL coated with MPL is significantly higher than that of the blank PTL. However, when the pore size is greater than 80 µm, there is no significant difference in the penetration between the two. This phenomenon confirms that the filling effect of the CeO₂ slurry on the PTL can form a gradient pore structure. The surface micropores match the mass transfer requirements at the catalyst layer interface, while the underlying layer retains the original macropores to guarantee overall fluid circulation. This gradient pore structure significantly enhances the electrode reaction kinetics and capillary pressure by improving the contact state at the PTL/CL interface, thereby optimizing the mass transfer efficiency [62, 132, 133]. Polytetrafluoroethylene (PTFE) is used as a hydrophobic agent and binder in the MPL [134, 135]. The Tafel slope increases with the increase in PTFE concentration (Figure 9d). Excessive PTFE content will reduce the electrode conductivity and increase the interface contact resistance, while too low a content will hinder the timely release of oxygen. Figure 9e shows the curves of mass activity and contact angle changes of the membrane electrode assembly (MEA) under different PTFE contents. This figure indicates that an appropriate amount of PTFE can achieve a balance between gas‐liquid exchange efficiency and the stability of the MPL structure, avoiding structural fracture and detachment during long‐term stability tests at industrial current densities. The researchers optimized the parameters through experiments, using the best parameters (52 nm CeO₂ and 40 µm MCC) to prepare the MPL and fill the PTL, and assembled it into a PEMWE for stability tests. The results show that under a constant current of 500 mA·cm⁻² at 60°C, this PEMWE exhibits excellent stability, with a voltage decay rate of only 88 µV·h⁻¹ within 210 h (Figure 9f). After long‐term stability tests, the structure of the catalyst array coated with MPL remains intact, without obvious collapse or severe agglomeration, further confirming that the efficient mass transfer mediated by the gradient pore structure is crucial for improving the long‐term stability of PEMWE (Figure 9g) [136].

FIGURE 9.

(a) Schematic illustration of the ordered MEA based on MPL. (b) Top‐view SEM micrograph of the MPL‐free PTL. Inset: CLSM image of the identical PTL without MPL. (c) Pore size distribution ranging from 0 to 300 µm. (d) Tafel slopes of MEAs with different PTFE content. (e) Mass activity and contact angle of MEAs prepared with various PTFE contents. (f) Constant current evaluation of PEMWE at 60°C using ordered MEA with 52 nm CeO₂/40 µm MCC MPL, without MPL and without ordered MEA and MPL. (g) SEM images of filled ordered MEA. Reproduced with permission [131]. Copyright 2025, American Chemical Society.

This gradient pore construction strategy can build directional gas‐liquid transport channels, improving capillary pressure to achieve efficient water supply while providing a low‐resistance escape path for reaction product gases. However, in the current stability tests, the continuous operation time of this PEMWE has not yet met the requirements for large‐scale application. The structural evolution and durability of the gradient pore structure during long‐term use still need further systematic verification and evaluation.

3.3. Artificial Intelligence Empowerment

In the rapidly evolving digital era, artificial intelligence is becoming the core driver for the large‐scale application of PEMWE, bringing disruptive changes. The current development cycle for new PEMWE catalysts is lengthy, significantly increasing overall research costs. On one hand, the synthetic feasibility of novel materials involves substantial uncertainty. On the other, obtaining sufficient high‐quality electrochemical experimental data requires substantial financial investment. Moreover, catalyst screening must simultaneously consider multiple criteria, including catalytic activity, stability, and electrode mass transport efficiency. Therefore, it is essential to adopt a scientifically accurate and cost‐effective approach in theoretical calculations and data screening to avoid the additional expenses associated with prolonged computational processes and high computing power. AI, with its powerful computational capabilities and autonomous learning ability, precisely addresses these challenges [137, 138]. This is the core rationale behind integrating AI—a cutting‐edge and highly efficient tool—into traditional experimental frameworks. It enables effective exploration of vast material and process parameter spaces within controlled computational budgets, rapidly identifying optimal solutions that balance catalytic activity, stability, and cost, thereby fundamentally reducing the time and monetary costs associated with conventional trial‐and‐error approaches [139, 140].

AI empowers traditional experiments across multiple dimensions. Table 2 lists the strategies enabled by various artificial intelligence technologies. At the atomic scale, AI can accurately analyze the microscopic mechanisms of reactions. In terms of performance, artificial intelligence can predict a material's synthesizability, activity, and stability [141]. From a data perspective, AI can integrate information and intelligently expand limited datasets, thereby cutting research costs [142, 143]. Furthermore, the self‐driving laboratories enable fully automated closed‐loop iteration across synthesis, fabrication, and performance testing. This not only reduces costs, but also simultaneously improves catalyst stability and electrode mass transfer efficiency. AI precisely matches and resolves key constraints in PEMWE systems, accelerating the iterative advancement of PEMWE technology and providing solid support for the large‐scale, low‐cost deployment of green hydrogen.

TABLE 2.

Summary of strategies empowered by artificial intelligence.

Applicable scenarios	Algorithm models	Advantages	Limitations	Refs.
Atomic‐scale structure analysis	MLIP	High computational efficiency Analyzing the mechanism of the catalyst from the atomic level	Fail to directly simulate dynamic conditions Prediction relies on the training model	[146]
Prediction of catalyst synthesis and performance	MA (XGBoost, SHAP, AutoEIS)	Reduce the cost waste caused by the synthesis of ineffective catalysts Multi‐objective collaborative optimization	Involves a “black‐box” model in the process	[155]
Expansion of low‐cost datasets	MLP, Bayesian algorithm	Address the pain points of high acquisition costs and scarce samples for high‐quality experimental data Improves generalization and prediction accuracy of small‐sample models	Involves a “black‐box” model in the process	[157]
Self‐driving laboratories	NN model, Bayesian optimization algorithm	Strong nonlinear fitting capability Can be embedded in an unmanned automation closed loop	Complex parameter tuning Depends on prediction model accuracy	[165]

3.3.1. Atomic‐Scale Structure Analysis

Atomic‐scale modeling and analysis provide an important theoretical basis for screening the optimal elemental combinations in multi‐component systems. However, traditional standard methods such as DFT face significant challenges when conducting quantum mechanical (QM) calculations and analyses of such materials [27, 144, 145]. The core reason lies in the vast configuration space that researchers need to explore, which leads to extremely high computational costs. The rapid development of AI technology offers an effective solution to this problem [27]. With higher computational accuracy and efficiency, AI can provide unconventional technical support for the design of multi‐component alloy catalysts in PEMWE through precise atomic‐scale analysis.

Yang's team [146] developed a machine‐learned interatomic potential (MLIP) and combined it with replica‐exchange molecular dynamics (REMD) and Monte Carlo atom swaps (MC) methods to conduct mechanistic analysis at the atomic scale for Ru_x(Ir,Fe,Co,Ni)_1−x multi‐component alloys and guide the design of catalysts [147]. This approach addresses the limitation of conventional density functional theory, which suffers from excessive computational costs and thus struggles to perform large‐scale and long‐timescale simulations of multi‐principal element alloys. They first determined through experiments that Ru_0.20(Ir,Fe,Co,Ni)_0.80 was the optimal composition among different Ru doping ratios. Based on the symmetry‐adapted perturbation theory – atomic position smooth overlap of atomic positions (SOAP) method, the team combined MLIP with REMD/MC techniques to deeply study the atomic‐scale segregation characteristics related to the stability of alloy phases [148]. The SOAP descriptor can quantitatively characterize the local atomic coordination environment and supply high‐dimensional structural feature inputs for MLIP to rapidly learn bonding and repulsive interactions within quinary transition metals. Through calculations, they obtained the atomic relaxation equilibrium configuration and the Cowley's short‐range order (SRO) parameter, as shown in Figure 10a [149]. The elements within the face‐centered cubic (fcc) main phase were well mixed [150]. At the same time, they found that Ni had the strongest segregation tendency, while the slight segregation of Ru was the main cause of the formation of the hexagonal close‐packed (hcp) phase [150]. Since the initial MLIP was only trained based on fcc/bcc phase data and could not effectively study the hcp phase, the researchers further trained the MACE MLIP to enable REMD/MC simulations of the Ru_0.20(Ir,Fe,Co,Ni)_0.80 alloy hcp phase bulk [151]. It was ultimately confirmed that the hcp phase also had excellent atomic mixing. Using the same research strategy, the team explored the element segregation behavior of this alloy system on the (111), (100), and (110) crystal planes of the fcc structure. The results in Figure 10b indicated that the acid‐resistant Ir element was enriched in the first few layers of the alloy, effectively inhibiting the dissolution and loss of Ru active sites. This is because surface‐enriched Ir readily forms a stable IrO_x passivation layer under the strongly oxidizing acidic OER environment, isolating direct contact between the electrolyte and subsurface Ru and blocking the corrosion pathway where Ru is oxidized into soluble RuO₄. To further verify the reliability of the MLIP model's calculation results, the researchers compared the total energy of the relaxed equilibrium alloy configuration with that of the corresponding special quasi‐random structures (SQSs) [152]. The results showed that all the alloy configurations after REMD/MC equilibrium were more stable than the corresponding SQS configurations (Figure 10c). Through atomic‐scale analysis, the research team expanded the study system to all RuIr‐based five‐component alloys (RuIr combined with three 3d transition metals). Leveraging the efficient computational advantage of MLIP, they conducted high‐throughput screening of 120 alloy combinations (Figure 10d). Compared with traditional computational methods, MLIP reduces the time cost for phase stability evaluation of a single alloy composition by several orders of magnitude, enabling rapid preliminary screening of massive component candidates. Except for Sc, Ti, and Cr, any three elements selected from Mn/Fe/Co/Ni could form uniform solid solutions when combined with Ru/Ir. As shown in Figure 10e, all 3d metals except Cu exhibited good mixing behavior with Ru and Ir, providing direct atomic‐scale theoretical guidance for the rational component design of multi‐component alloy catalysts in the OER reaction.

FIGURE 10.

(a) Average SRO for atomic pairs in bulk Ru_x(Ir,Fe,Co,Ni)_1−x alloys configurations from the final 3000 REMD/MC steps. Equilibrium structure model of Ru_0.20(Ir,Fe,Co,Ni)_0.80 is also displayed. (b) Average proportions and corresponding standard deviations of each metallic element in the surface and subsurface layers of Ru_x(Ir,Fe,Co,Ni)_1−x (111), Ru_0.20(Ir,Fe,Co,Ni)_0.80 (100), and Ru_0.20(Ir,Fe,Co,Ni)_0.80 (110) surface structures, obtained from the final 3000 REMD/MC sampling steps. Equilibrium structure models of Ru_0.20(Ir,Fe,Co,Ni)_0.80 (111), (100), and (110) surfaces are also provided. (c) Comparison of the total energies for the relaxed SQS bulk and (111) surface of Ru_x(Ir,Fe,Co,Ni)_1−x alloys with those of the relaxed equilibrium structures derived from REMD/MC simulations, with all energy values computed via DFT. (d) Scatter plot correlating the minimal mean SRO with its standard deviation. (e) Correlation plot showing the mean value (represented by circle color) and standard deviation (represented by circle size) of the average SRO (obtained from the final 3000 REMD/MC steps) for atomic pairs in equimolar RuIr‐based bulk HEAs containing three additional 3d transition metals. Reproduced with permission [146]. Copyright 2025, American Chemical Society.

This strategy, which utilizes AI for atomic‐scale analysis, mainly focuses on the static atomic structure of the catalyst after preparation, and can assist in explaining the intrinsic mechanism by which the catalyst enhances the performance of PEMWE. In the future, the core breakthrough direction of AI atomic‐scale analysis lies in achieving the simulation of atomic behavior under dynamic working conditions, making it closer to the actual working environment of catalysts, and further promoting the design of multi‐component alloy catalysts.

3.3.2. Prediction of Catalyst Synthesis and Performance

During the research process of PEMWE, the feasibility of material synthesis often lacks precise pre‐prediction methods, resulting in a large amount of experimental resources being wasted on ineffective trial‐and‐error exploration, which seriously slows down the pace of its development [37, 38]. At the same time, the core performance of PEMWE catalysts, namely catalytic activity and stability, have a complex coupling relationship, making it difficult to achieve precise and coordinated prediction of both, which has become one of the bottlenecks for the large‐scale application of PEMWE technology [153, 154]. Nowadays, the targeted development of AI algorithms has effectively solved these two major problems, significantly accelerating the scientific research process for the industrialization and large‐scale application of PEMWE technology.

Sargent and his team [155] developed a hierarchical multi‐modal workflow called mixed acceleration (MA). All three layers of predictive models within the entire MA framework uniformly adopt the XGBoost gradient boosting decision tree algorithm to complete classification and regression tasks, which can effectively avoid overfitting. This workflow can accurately predict complex material properties and incorporate heterogeneous materials into the machine learning‐guided material discovery system. The MA workflow includes a modular high‐throughput (HT) platform for material synthesis, characterization, and electrocatalytic performance testing, as well as three independent ML models, which are used to predict the synthesis results of sol‐gel reactions, evaluate the OER activity of materials, and assess the stability of catalysts (Figure 11a) [146, 156]. From the vast chemical space, the gel space with synthetic feasibility is first screened out, followed by the screening of high‐activity OER material space through the activity model, and finally, the OER material space with both activity and stability is locked down with the help of the stability model (Figure 11b). SHapley Additive exPlanations (SHAP) was adopted in this research to unpack the black‐box XGBoost model and realize interpretable analysis of the structure‐activity relationship of materials. This screening process intuitively demonstrates the efficient screening advantage of the MA workflow. To verify the effectiveness of the MA framework, the researchers designed and compared five rounds of iterative experiments. I1: initial ternary data set; I2: random quaternary composition combinations; I3: expert experience guidance; I4: traditional ML guidance; I5: MA framework guidance. The results in Figure 11c show that without introducing the gel model, the gelation success rate of I1 was 80%, while that of I2 and I3 was only about 60%. In contrast, after introducing the gelation prediction model, I4 achieved a 100% gelation reaction success rate in 50 synthetic materials, fully verifying the reliability of the prediction model. Figure 11d indicates that under the assistance of the MA framework, more than 90% of the materials in the I5 group had a ruthenium dissolution rate of less than 1%. The activity and stability of catalysts usually have a mutually restrictive relationship, which is manifested as a Pareto frontier in multi‐objective optimization. Seven catalysts in the I5 group successfully broke through this Pareto frontier, namely Ru_0.5Zr_0.1Zn_0.4O_x, Ru_0.5Zn_0.4Ni_0.1O_x, Ru_0.6Zn_0.3Cu_0.1O_x, Ru_0.6Zr_0.1Zn_0.3O_x, Ru_0.5Mg_0.1Zn_0.4O_x, Ru_0.8Fe_0.1Ni_0.1O_x, and Ru_0.7Cs_0.2Rb_0.1O_x. The MA strategy proposed in this study does not rely on high‐precision quantum mechanics simulations but only requires DFT calculation results of medium precision that can capture the experimental change trends to achieve efficient predictions. The Pearson correlation coefficient (R_Pearson) was used to evaluate the predictability of the OER stability model. The results show that with the MA framework, the R_Pearson value of this model reached 0.725, and the average absolute percentage error was 65.3% (Figure 11e). In addition, the maximum relative prediction uncertainty of the MA strategy for each material is less than 14%, which is much lower than the 64% of models constructed only based on pre‐synthesis information. Traditionally, testing one material takes hundreds of hours. The MA framework can reduce the time required to obtain the structure, composition, and OER activity‐related data of each material to approximately 10 min, significantly lowering the time and economic costs of experiments. AutoEIS was further matched and utilized in this work to batch‐extract double‐layer capacitance and charge transfer resistance with high precision, so as to quantitatively track the evolution law of active interfaces of catalysts during long‐term cycling.

FIGURE 11.

(a) Schematic diagram of MA layered multi‐modal workflow. (b) Illustration showing the progressive reduction of the accessible stable OER catalyst space with the aid of MA. (c) Comparison of synthetic results under the guidance of the gelation model versus without it: purple bars denote successful gelation; pink bars signify gelation failure. (d) Correlation plot of overpotential against ruthenium dissolution for catalysts. (e) Comparison diagram of stability results obtained by MA and experiments. Reproduced with permission [155]. Copyright 2026, Springer Nature.

Relying on the powerful data processing and intelligent prediction capabilities of artificial intelligence technology, it can effectively reduce the excessive reliance on human experience in catalyst research and development, and greatly shorten the development cycle of new catalytic materials. At the same time, AI can accurately predict the OER activity and long‐term stability of catalysts in acidic environments, precisely meeting the core requirements of PEMWE large‐scale applications for high activity and high durability of catalysts. The assistance of AI can help researchers quickly screen out catalytic systems with excellent comprehensive performance, laying a key material foundation for the large‐scale promotion and industrialization of PEMWE technology.

3.3.3. Expansion of Low‐Cost Datasets

Capturing the complex nonlinear relationships among variables using AI typically relies on large‐scale high‐quality datasets. When the data volume is insufficient, the model is prone to overfitting, making it difficult to learn universal rules and resulting in poor generalization ability, which hinders its effective application in new scenarios. Traditional experimental methods for data acquisition suffer from long cycles, high costs, and limited conditions, making it difficult to quickly accumulate a sufficient number of samples. To address the bottleneck of difficult model training and insufficient generalization in small sample scenarios, Liu et al. [157] introduced transfer learning into the modeling process. By leveraging existing knowledge from related fields and pre‐training information, they effectively mitigated the overfitting problem caused by small datasets, providing a feasible approach for rule mining of PEMWE under limited data conditions [142].

The researchers first built a pre‐training model based on 11 025 large‐scale source data [138]. Large‐scale source datasets provide abundant sample support for neural networks to fully capture the underlying nonlinear correlation between electrolysis voltage and fabrication parameters. After optimizing the hyperparameters through the Bayesian algorithm, they determined to use a Multilayer perceptron (MLP) structure with 6 hidden layers (32 nodes per layer) [158]. The stacked multi‐hidden‐layer structure is well‐suited to describe the complex response behavior arising from multivariate coupling in electrocatalytic systems. The input variables were 15 preparation and operation parameters, and the output variable corresponded to the electrolysis voltage (Figure 12a). They divided the source data into a training set and a test set in a 4:1 ratio and completed the model training under a learning rate of 0.01. The final model's coefficient of determination R² reached 0.99655, verifying the good reliability of the pre‐training model (Figure 12b). Since the exclusive target data set for the optimization of low‐Ir MEA catalyst ink was only 12 groups, the researchers further constructed a target MLP model fine‐tuned by transfer learning (Figure 12c). This model froze the first two fully connected hidden layers, adjusted the input layer dimension to 8 (7 catalyst ink formula parameters + current density), and the output layer was the electrolysis voltage. During the fine‐tuning process, the learning rate was reduced to 0.001, and it was set to decay by 5 times every 10 iterations. Eventually, 300 target data points were expanded to complete the model training. As shown in Figure 12d, the R² of the target model reached 0.99507, significantly better than the MLP model directly trained with a small data set (R² was only 0.87489) [159]. Subsequently, the researchers used the Harris Hawk Optimization (HHO) algorithm to optimize the catalyst ink formula [160]. This algorithm took the electrolysis voltage of the low‐Ir MEA component at 3.0 A·cm⁻² predicted by the MLP model as the objective function. After more than 250 iterations, it converged to the global minimum and determined the optimal ink formula (Figure 12e). Based on the best ink formula obtained through this AI optimization, MEA‐opt was prepared (Figure 12f). Compared with MEA9 and MEA10, the catalyst‐ionomer particle size of MEA‐opt was the smallest and less prone to agglomeration (Figure 12g). The viscosity of MEA‐opt was within the range of 0.1 to 0.15 Pa·s, which was more conducive to the uniform spraying of the catalytic layer [161, 162]. As shown in Figure 12h, MEA‐opt presented the most uniform color distribution, with a standard deviation (σ_SD) of 4.96. It should be noted that the performance verification of MEA‐opt was only a short‐term static test, and long‐term durability cycling experiments have not been conducted. This direction still needs to be further improved in subsequent research.

FIGURE 12.

(a) Schematic structure diagram, (b) training loss curve and scatter plot for predictive performance of the pre‐trained model. (c) Schematic structure diagram, (d) training loss curve and scatter plot for predictive performance of the target model. (e) Objective function variation during iterative processes. (f) Size distribution of particles and (g) steady‐shear viscosity profiles of the inks used for fabricating MEA9, MEA10 and MEA‐opt. (h) Schematic diagram of catalyst ionomer particle dispersion in MEA‐opt after optimization of ink formula. Reproduced with permission [157]. Copyright 2024, Elsevier BV.

Different from the traditional “black box method”, this dataset expansion approach provides clear physical meaning guidance for the optimization of PEMWE. Additionally, this strategy can be transferred to the research and development process of other components of PEMWE. Only a small number of data samples are needed for fine‐tuning to expand the dataset, significantly reducing the cost of dataset construction in similar studies. However, this strategy still has certain limitations. Currently, it only realizes a one‐way process from dataset expansion to optimization verification. Based on this, the optimized experimental data can be fed back to the model for secondary fine‐tuning to further improve the model performance.

3.3.4. Self‐Driving Laboratories

The self‐driving laboratory is a new research paradigm centered on AI, automated platforms, and closed‐loop experimental systems. It can independently complete the entire process of scientific research tasks, including experimental design, sample preparation, performance testing, and data analysis, without human intervention [163, 164]. Luo et al. [165] developed a self‐driving laboratory for the automatic synthesis and intelligent optimization of catalysts for the OER reaction, which increased the research and development efficiency by five orders of magnitude.

Figure 13a shows the two‐layer workflow of this self‐driving laboratory. The outer layer consists of mobile robots and “intelligent” chemical workstations, covering 12 steps of automated experimental operations and data management processes. The inner layer's nine consecutive digital operations are regulated by an intelligent computing “brain”. This self‐driving laboratory first automatically detects the elemental content of five Martian meteorites through laser‐induced breakdown spectroscopy (LIBS) technology, constructing an available elemental library. LIBS relies on pulsed laser ablation of ores to generate characteristic plasma spectra, enabling accurate quantification of key metals and restricting the relative error of elemental content detection within ±5%. Then, it combines molecular dynamics simulation and the DFT calculations to theoretically predict the OER reaction activity of the catalysts and builds a neural network (NN) model based on this. The verification results show that the constructed NN model has high accuracy and reliability in predicting the OER catalytic performance (Figure 13b). Subsequently, the robot automatically synthesizes 243 catalysts according to a random ratio scheme, and uses the obtained experimental data to retrain and optimize the NN model. At the same time, it incorporates the Bayesian optimization algorithm. It adopts a Gaussian process surrogate model to quantify uncertainties in the formulation space, balances exploration of untested components and exploitation of well‐known high‐activity species, and greatly accelerates the screening of numerous combinatorial candidates. Based on this workflow, the optimal catalyst formula (Model‐guided OPT) is identified from 3.76 million elemental combinations in total [166]. As shown in Figure 13c, the elemental ratios of Model‐guided OPT are significantly different from those obtained by relying solely on simulation or experimental data for optimization, confirming that this self‐driving laboratory can effectively escape local optimal solutions and achieve global optimal search for OER catalytic formulas. Figure 13d–f shows the real‐time process of fully automated electrochemical testing, intuitively demonstrating the self‐driving laboratory's full‐chain automation capabilities in the device and application stages.

FIGURE 13.

(a) The workflow for designing and producing OER electrocatalysts using self‐driving laboratories. (b) The prediction results of OER overpotentials by predictive model. (c) Kiviat diagram of elemental ratios. (d) Location of the electrochemical workstation. (e) Insertion of working electrodes into the electrolyser. (f) Solar‐driven OER process. Reproduced with permission [165]. Copyright 2023, Springer Nature.

Although this study has not yet been tested in actual PEMWE, it points out a good direction for the future development of PEMWE. The self‐driving laboratory, relying on high‐throughput experiments and the efficient iterative optimization of AI algorithms, has built an autonomous closed‐loop research and development system, providing an intelligent and sustainable technical path to break through the bottlenecks of PEMWE industrial application. With the continuous maturation of related technologies, the construction of a self‐driving laboratory for PEMWE is clearly feasible, which can significantly reduce research and development exploration costs and achieve a high degree of standardization in the experimental process.

Overall, the above strategies provide brand‐new solutions for the large‐scale application of PEMWE from different technical paths. This strategy system, which combines traditional experimental optimization with AI empowerment in a complementary and collaborative manner, can meet the demands of the technological upgrade for the large‐scale application of PEMWE. It not only accurately identifies the core engineering bottlenecks in the industrialization process of PEMWE but also exhibits excellent practicality and prominent innovation. Traditional experiments provide real and reliable scientific training samples and verification scenarios for the construction of AI models. Meanwhile, AI empowerment effectively enhances the research accuracy and efficiency of traditional experiments, making the experimental optimization more targeted. This strategy system further strengthens the connection between technological research and the demands of large‐scale application, and can effectively adapt to the technological upgrade needs of the green hydrogen industry.

4. Summary and Outlooks

In the previous sections, this paper systematically reviewed the three core challenges faced in the large‐scale application of PEMWE. Based on the latest research progress in this field, the corresponding technical improvement strategies were summarized from three aspects: “Precious‐metal optimization & non‐precious‐metal exploration”, “Mass transfer enhancement”, and “Artificial intelligence empowerment”. By establishing a strategic system that integrates traditional experimental research with AI‐assisted design, the comprehensive performance of PEMWE can be significantly enhanced, providing a systematic and feasible technical solution for its large‐scale, low‐cost, and high‐stability operation. Currently, the global hydrogen energy industry is accelerating its transition from gray hydrogen and blue hydrogen to green hydrogen. The paper can provide scientific, systematic, innovative, and feasible theoretical guidance for the large‐scale application of PEMWE. The large‐scale application of PEMWE will further unleash its broad industrial application value and provide efficient, economical, and highly applicable technical solutions for the preparation of green hydrogen.

Although many new strategies have been proposed to improve PEMWE, there is still room for further optimization and improvement in its large‐scale application. Therefore, this paper outlines the research directions that should be paid attention to in the future for the large‐scale application of PEMWE.

First, in the current performance evaluation system of PEMWE, there are differences in the test conditions (such as electrolyte concentration, test temperature, etc.) and the selection of core components adopted by different research teams, which leads to insufficient comparability of research results. At the same time, the existing evaluation indicators mainly focus on the performance of individual components, while lacking unified, multi‐dimensional, and engineering‐oriented evaluation criteria for the large‐scale application of PEMWE. This has become a key bottleneck restricting the transformation of technological achievements and industrial promotion. In the future, it is urgent to establish a standardized evaluation system suitable for the large‐scale application of PEMWE, providing a normative performance evaluation and technical screening basis for the transformation of laboratory achievements to industrial applications.

Second, building a multi‐scenario and multi‐dimensional full‐chain database for PEMWE can provide solid data support and rich training and validation samples for the in‐depth application of AI in this field. This database should not only integrate multi‐dimensional data such as materials, devices, and systems, but also incorporate industrial‐end information such as operation and maintenance and market. It will realize data interoperability and sharing throughout the whole process of research, manufacturing, operation and services, thereby accelerating the progress of large‐scale application. Promoting the open sharing of data platforms helps eliminate cooperation barriers between universities, research institutes and enterprises, further accelerating the iteration of new technologies and their industrial application.

Thirdly, at present, the application of AI in the PEMWE field is still focused on individual components and has not yet achieved multi‐physical field coupling simulation and collaborative optimization of the entire electrolyzer system, which is inconsistent with the actual needs of industrial‐scale application. Subsequent research should integrate multi‐physical field information such as electrochemistry, fluid mechanics, thermodynamics, and mechanics, and rely on AI algorithms to build multi‐physical field coupling models to achieve high‐fidelity dynamic simulation of the operation state of the electrolyzer, thereby solving the actual engineering problems caused by multi‐factor coupling in large‐scale application. The expanded application of AI will promote the design of PEMWE from component optimization to system‐level collaborative design, providing forward‐looking solutions for large‐scale green hydrogen production.

Finally, the actual implementation of PEMWE still faces many engineering challenges. In terms of electrolyzer integration and stack assembly, the interface contact resistance and sealing performance of the stack structure are key factors restricting the overall performance of the system. The stable operation of multiple electrolyzers in coordination is also a core difficulty in the integration of large‐scale equipment. Coupling hydrogen production with renewable energy is an important development direction for PEMWE. In response to the fluctuating energy input from renewable sources, PEMWE needs to have a millisecond‐level rapid response capability to support the gradual replacement of fossil fuels by green hydrogen. Additionally, in terms of system operation and maintenance as well as industrialization support, large‐scale PEMWE devices should move toward precision, intelligence, and convenience. Fault diagnosis should shift from passive detection to active intelligent early warning, and AI should be utilized to enhance operation and maintenance efficiency. Key components should adopt modular design and targeted maintenance to reduce operation and maintenance costs and minimize downtime.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

Data sharing not applicable to this article as no datasets were generated or analysed during the current study.