Deep learning-based structural characterization and mass transport analysis of CO₂ reduction catalyst layers

Abstract

Electrochemical CO₂ reduction (CO₂RR) is a promising technology for mitigating global climate change. The catalyst layer (CL), where the reduction reaction occurs, plays a pivotal role in determining mass transport and electrochemical performance. However, accurately characterizing local structures and quantifying mass transport remains a significant challenge. To address these limitations, a systematic characterization framework based on deep learning (DL) is proposed. Five semantic segmentation models, including Segformer and DeepLabV3plus, were compared with conventional image processing techniques, among which DeepLabV3plus achieved the highest segmentation accuracy (> 91.29%), significantly outperforming traditional thresholding methods (72.35%–77.42%). Experimental validation via mercury intrusion porosimetry (MIP) confirmed its capability to precisely extract key structural parameters, such as porosity and pore size distribution. Furthermore, a series of ionomer content gradient experiments revealed that a CL with an ionomer/catalyst (I/C) ratio of 0.2 had the optimal pore network structure. Numerical simulations and electrochemical tests demonstrated that this CL enabled a twofold increase in gas diffusion distance, thereby promoting long-range mass transport and significantly enhancing CO production rates. This work establishes a multi-scale analysis framework integrating “structural characterization, mass transport simulation, and performance validation,” offering both theoretical insights and practical guidance for the rational design of CO₂RR CLs.

Graphical abstract

Keywords

carbon dioxide reduction / deep learning (DL) / semantic segmentation / catalyst layer (CL) / mass transport

1 Introduction

The electrochemical CO₂ reduction reaction (CO₂RR) is a key technology for achieving carbon neutrality, enabling the conversion of CO₂ into high-value-added chemicals and fuels [1–3]. Among the various limiting factors, CO₂ mass transport process has been identified as the rate-determining step, significantly impacting overall system performance. [4–7]. To enhance reaction activity, CO₂RR adopts design principles from proton exchange membrane fuel cells (PEMFCs), employing a membrane electrode assembly (MEA) electrolyzer [8–10]. The core component, the gas diffusion electrode (GDE), incorporates a catalyst layer (CL) that facilitates direct gas-phase transport of CO₂ to the active reaction sites. [11,12]. As the primary reaction zone, the CL’s microscopic porous structure, defined by factors such as catalyst distribution and ionomer content, plays a crucial role in governing mass transport and electrochemical activity [13–15]. However, the accurate study and characterization of gas transport within the local CL structure remain unresolved challenges.

Previous studies on mass transport within the CL of CO₂RR primarily relied on numerical simulations that treat the CL as a homogeneous medium, thus overlooking its complex and heterogeneous porous structure. For instance, Lees et al. [16] employed a continuum model to simulate an entire electrolyzer and demonstrated that adjusting pH, porosity, and thickness could enhance CO selectivity. Weng et al. [17] developed localized models to investigate the role of GDEs in enhancing CO₂RR, yet did not incorporate real pore structures. Similarly, although some studies on PEMFCs have considered local mass transport within the CL, they primarily rely on artificially generated frameworks [17–19], which fail to accurately replicate the true pore structure, leading to deviations in transport predictions. Therefore, integrating precise structural characterization with simulations is essential for realistically representing local pore networks and accurately evaluating mass transport behavior.

Current characterization techniques for CL structures have notable limitations. Mercury intrusion porosimetry (MIP), for example, provides bulk pore property data but lacks the resolution to capture local features. Scanning electron microscopy (SEM) offers detailed morphological visualization but cannot quantify measurements. [20]. Furthermore, conventional image segmentation methods, such as threshold binarization, Otsu’s algorithm, and K-means clustering, exhibit limited efficacy in distinguishing regions with overlapping grayscale intensities [21–26]. With advances in artificial intelligence, deep learning (DL), with its rapid progress in image analysis, has demonstrated significant advantages in overcoming these challenges. Traditional continuum models also assume homogeneous porous structures with idealized parameters, further constrain realistic analysis. For instance, Berg et al. [27] developed a non-isothermal cathode model that neglected pore-scale heterogeneity, while Armstrong and von Spakovsky [28] employed oversimplified structural assumptions to study oxygen transport. Similarly, Wang et al. [29]’s molecular-scale simulations of PEMFCs failed to capture practical electrode complexities. Although algorithmically generated or randomly packed artificial structures can replicate certain statistical features, they often fail to reproduce authentic electrode morphologies.

In contrast, DL-based semantic segmentation has demonstrated superior capabilities in reconstructing realistic microstructures across diverse domains, including porous rocks [30], PEMFC CLs [31], solid oxide fuel cells [32], and lithium-ion battery cathodes [33]. However, its application to CO₂RR CLs remains largely unexplored. This study aims to fill that gap by integrating semantic segmentation into CO₂RR CLs and systematically identifying the most effective model for accurate structural characterization.

Specifically, this work aims to precisely characterize the local porous structure of CO₂RR CLs and to investigate the mass transport properties. First, the application of semantic segmentation in CO₂RR CL analysis was explored, leading to the development of a systematic characterization approach consisting of sample preprocessing-semantic segmentation-pore parameter calculation. Through comprehensive evaluation, the most suitable semantic model for CO₂RR CL characterization was identified. Based on the systematic method, CL samples with varying ionomer contents were characterized to construct realistic local pore models. The influence of ionomer loading on mass transport is then assessed through numerical simulations, which are further validated by electrochemical experiments. Overall, this work presents a practical tool for studying real microstructures and provides insights into the understanding of the structure-performance relationship in CO₂RR systems.

2 Experimental

2.1 Sample fabrication

Commercial nanosilver catalysts (RHAWN, 60–120 nm), Nafion ionomer solution (DuPont, 10 wt%), anion exchange membrane (dioxide materials, Sustainion X37-50), and gas diffusion layers (GDLs, Ballard, GDS3250) were used as received for GDE fabrication. Isopropanol and ultrapure water (Millipore, > 18.2 MΩ·cm) served as solvents for preparing the cathode catalyst inks.

To prepare the catalyst ink, nanosilver particles, isopropanol, and Nafion ionomer solution were mixed at ionomer/catalyst (I/C) mass ratios of 0.2, 0.4, and 0.6 to produce three distinct catalyst slurries. These slurries were ultrasonicated for 24 h in an ultrasonic bath to ensure homogeneous mixing and uniform dispersion.

The resulting ink was then applied to the GDL via electrostatic spraying, achieving a silver (Ag) loading of 1 mg/cm², which was confirmed by measuring the weight difference of the GDL before and after spraying and drying. Commercial nickel foam (0.15 mm thick) was employed as the anode and was replaced with fresh material for each experiment to minimize corrosion.

2.2 Microstructural characterization

To minimize depth-of-field limitations in SEM imaging and reduce interference from underlying structures, focused ion beam (FIB) [34] and triple ion beam (TIB) [35] techniques were employed for CL preprocessing. A high-resolution FIB-SEM system (TESCAN GAIA3) was utilized, operating in secondary electron (SE) mode with an FIB voltage of 30 kV, an SEM voltage of 5.0 kV, and a magnification of 50 kx. The FIB enabled precise ion beam milling of target regions, while SEM provided high-resolution cross-sectional imaging.

To further ensure high-quality structural characterization, a TIB cutter (Leica EM TIC 3X) was used to produce smooth cross-sections, with ion beam voltages ranging from 5 to 9 kV, a milling rate of 150 μm/h, and a cutting duration of 2 h. Additionally, mercury intrusion porosimetry (MIP, MicroActive AutoPore V 9600) was conducted to quantify porosity and pore structure, providing detailed insights into the CL’s morphological characteristics.

2.3 Image segmentation and pore structure quantification

To accurately extract the local pore structure characteristics of the CL, this study employed image segmentation techniques, utilizing the results for pore parameter calculation and numerical simulations. Three conventional segmentation methods (binary thresholding, Otsu thresholding, and K-means clustering) and five DL-based semantic segmentation models (UNet, PSPNet, DeepLabV3plus, Segformer, and FastSCNN) were evaluated (detailed in SI-1, Electronic Supplementary Material). To enhance the robustness and generalization of the segmentation models, several data augmentation strategies were applied using the OpenCV library. These included random cropping (with a maximum area ratio of 0.75), horizontal or vertical flipping with a probability of 0.5, and random resizing with a scaling factor ranging from 0.5 to 2.0. Such augmentations helped the models adapt to diverse pore morphologies and variations in local structures during training.

Conventional segmentation methods were implemented via OpenCV, where binary thresholding used fixed thresholds, Otsu’s method determined thresholds automatically, and K-means clustering applied unsupervised learning to classify pixels, requiring predefined cluster numbers and iteration limits to adapt to the pore structure. All images used in this study were obtained via high-resolution SEM, with a resolution of 1280 × 896 pixels. For each I/C ratio (0.2, 0.4, and 0.6), 50 SEM images were collected from randomly selected areas across multiple electrode cross-sections to ensure statistical representativeness of the CL morphology. In total, 150 images were included in the data set.

Image annotation was conducted using the Labelme software. Regions of interest were manually labeled into two categories: “Ag” (representing silver catalyst particles) and “Background” (representing pore space and ionomer matrix). The data set was randomly split into a 4:1 training-to-test ratio, and models were trained using the MMsegmentation toolkit [36] with an SGD optimizer, an initial learning rate of 0.001, a batch size of 4, and 20000 iterations. The experiments were conducted on a workstation equipped with an NVIDIA RTX 3090 GPU (25.4 GB VRAM) and a 6-core E5-2680 v4 CPU, using Python 3.7, PyTorch 1.10, and CUDA 11.3.

To evaluate segmentation performance, key metrics including mean accuracy (aAcc), precision, recall, and F1 score (Fscore) were compared. For DL-based models, binary cross-entropy loss was used as the optimization objective, while intersection over union (IoU) and mean IoU (mIoU) were employed to assess segmentation accuracy, with IoU measuring the overlap between predicted and ground truth segments, and mIoU providing an overall segmentation evaluation. Based on the segmentation results, key pore structure parameters—including porosity, pore size distribution, cumulative distribution function, and average equivalent diameter—were extracted. Details on segmentation parameters and pore parameter calculations are provided in SI5-1, Electronic Supplementary Material.

2.4 Local pore model construction

2.4.1 Structural simulation based on QSGS method

The CL consists of silver catalyst particles and the ionomer. Key parameters include catalyst density, ionomer density, the I/C ratio, porosity, and pore size distribution. Among these, catalyst density, ionomer density, and the I/C ratio serve as the primary input parameters for layer fabrication, while porosity and pore size distribution are obtained from the pore structure analysis described in Section 2.3. To reconstruct the microscopic structure of the CL, this study combines semantic segmentation results with the QSGS (quadruple-parameter stochastic growth) method, implemented in MATLAB. The detailed process includes the following steps:

1) Input of segmentation image and pore parameters: The semantic segmentation image classifies each pixel as either “Background” (pore, labeled 0) or “Ag” (catalyst, labeled 1). The image is converted into a binary matrix. The matrix, along with porosity and pore size parameters, serves as input for model construction.

2) Calculation of reconstruction parameters: Based on the I/C ratio, catalyst density, and ionomer density, the required number of silver and ionomer nodes is calculated, using Eqs. (1) and (2).

3) Generation of silver catalyst: The silver catalyst region is assumed to consist of overlapping circular particles [37] with radii ranging from 20 to 40 nm. Within the catalyst-labeled area of the segmented image, silver circle‘s diameter with random centers and radii are generated until the total area meets the target area fraction.

4) Generation of ionomer: Assuming the ionomer uniformly coats the surface of silver particles, ionomer nodes are generated using the QSGS method. Each potential node is assigned a random value from 1 to 10000. If the value is below a preset threshold, the node is converted into an ionomer node. This process continues until the ionomer area fraction reaches the target.

(1){\epsilon }_{\mathrm{A}\mathrm{g}}=\frac{\left(1-\epsilon \right)/{\rho }_{\mathrm{A}\mathrm{g}}}{{\theta }_{\mathrm{I}\mathrm{o}\mathrm{n}}/{\rho }_{\mathrm{I}\mathrm{o}\mathrm{n}}+{\theta }_{\mathrm{A}\mathrm{g}}/{\rho }_{\mathrm{A}\mathrm{g}}},

(2){\epsilon }_{\mathrm{I}\mathrm{o}\mathrm{n}}=\frac{\left(1-\epsilon \right){\theta }_{\mathrm{I}\mathrm{o}\mathrm{n}}/{\rho }_{\mathrm{I}\mathrm{o}\mathrm{n}}}{{\theta }_{\mathrm{I}\mathrm{o}\mathrm{n}}/{\rho }_{\mathrm{I}\mathrm{o}\mathrm{n}}+{\theta }_{\mathrm{A}\mathrm{g}}/{\rho }_{\mathrm{A}\mathrm{g}}},

where ε represents the overall porosity, while ε_Ag and ε_Ion are the area fractions of silver and ionomer, respectively; {\theta }_{\mathrm{I}\mathrm{o}\mathrm{n}} denotes the mass ratio of ionomer to silver, defined as m_Ion/m_Ag; The density of silver and ionomer are {\rho }_{\mathrm{A}\mathrm{g}} (10.49 g/cm³) and {\rho }_{\mathrm{I}\mathrm{o}\mathrm{n}} (2 g/cm³), respectively.

2.4.2 Local pore mass transport model construction

The locally reconstructed microstructure from Section 2.4.1 is imported into COMSOL to build a geometric model and define the computational domain. A mass transport simulation model is developed by meshing, setting governing equations, defining boundary conditions, and assigning physical parameters. The detailed steps are as follows:

1) Model description: The local CL is divided into three regions: pores, ionomer, and silver catalyst. All three regions are included in the computational domain.

2) Model assumptions:

(a) The system operates under isothermal and steady-state conditions.

(b) Gas flow in the pore region follows Brinkman flow.

(c) The reactant gas is CO₂, and only its electrochemical reduction to CO is considered. Both CO₂ and CO are treated as incompressible ideal gases, with only binary diffusion between the models.

(d) The pore and ionomer regions are assumed to be isotropic porous media.

3) Governing equations: Gas flow is described by the Brinkman equation [38], while mass transport is described by the Maxwell-Stefan equation [39], which simulates multicomponent gas diffusion. The full set of equations, along with boundary conditions and key parameters, is provided in Table SI3-1, Electronic Supplementary Material.

Investigating the effective diffusion coefficient of CO₂ is essential for understanding the gas transport mechanism, optimizing pore structures, and improving the performance of CO₂ reduction. Using the COMSOL model, the mass transport coefficient can be obtained via numerical simulation. According to Fick’s law and Faraday’s law, the molar flux of CO₂ in the catalytic layer, J_{\mathrm{C}{\mathrm{O}}_2}, is given by Eqs. (3) and (4).

(3)J_{\mathrm{C}{\mathrm{O}}_2}=-D_{\mathrm{C}{\mathrm{O}}_2}^{\mathrm{e}\mathrm{f}\mathrm{f}}\frac{\mathrm{d}{c}_{\mathrm{C}{\mathrm{O}}_2}}{\mathrm{d}x},

(4)J_{\mathrm{C}{\mathrm{O}}_2}=\frac{i}{nF}.

Thus, the effective diffusion coefficient of CO₂ in the CL structure, D_{\mathrm{C}{\mathrm{O}}_2}^{\mathrm{e}\mathrm{f}\mathrm{f}}, is given by Eq. (5).

(5)D_{\mathrm{C}{\mathrm{O}}_2}^{\mathrm{e}\mathrm{f}\mathrm{f}}=-\frac{i}{nF}{\left(\frac{\mathrm{d}{c}_{\mathrm{C}{\mathrm{O}}_2}}{\mathrm{d}\mathit{x}}\right)}^{-1},

where c_{\mathrm{C}{\mathrm{O}}_2} is the molar concentration of CO₂, x is the transport distance along the x-axis, i is the current density, n is the number of electrons involved in the reaction (with n = 2 in this study), and F is the Faraday constant [40].

2.5 Electrochemical validation

The CO₂RR performance was evaluated using an MEA reactor connected to an electrochemical workstation (Autolab PGSTAT302N). Electrolysis was conducted in a constant current mode, and gas-phase products were collected for quantitative analysis. Gas chromatography (GC, SHIMADZU 2014) with high-purity argon (Air Liquide, 99.99%) as the carrier gas was employed for product detection. The system was calibrated using standard gas mixtures containing CO, H₂, CH₄, C₂H₄, CO₂, and C₂H₆ at concentrations ranging from 100 to 50000 ppm. Carbon-containing products such as CO, CH₄, and C₂H₄ were quantified using a hydrogen flame ionization detector (FID), while a thermal conductivity detector (TCD) measured H₂ concentration.

Faradaic efficiency (FE) analysis was performed to assess the selectivity of CO₂RR products. FE was determined as the ratio of the actual charge consumed in generating a specific product to the total theoretical charge required, as described by Eq. (6) [41]. A higher FE indicates enhanced catalytic selectivity, minimizing side reactions and competing pathways [42,43]. To investigate the effect of ionomer content on catalytic performance, GDEs with varying I/C ratios were assembled into MEA reactors. The FE of each product was calculated based on its concentration at different current densities, providing a quantitative assessment of mass transport and electrochemical performance in relation to CL composition.

(6)F{E}_{\mathrm{C}\mathrm{O}}=\frac{nFcV}{Q}\times 100\mathrm{\%},

where n is the number of electrons transferred in the reduction reaction (n = 2 for CO₂ reduction to CO), F is the Faraday constant, c is the concentration of the CO product (mol/L), V is the electrolyte volume (L), and Q represents the total charge consumed in the reduction reaction (C).

3 Results and discussion

Figure 1 provides the research framework for the structural characterization and mass transport investigation of the CL in the CO₂RR, consisting of four modules: sample microstructural characterization, image segmentation and pore structure quantification, local pore model construction, and mass transport investigation.

3.1 Systematic characterization of CL microstructure

3.1.1 Characterization interface formation

To mitigate depth-of-field limitations in SEM imaging and achieve a smooth, well-defined characterization interface, appropriate preprocessing techniques such as FIB-SEM and TIB were applied to the CL. However, due to the high ductility of Ag, exposure to high-energy ion beams may induce sample deformation, compromising characterization accuracy. Therefore, this study systematically compared the effects of FIB and TIB preprocessing on Ag CLs with an I/C ratio of 0.4 (as shown in Fig. 2) and identified the optimal ion beam energy conditions.

Figures 2(a) and 2(b) present the FIB-SEM characterization results, revealing significant plastic deformation of the Ag CL under high-energy ion beam exposure. Catalyst particles appeared compressed, with localized pore structures either reduced or completely collapsed. Additionally, interparticle boundaries became blurred, leading to unnatural agglomeration, a phenomenon attributed to mechanical stress and thermal effects induced by high-energy ion irradiation. The localized heating further exacerbated particle melting and adhesion, resulting in structural distortions. Consequently, FIB-SEM was found to be unsuitable for accurately characterizing the Ag CL due to its pronounced impact on structural integrity.

In contrast, TIB, which employs an Ar ion beam for polishing, was able to mitigate these high-energy-induced artifacts, as shown in Figs. 2(c)‒2(e). When the ion beam energy was set to 5 or 6 keV, it was insufficient to effectively penetrate the porous structure of the Ag CL, leading to irregular cutting surfaces and structural defects. Conversely, at 8 and 9 keV, excessive beam energy caused sample damage, introducing cutting marks and interparticle fusion, which distorted pore structure and particle distribution. At an optimized energy of 7 keV, the treated interface remained smooth and free of cutting artifacts, making it the most suitable condition for Ag CL preprocessing.

It is acknowledged that the FIB polishing process may influence the local distribution of ionomer materials due to possible localized heating or sputtering. However, in this study, all samples were prepared under the same optimized TIB condition (7 keV), which was selected based on prior reports to minimize damage while ensuring high-fidelity cross-sectional imaging. The uniformity of treatment across all samples ensures internal consistency and minimizes potential bias in structural comparison. Based on these findings, all subsequent experiments employed a 7 keV ion beam energy for preprocessing Ag CLs with different I/C ratios, ensuring a smooth surface and accurate representation of their microstructural characteristics.

3.1.2 Image segmentation comparison

Although TIB-treated samples provide a relatively smooth characterization interface in SEM imaging, depth-of-field limitations still introduce structural interference. Therefore, precise extraction of the CL surface structure requires additional image segmentation techniques. To rapidly compare the effectiveness of semantic segmentation against conventional methods—including binarization thresholding, Otsu’s thresholding, and K-means clustering—this study employed FastSCNN as the semantic segmentation model due to its high processing speed. The ground truth was defined based on manually labeled data, and all four segmentation methods were applied to the same SEM image of the CL. The results are presented in Fig. 3.

A qualitative comparison of different segmentation models is shown in Fig. 3(a). Conventional methods resulted in a significantly higher proportion of the yellow-labeled regions compared to the ground truth, indicating misidentification of underlying catalyst structures as surface features due to SEM depth-of-field limitations. In contrast, semantic segmentation closely matched the ground truth, accurately identifying the surface structure of the CL.

For quantitative evaluation, Fig. 3(b) compares the overall segmentation accuracy using average accuracy (aAcc) as a performance metric. Semantic segmentation achieved the highest accuracy of 91.29%, considerably outperforming the three conventional methods (72.35%–77.42%), highlighting its superiority in overall segmentation performance. Further analysis in Fig. 3(c) assessed the ability of each method to distinguish pores from catalyst regions. Semantic segmentation exhibited a recall rate of 92.02% for pore identification, significantly higher than conventional methods (45.43%–58.36%), demonstrating its effectiveness in addressing misclassification caused by depth-of-field interference.

Considering precision, recall, and F1-score, semantic segmentation consistently exhibited higher accuracy in distinguishing pore structures and CLs, underscoring its critical role in CL microstructural characterization. Therefore, integrating semantic segmentation into CL analysis holds significant value. Furthermore, a systematic evaluation of five semantic segmentation models—Segformer, DeepLabV3plus, PSPNet, UNet, and FastSCNN—was conducted using both qualitative and quantitative assessments, with detailed results presented in SI-4, Electronic Supplementary Material.

3.1.3 Pore structure quantification

Based on the image segmentation results, key geometrical features of the local pore structure in the CL were extracted, including porosity, pore size distribution, cumulative distribution function (CDF), and average equivalent diameter. These parameters were further compared with MIP experimental data. As shown in Figs. 4(a) and 4(b), the pore size distributions obtained from the five semantic segmentation models exhibited high consistency with the MIP results, with DeepLabV3plus demonstrating the best performance. Within the 5–40 nm range, its relative error was as low as 1.7%–7.1%. Unlike MIP, which is limited in measuring small pores (< 5 nm), the semantic segmentation approach provided a more accurate characterization of microporous structures [44,45].

Porosity calculations (Fig. 4(c)) further highlighted the advantages of DL-based methods (39.10%–41.89%), which more closely matched the MIP result (37.87%) compared to conventional segmentation methods (25.84%–31.27%). This improvement stems from the ability of DL models to accurately extract surface catalyst-pore structures while minimizing interference from underlying layers. Additionally, as shown in Fig. 4(d), the pore parameters derived from semantic segmentation models exhibited a significantly higher degree of agreement with experimental data than conventional methods, as demonstrated by the average equivalent diameter obtained via CDF analysis. A detailed error analysis comparing image segmentation-derived pore parameters with experimental results is presented in Table SI-5, Electronic Supplementary Material.

Furthermore, a comprehensive evaluation of different semantic segmentation models was conducted based on six criteria: segmentation accuracy, model simplicity, resource consumption, computational efficiency, learning efficiency, and consistency with MIP results (see SI-5 for details, Electronic Supplementary Material).

Figure 4(e) presents a comparative radar chart that visualizes the performance of five semantic segmentation models across six evaluation criteria: segmentation accuracy, model simplicity, resource consumption, time efficiency, learning efficiency, and consistency with MIP results. Each axis is scored from 0 (worst) to 5 (best), providing a holistic overview of model strengths and trade-offs. As shown, DeepLabV3plus exhibits a well-balanced and expansive profile, indicating strong overall performance. This is attributed to its use of atrous convolutions and an ASPP (atrous spatial pyramid pooling) module, which effectively captures multiscale pore features, making it particularly suitable for complex microporous structures.

While Segformer exhibited strong segmentation accuracy and learning efficiency, its resource-intensive Transformer architecture limited its applicability, particularly for micropore characterization. FastSCNN demonstrated high model simplicity and resource efficiency but was less effective in representing intricate microporous structures. PSPNet and UNet showed slightly lower overall performance compared to DeepLabV3plus.

Based on the analysis, this study establishes a systematic characterization framework for CLs, integrating sample preparation, image segmentation, and pore parameter extraction. This approach enables both visualization and quantitative analysis of local microstructures in CLs. Moreover, DeepLabV3plus was identified as the optimal solution for semantic segmentation and pore parameter calculation in CO₂RR CLs, providing a reliable methodological foundation for the microstructural characterization of electrocatalytic CO₂ reduction systems.

3.2 Mass transport impacts of CL with different ionomer contents

3.2.1 Systematic characterization of CL

The ionomer content is a key factor in regulating the microstructure of the CL, significantly influencing pore distribution and mass transport properties, which in turn determine the efficiency and rate of the CO₂ reduction reaction [46]. Therefore, understanding the impact of ionomer content on pore morphology and mass transport is of critical importance. In this study, CLs with I/C ratios of 0.2, 0.4, and 0.6 were systematically characterized. The microstructures were reconstructed using the QSGS method and subsequently imported into COMSOL for domain discretization, as illustrated in Fig. 5(a).

Figure 5(b) presents the segmentation accuracy results, demonstrating that the DeepLabV3plus model consistently achieved high accuracy across different I/C ratios (88.30%–91.14%). This confirms its robust applicability in accurately identifying and segmenting key structural features, even under varying ionomer contents.

Figures 5(c)–5(f) depict the pore parameters for CLs with different ionomer contents. The pore size distribution primarily falls within the 2–150 nm range, with mesopores and macropores (10–50 nm) dominating the structure. This indicates an open pore network that facilitates mass transport and reaction kinetics [13]. As the I/C ratio increases, the filling effect of the ionomer becomes more pronounced, preferentially covering or obstructing larger pores. This leads to a decrease in macropore fraction and an overall reduction in porosity [47].

Figure 5(g) further elucidates the underlying mechanism of this phenomenon. The selective filling behavior of the ionomer demonstrates a tendency to occupy larger pore spaces, thereby altering the overall pore distribution. This highlights the critical role of ionomer content in shaping the microstructure of the CL.

3.2.2 Local mass transport simulation of CL

Numerical simulations were conducted based on the reconstructed microstructures of CLs with different ionomer contents (as described in Section 3.2.1) under identical conditions to compare and analyze mass transport characteristics. The simulation results are presented in Fig. 6. The pressure field results indicate that CO₂ preferentially diffuses through the pore and ionomer regions due to their significantly higher diffusion coefficients compared to the silver catalyst region. Consequently, high-pressure zones primarily form within the interconnected pores. However, as the gas encounters an increasing amount of Ag catalyst along the diffusion path, mass transport becomes progressively hindered. This occurs for two reasons: first, the lower diffusion coefficient of the Ag catalyst region slows gas flow; second, since the CO₂ reduction reaction occurs within the catalyst region, part of the CO₂ is consumed during diffusion, leading to local variations in gas concentration and pressure.

Additionally, as the I/C ratio increases from 0.2 to 0.6, the extent of the high-pressure region decreases, indicating increased mass transport resistance. Although the distribution image for I/C = 0.4 is smoother, I/C = 0.2 has stronger convection with higher CO₂ accessibility and is therefore more advantageous in terms of mass transport performance.

Velocity field simulations further reveal that gas flows more rapidly through interconnected pores, with a noticeable velocity surge at narrow constrictions. This suggests that irregular pore structures can enhance localized mass transport and reaction efficiency. The highest gas velocity observed in the CL with an I/C ratio of 0.2 is 470 nm/s, whereas for I/C ratios of 0.4 and 0.6, the maximum velocity decreases significantly to 27 and 26 nm/s, respectively, highlighting the detrimental effect of ionomer filling on gas diffusion.

The distributions of CO₂ and CO concentrations exhibit trends consistent with the pressure field results. In the CL with an I/C ratio of 0.2, the high-concentration region is the most extensive, indicating optimal mass transport. As the I/C ratio increases, the extent of the high-concentration region diminishes, further confirming increased mass transport resistance. The CO concentration distribution closely mirrors that of CO₂, as CO is primarily generated from CO₂ reduction. Regions with higher CO₂ concentration and better mass transport performance also exhibit higher CO concentrations.

In summary, appropriately reducing the ionomer content helps maintain pore connectivity, thereby enhancing mass transport and reaction efficiency.

Based on the localized pore structure model of the CL, this study further analyzes mass transport characteristics. The gas diffusion distance (D_diff) was defined as the farthest coordinate from the origin where system pressure reached 0.80 Pa. The CL with an I/C ratio of 0.2 exhibited the longest D_diff (672.88 nm), while increasing the I/C ratio to 0.6 significantly reduced it to 321.38 nm. This indicates that a lower ionomer content (I/C = 0.2) facilitates long-range gas diffusion through open macropores and mesopores, whereas higher ionomer content (I/C = 0.6) restricts diffusion pathways due to pore filling, highlighting the importance of optimizing ionomer content and pore structure.

Additionally, CO concentration profiles were extracted along multiple cross-sections and normalized using the Minimax method to ensure comparability across different scales and units. Results show that in the 0–D_diff range, the CLs with I/C ratios of 0.2 and 0.4 exhibited higher average CO concentrations than those with I/C = 0.6, confirming more efficient gas diffusion and reaction kinetics, consistent with the findings in Fig. 7(a).

The effective CO₂ diffusion coefficient was calculated to assess the ease with which CO₂ molecules traverse the porous structure and reach catalytic active sites. As shown in Fig. 7(c), the computed values align well with experimental data from Sun et al. [48], validating the reliability of the simulations. The effective diffusion coefficient decreases with increasing I/C ratio. Notably, at y = 336 nm, the structure with I/C = 0.6 exhibits a slightly higher CO₂ diffusion coefficient than that with I/C = 0.4. This may be attributed to a relatively more open local pore structure at this position, reducing gas diffusion resistance.

These findings suggest that structural heterogeneities within the CL significantly influence mass transport, with local pore openness playing a critical role. To ensure a focused evaluation of the impact of ionomer content and pore morphology on gas-phase diffusion, the present simulation assumes steady-state conditions and excludes pore flooding or ionomer swelling throughout the CL. These boundary conditions were selected to exclude transient multiphase phenomena and isolate structure-transport relationships. Future work will develop time-dependent multiphysics models incorporating water distribution, ionomer swelling dynamics, and multi-step reaction pathways to provide a more comprehensive understanding of the realistic operating environment in membrane electrode assemblies.

3.2.3 Electrochemical validation of CL

To validate the accuracy of the simulation results, electrochemical performance was assessed using MEA reactors fabricated with GDEs featuring varying ionomer-to-catalyst (I/C) ratios. The MEA reactor employed in this study offers not only robust performance at the laboratory scale, but also shows strong potential for industrial applications due to its compact structure, efficient gas transport, and high current density capability. Importantly, the DL-based segmentation and multi-scale modeling framework developed herein can be readily extended to scaled-up MEA systems, providing morphology-informed guidance for industrial cell design and optimization.

Electrochemical measurements revealed a clear dependence of performance on I/C ratio. CLs with I/C = 0.2 exhibited the highest CO selectivity, reaching 89% at 250 mA/cm², while maintaining the lowest cell voltage of 2.63 V among all tested samples at equivalent current densities. In comparison, I/C = 0.4 and 0.6 resulted in reduced CO selectivity of 86% and 77%, respectively, accompanied by correspondingly higher cell voltages. These experimental trends align well with the simulation results, which showed that lower ionomer loading facilitates gas diffusion by minimizing pore blockage and transport resistance, while also suppressing interfacial charge transfer impedance. Thus, the optimized microstructure at I/C = 0.2 effectively enhances CO₂ accessibility and balances mass transport with reaction kinetics, leading to improved energy efficiency and product selectivity.

4 Conclusions

This study successfully characterized the local structure of CO₂RR CLs and provided a comprehensive analysis of their mass transport properties. By optimizing sample preparation using a 7 keV TIB technique and integrating a DL-based segmentation model, segmentation accuracy was improved by 13.87%–18.94% compared to traditional methods. This advancement enabled precise characterization of CL morphology and pore structures. Among the evaluated models, DeepLabV3plus demonstrated the highest accuracy, facilitating the establishment of a systematic framework encompassing “sample preparation, semantic segmentation, and pore parameter analysis”. Local pore models for CLs with I/C ratios of 0.2, 0.4, and 0.6 revealed that excessive ionomer content severely hinders mass transport, a finding further validated by electrochemical experiments. These findings offer a quantitative foundation for optimizing CL design to enhance mass transport and improve CO₂RR performance at high current densities.

References

[1] Niu J , Zhang C , Liu H . . Enhanced performance of oxygen vacancies on CO₂ adsorption and activation over different phases of ZrO₂. Frontiers in Energy, 2023, 17(4): 545–554

[2] Deng T , Jia S , Han S . . Electrochemical CO₂ reduction to C2+ products over Cu/Zn intermetallic catalysts synthesized by electrodeposition. Frontiers in Energy, 2024, 18(1): 80–88

[3] Lee W H , Ko Y J , Choi Y . . Highly selective and scalable CO₂ to CO electrolysis using coral-nanostructured Ag catalysts in zero-gap configuration. Nano Energy, 2020, 76: 105030

[4] Rudnev A V , Fu Y C , Gjuroski I . . Transport matters: Boosting CO₂ electroreduction in mixtures of [BMIm][BF₄]/water by enhanced diffusion. ChemPhysChem, 2017, 18(22): 3153–3162

[5] Raciti D , Mao M , Park J H . . Mass transfer effects in CO₂ reduction on Cu nanowire electrocatalysts. Catalysis Science & Technology, 2018, 8(9): 2364–2369

[6] Yang K , Kas R , Smith W A . In situ infrared spectroscopy reveals persistent alkalinity near electrode surfaces during CO₂ electroreduction. Journal of the American Chemical Society, 2019, 141(40): 15891–15900

[7] Shi R , Guo J , Zhang X . . Efficient wettability-controlled electroreduction of CO₂ to CO at Au/C interfaces. Nature Communications, 2020, 11(1): 3028

[8] Rabiee H , Ge L , Zhang X . . Gas diffusion electrodes (GDEs) for electrochemical reduction of carbon dioxide, carbon monoxide, and dinitrogen to value-added products: A review. Energy & Environmental Science, 2021, 14(4): 1959–2008

[9] Nguyen T N , Dinh C T . Gas diffusion electrode design for electrochemical carbon dioxide reduction. Chemical Society Reviews, 2020, 49(21): 7488–7504

[10] Wu X , Zhao J Y , Sun J W . . Isolation of highly reactive cobalt phthalocyanine via electrochemical activation for enhanced CO₂ reduction reaction. Small, 2023, 19(23): 2207037

[11] Xiang H . Rasul S. Scott K, et al. Enhanced selectivity of carbonaceous products from electrochemical reduction of CO₂ in aqueous media. Journal of CO₂ Utilization, 2019, 30: 214–221

[12] Lu S , Wang Y , Xiang H . . Mass transfer effect to electrochemical reduction of CO₂: Electrode, electrocatalyst and electrolyte. Journal of Energy Storage, 2022, 52: 104764

[13] He Y , Jiang W J , Zhang Y . . Pore-structure-directed CO₂ electroreduction to formate on SnO₂/C catalysts. Journal of Materials Chemistry. A, Materials for Energy and Sustainability, 2019, 7(31): 18428–18433

[14] Goyal A , Bondue C J , Graf M . . Effect of pore diameter and length on electrochemical CO₂ reduction reaction at nanoporous gold catalysts. Chemical Science, 2022, 13(11): 3288–3298

[15] Jia C , Zhao Y , Song S . . Highly ordered hierarchical porous single-atom Fe catalyst with promoted mass transfer for efficient electroreduction of CO₂. Advanced Energy Materials, 2023, 13(37): 2302007

[16] Lees E W , Bui J C , Song D . . Continuum model to define the chemistry and mass transfer in a bicarbonate electrolyzer. ACS Energy Letters, 2022, 7(2): 834–842

[17] Weng L C , Bell A T , Weber A Z . Modeling gas-diffusion electrodes for CO₂ reduction. Physical Chemistry Chemical Physics, 2018, 20(25): 16973–16984

[18] Mukherjee P P , Wang C Y , Kang Q . Mesoscopic modeling of two-phase behavior and flooding phenomena in polymer electrolyte fuel cells. Electrochimica Acta, 2009, 54(27): 6861–6875

[19] Chen L , Wu G , Holby E F . . Lattice Boltzmann pore-scale investigation of coupled physical-electrochemical processes in C/Pt and non-precious metal cathode catalyst layers in proton exchange membrane fuel cells. Electrochimica Acta, 2015, 158: 175–186

[20] Chen M , Zhao C , Sun F . . Research progress of catalyst layer and interlayer interface structures in membrane electrode assembly (MEA) for proton exchange membrane fuel cell (PEMFC) system. eTransportation, 2020, 5: 100075

[21] Lang J T , Kulkarni D , Foster C W . . X-Ray tomography applied to electrochemical devices and electrocatalysis. Chemical Reviews, 2023, 123(16): 9880–9914

[22] Münch B , Gasser P , Holzer L . . FIB-nanotomography of particulate systems—Part II: Particle recognition and effect of boundary truncation. Journal of the American Ceramic Society, 2006, 89(8): 2586–2595

[23] Otsu N . A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics, 1979, 9(1): 62–66

[24] White R T , Ramani D , Eberhardt S . . Correlative X-ray tomographic imaging of catalyst layer degradation in fuel cells. Journal of the Electrochemical Society, 2019, 166(13): F914–F925

[25] Ariff F N M , Nasir A S A , Jaafar H . . Sauvola and Niblack techniques analysis for segmentation of vehicle license plate. IOP Conference Series: Materials Science and Engineering, 2020, 864(1): 012136

[26] Likas A , Vlassis N , Verbeek J . The global K-means clustering algorithm. Pattern Recognition, 2003, 36(2): 451–461

[27] Berg P , Novruzi A , Promislow K . Analysis of a cathode catalyst layer model for a polymer electrolyte fuel cell. Chemical Engineering Science, 2006, 61(13): 4316–4331

[28] Armstrong K W , von Spakovsky M R . A microscopic continuum model of a proton exchange membrane fuel cell electrode catalyst layer. In: ASME 2006 International Mechanical Engineering Congress and Exposition, 2006, 11(5-10): 349–358

[29] Wang W , Qu Z , Wang X . . A molecular model of PEMFC catalyst layer: Simulation on reactant transport and thermal conduction. Membranes, 2021, 11(2): 148

[30] Wang Y D , Blunt M J , Armstrong R T . . Deep learning in pore scale imaging and modeling. Earth-Science Reviews, 2021, 215: 103555

[31] Colliard-Granero A , Batool M , Jankovic J . . Deep learning for the automation of particle analysis in catalyst layers for polymer electrolyte fuel cells. Nanoscale, 2021, 14(1): 10–18

[32] Hwang H , Choi S M , Oh J . . Integrated application of semantic segmentation-assisted deep learning to quantitative multi-phased microstructural analysis in composite materials: Case study of cathode composite materials of solid oxide fuel cells. Journal of Power Sources, 2020, 471: 228458

[33] Müller S , Sauter C , Shunmugasundaram R . . Deep learning-based segmentation of lithium-ion battery microstructures enhanced by artificially generated electrodes. Nature Communications, 2021, 12(1): 6205

[34] Young R J . Micro-machining using a focused ion beam. Vacuum, 1993, 44(3-4): 353–356

[35] Ji G , Tan Z , Shabadi R . . Triple ion beam cutting of diamond/Al composites for interface characterization. Materials Characterization, 2014, 89: 132–137

[36] Ren H , Meng X , Lin Y . . Structural stability of catalyst ink and its effects on the catalyst layer microstructure and fuel cell performance. Journal of Power Sources, 2022, 517: 230698

[37] Wang M , Wang J , Pan N . . Three-dimensional effect on the effective thermal conductivity of porous media. Journal of Physics. D, Applied Physics, 2007, 40(1): 260–265

[38] Hou Y , Deng H , Pan F . . Pore-scale investigation of catalyst layer ingredient and structure effect in proton exchange membrane fuel cell. Applied Energy, 2019, 253: 113561

[39] Matar S , Ge J , Liu H . Modeling the cathode catalyst layer of a direct methanol fuel cell. Journal of Power Sources, 2013, 243: 195–202

[40] Dong F , Hu Z , Wang Z . . Modeling and optimizing anode catalyst layer for direct ammonia fuel cell. Energy Nexus, 2023, 9: 100170

[41] Kim C , Bui J C , Luo X . . Tailored catalyst microenvironments for CO₂ electroreduction to multicarbon products on copper using bilayer ionomer coatings. Nature Energy, 2021, 6(11): 1026–1034

[42] Qu Y , Yang K , Li W . . Operando diagnosis of MEA-type CO₂ electrolyzer via distribution of relaxation times analysis. ACS Energy Letters, 2024, 9(6): 3042–3048

[43] Ren S , Joulié D , Salvatore D . . Molecular electrocatalysts can mediate fast, selective CO₂ reduction in a flow cell. Science, 2019, 365(6451): 367–369

[44] Senocrate A , Bernasconi F , Rentsch D . . Importance of substrate pore size and wetting behavior in gas diffusion electrodes for CO₂ reduction. ACS Applied Energy Materials, 2022, 5(11): 14504–14512

[45] Goyal A , Bondue C J , Graf M . . Effect of pore diameter and length on electrochemical CO₂ reduction reaction at nanoporous gold catalysts. Chemical Science, 2022, 13(11): 3288–3298

[46] Möller T , Ngo Thanh T , Wang X . . The product selectivity zones in gas diffusion electrodes during the electrocatalytic reduction of CO₂. Energy & Environmental Science, 2021, 14(11): 5995–6006

[47] Hao S , Zhou Y , Guo Y . A brief survey on semantic segmentation with deep learning. Neurocomputing, 2020, 406: 302–321

[48] Sun M , Cheng J , Yamauchi M . Gas diffusion enhanced electrode with ultrathin superhydrophobic macropore structure for acidic CO₂ electroreduction. Nature Communications, 2024, 15(1): 491

RIGHTS & PERMISSIONS

Higher Education Press 2025