pubs.acs.

org/IECR Article

Data-Driven Optimization of High-Dimensional Variables in Proton

Exchange Membrane Water Electrolysis Membrane Electrode
Assembly Assisted by Machine Learning
Yipeng Zhang, Aidong Tan,* Zhuolin Yuan, Kun Zhao, Xiaoyun Shi, Ping Liu, and Jianguo Liu*
Cite This: Ind. Eng. Chem. Res. 2024, 63, 1409−1421 Read Online

ACCESS Metrics & More Article Recommendations *

sı Supporting Information
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

ABSTRACT: The optimization of the membrane electrode assembly (MEA) is crucial for
enhancing the performance of proton exchange membrane water electrolysis. Nevertheless,
Downloaded via ZHEJIANG UNIV on February 13, 2025 at 09:29:48 (UTC).

achieving global optimization of all manufacturing parameters of the MEA poses challenges
due to their high-dimensional complexity and limited experimental data. In this study,
machine learning (ML) techniques were introduced to tackle this intricate engineering
challenge. 58 MEAs were fabricated and tested to construct a comprehensive database
enriched with features and ample data. This was achieved through a data expansion
method that involves altering the operating temperature of the electrolyzer. The XGBoost
was employed to perform regression predictions on high-dimensional variables, achieving a
remarkable coefficient of determination (R2) value of 0.99926. The SHAP (SHapley
Additive exPlanations) method and the genetic algorithm were applied for model
interpretation and global optimization, respectively. By utilizing the insights provided by
the SHAP method, we could narrow the decision variable dimensionality down to 5 key
variables, achieving results that are comparable to full-variable optimization while notably
reducing time costs by 67.9%. Guided by ML, the MEA with globally optimized variables achieved a voltage of only 1.828 V at 3 A
cm−2. The study presents an approach that integrates intelligent optimization techniques with data-driven methods for high-
dimensional variable optimization. This contribution provides valuable insights into energy conversion and storage technologies in
the chemical industry.

1. INTRODUCTION mulation,12−16 porosity of the PTL,17−19 and MEA config-

Hydrogen energy, a promising and environmentally friendly uration.10 In this case, it is necessary to conduct global
optimization of all variables and high-dimensional analysis to
power source, has garnered considerable attention.1,2 The
fully realize the performance potential of the MEA.
production of green hydrogen through water electrolysis is the
Unfortunately, due to the limitations of the human brain’s
foundation of the hydrogen energy industry chain.3 Proton
data processing capabilities, current research can only focus on
exchange membrane water electrolysis (PEMWE) stands out
optimizing individual variables through controlled experiments.
as an environmentally friendly, high-efficiency hydrogen
In recent years, machine learning (ML), a powerful artificial
production method.4,5 Comparing with other hydrogen
intelligence (AI) technique, has become increasingly relevant
production technic like alkaline water electrolysis and high-
in the energy sector. This approach encompasses various tasks
temperature solid oxide water electrolysis, PEMWE offers
such as data collection and comprehensive databases establish-
advantages including flexible operating temperatures, enhanced
ment, model selection, hyperparameters tuning, and model’s
hydrogen purity, rapid response, superior controllability, and
application. By employing this methodology, researchers can
heightened safety.6 The membrane electrode assembly (MEA),
effectively address complex system challenges while gaining
composed of a proton exchange membrane, catalytic layers profound insights into the subject matter.
(CLs), and porous transport layers (PTLs), serves as the During the data preparation stage for training, the
central components in PEMWE. Numerous complex phys- experiment-based methods were used to gather a large amount
icochemical processes occur on the MEA that directly affect
the electrolysis efficiency, including electrochemical reactions,
electron and proton transfer, as well as gas−liquid transport. Received: October 9, 2023
Therefore, regulating the structure and composition of the Revised: December 25, 2023
MEA to optimize these processes is crucial for improving its Accepted: December 28, 2023
performance. Published: January 11, 2024
In practice, the performance of the MEA is affected by
different variables, including catalyst loading,7−11 ink for-

© 2024 American Chemical Society https://doi.org/10.1021/acs.iecr.3c03546

1409 Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

Figure 1. Illustration of the ML workflow.

of reliable data, providing a strong foundation for further contrast to a comprehensive variable optimization, this method
optimization.20−22 Combining Sobol sampling with multilayer uses only 5 key features selected through the SHAP
Artificial Neural Networks (ANN) for the modeling has method29−31 as decision variables. It yielded results com-
significantly reduced sample and computation resources, parable to those of full-variable optimization while substantially
increased optimization speed 10-fold, improved predictive improving time efficiency. The optimal variables for
accuracy, and effectively prevented model overfitting.23 For manufacturing MEA were determined, achieving an electrolysis
tackle complex tasks, it is beneficial to evaluate a variety of voltage of merely 1.828 V, with a current density of 3 A cm−2.
models and select the one that best suit the specific
requirements of the task.20,24,25 Afterward, the focus shifts to 2. EXPERIMENT
the pivotal role of hyperparameters in enhancing the model
2.1. Preparation of Membrane Electrode Assemblies.
performance. Recent studies have demonstrated the effective-
The sprayer (XF400 from Xifeng, China) was utilized to coat
ness of multiobjective optimization algorithms in hyper-
the membrane electrode assemblies (MEAs) with an active
parameter tuning. This method involves training surrogate
area measuring 2.5 × 2.5 cm2. This process resulted in the
models over a broad range, which improves performance and
creation of four identical MEAs following each spraying
simplifies the model, ultimately reducing computational procedure. The spraying area was set at 8 × 8 cm2 and
costs.23,26,27 In research aimed at optimization-oriented tasks, oversprayed to ensure that there was no loading loss at the
various optimization algorithms such as differential evolution, edge of the CL. The sprayer nozzle was precalibrated to
particle swarm optimization, and genetic algorithms (GAs) produce a uniform, fine spray. For the formation of a flat
have been widely studied and implemented for model catalysis layer (CL), the proton exchange membrane was
application.20,24,25 Recent advancements have highlighted the vacuumed and heated before spray ink was applied.
introduction of an online multiobjective Bayesian optimization Platinum catalysts supported on Vulcan XC72 carbon (60
approach. This approach has effectively reduced the calls for wt % platinum, TEC10E60TPM, and 20 wt % platinum,
high-fidelity model utilization in the optimization of crystal- TEC10E20E, both from Tanaka Kikinzoku Kogyo, Japan) and
lization processes, significantly enhancing the computational IrO2 catalyst supported on TiO2 (IrO2/TiO2 with 75 wt %
efficiency.28 This development has significantly contributed to Iridium; Elyst Ir75 0480 from Umicore, Germany) were
the application of data-driven models in the field of chemical utilized for the hydrogen evolution reaction (HER) and
engineering. oxygen evolution reaction (OER). The catalyst and ionomer
In this paper, we focused on high-dimensional analysis of the (5 wt %, D520 from Chemours, USA) were dispersed in
MEA fabrication process to optimize its performance via ML deionized (DI) water and isopropyl alcohol. After an
(Figure 1). First, we executed high-dimensional orthogonal ultrasonication for 10 min in an ice bath, the prepared ink
experiments to optimize the variables involved in MEA was loaded into a syringe pump, and the ink was sprayed onto
fabrication. Optimal ML performance relies on training with solid electrolytes or PTL of varying thicknesses using different
ample high-quality data. However, gathering such extensive technical parameters for catalyst-coated membrane (CCM)
data within a laboratory context is a costly endeavor. For this configuration or porous transport electrode (PTE) config-
purpose, temperature variations were introduced to enrich the uration. The catalyst loading was determined by the number of
data set size. Second, a data set was constructed using 58 spraying cycles and the ink flow rate.
MEAs, comprising 16 input features and 11,025 experimental N115, N117, and NR212 (Chemours, USA) were utilized as
data points. This data set was then employed for the regression proton exchange membranes. Initially, N115 and N117 were
model to develop a high-performance model. Third, the SHAP boiled in a 5.0 wt % H2O2 solution at 80 °C for 1 h.
(SHapley Additive exPlanations) interpretation method was Subsequently, the two samples were immersed and boiled in a
utilized for qualitative and quantitative analyses of the ML 0.50 M H2SO4 solution at the same temperature for 1 h. At
model. The method enhanced our understanding of how last, the samples were stored in DI water.
different variables influence MEA performance. Subsequently, Sintered fiber titanium felt with a thickness of 250 μm
we applied the GA to optimize the electrolytic voltage. In (2GDL10-0.25 and 2GDL08-0.25, both from Bekaert,
1410 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

Figure 2. Voltage variation for the as-prepared MEAs with different (a) water content and (b) ionomer content in the dispersion.

Belgium) and carbon paper, also with a thickness of 250 μm net regression (ENR). The ML models and optimization
and a microporous transport layer (H24CX483 from algorithms were implemented in the Jupyter Notebook using
Freudenberg, Germany), were employed as the PTLs for the Python. Data processing, model training, and optimization
anode and cathode, respectively. The anode PTL was algorithms were implemented using several open-source
ultrasonically cleaned in acetone, ethanol, and DI water for libraries, including numpy 1.23.5, pandas 1.5.3, scikit-learn
30, 20, and 10 min, respectively, and then air-dried at room 1.2.1, matplotlib 3.7.0, shap 0.41.0, xgboost 1.7.6, and
temperature for future experiments. The PTL or PTE was geneticalgorithm 1.0.2, pyswarms 1.3.1, scipy 1.11.3, installed
securely bonded to the CL or proton exchange membrane using Anaconda 3. First, the data sets were randomly assigned
through a hot-pressing process at a temperature of 140 °C. as training, validation, and test sets in the ratio of 0.6:0.2:0.2,
The final 5-in-1 assembly consisted of the MEA sandwiched and then the best model parameters were determined by grid
between two serpentine flow field plates and sealed with PTFE search and fourfold cross-validation. The coefficient of
gaskets. The 5-in-1 MEA production process is shown in determination (R2), root-mean-square error (RMSE), and
Figure S1a. mean absolute error (MAE) were used for the comprehensive
2.2. Electrochemical Performance Testing of Mem- evaluation of the 8 machine modeling algorithms. R2, RMSE,
brane Electrode Assemblies. A homemade test station and MAE were used to comprehensively evaluate the 8
equipped with a DC power supply (16 V/250 A/1.8 kW, machine modeling algorithms. Subsequently, we established a
N3618-016-250 from NGI, China) was utilized to measure the GA to optimize the performance of MEA. The objective
I−V polarization curves. The activation and testing procedures function used in the optimization process of GA was defined as
of the electrolyzer were by our previous study.32 In this the predicted voltage of the ML model at 3 A cm−2. The
procedure, the anode of the electrolyzer was supplied with 80 optimization algorithm is governed by several key parameter
°C DI water for 8 h, facilitated by a peristaltic pump. settings, including num_iteration, population_size, mutation_-
Subsequently, the electrolysis voltage was set to 1.6 V, and the probability, elit_ratio, crossover_probability, parents_portion,
operating current was continuously monitored until it and crossover_type (Supporting Note 1). These parameters
stabilized. The polarization curves were then recorded using are fine-tuned to minimize the objective function. Ultimately,
the galvanostatic method with step durations of 0.5 min. The we obtained optimal performance results and corresponding
temperature change was realized by employing a dual feature values through the GA, which were then outputted in
mechanism comprising a preheater and a thermostatic water the Jupyter Notebook.
bath.
Electrochemical impedance spectroscopy (EIS) and cyclic 3. RESULT AND DISCUSSION
voltammogram (CV) were performed with the electrochemical 3.1. Collecting and Analyzing Data. The performance of
workstation (Zennium Pro from Zahner, Germany). After each membrane electrode assemblies (MEAs, 5-in-1) is affected by
polarization step, EIS measurements were conducted at a their preparation, assembly, and testing conditions, where
frequency range of 100 kHz to 100 mHz, using the dozens of variables are involved. Among them, the following 16
galvanostatic mode at 2.0 A. To ensure a suitable signal-to- key variables were selected as input features of the modeling, as
noise ratio while preserving the linear system response, the shown in Figure S1b. Ink formulation is the first and
amplitude of the current perturbation was set to 0.2 A. Before fundamental step in MEAs preparation. The ink consists of
conducting the CV test, the electrolyzer’s anode was flushed catalyst, ionomer, and dispersant (water and alcohols), and in
with humid N2, while the cathode was placed in a H2-saturated this investigation, the water content and ionomer content were
environment, serving as both the reference and counter chosen as two key variables. Second, ultrasonic spraying, as a
electrodes. The CV test was conducted at a scan rate of 50 commonly used technique for laboratory-scale investigations,
mV/s in the potential range of 0−1.40 V. was utilized here to prepare CLs. The following variables were
2.3. Machine Learning Modeling. 8 popular ML models selected: vacuum-heated plate temperature, ink flow rate,
for small data sets were trained on the same computing cluster. cathode and anode catalyst loadings, MEA configuration, and
The Supporting Information contains brief introductions and proton exchange membrane thickness. Third, we evaluated
code implementations of the following applied algorithms: how different assembly variables affect MEAs performance,
artificial neural network (ANN), XGBoost, AdaBoost, K- including the Ti felt porosity, whether it is platinum-coated,
nearest neighbors (KNN), random forest (RF), Lasso hot press duration, hot press pressure, and compression rate of
regression (LR), support vector regression (SVR), and elastic the PTL. Fourth, the impact of operating variables, including
1411 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

Figure 3. Voltage variation for the as-prepared MEAs with different (a) vacuum heating plate temperature, (b) ink flow rate of the sprayer, (c)
cathode catalyst loading, (d) anode catalyst loading, (e) MEAs configuration, and (f) types of the proton exchange membrane.

DI-water flow rates and operating temperatures, was proton conductivity and active sites accessibility.14,16 Yet high
investigated. In addition to the above 15 variables, the current water content (>20 wt %) in the ink causes localized swelling
density in the I−V curves was regarded as another input of the proton exchange membrane, which results in the
feature. For each of these 16 variables, a set of experiments formation of defects in the CL and leads to activation and mass
were designed to determine how it affects the final perform- transport losses.12 Therefore, there is an optimal water content
ance. Ultimately, a comprehensive data set was compiled that that minimizes the electrolysis voltage.
contains extensive information and a variety of characteristics. Ionomer plays the role of binding catalyst particles,
Refer to Table S1 for detailed MEAs fabrication variables, stabilizing catalyst ink, and conducting protons. The
respectively. electrolysis voltage and polarization curve for the MEAs (No.
3.1.1. Ink Preparation. Water is incorporated into the 6−34 and 4) with different ionomer contents are shown in
dispersion to diminish the contact concentration between the Figures 2b and S2b. The performance of the MEAs improves
catalyst and the organic solution by employing water’s dilution as the ionomer content increases from 5 to 10 wt % then
effect, which can reduce the risk of deflagration.20 Figures 2a declines as it further increases to 25 wt %. At low Nafion
and S2a illustrate how the water content affects the MEAs’ content (5 wt %), the effective proton conduction in CL is
performance. The electrolysis voltage of the MEAs (Nos. 1−5) hampered due to the absence of a continuous ionomer
descends as the water content rises from 5 to 20 wt % and network. With an increasing ionomer content, the proton
subsequently increases with the water content (20 to 60 wt %). transport channels gradually form, while the resistance to
Introducing water into the ink helps unfold the hydrophilic electron transport rises. Further increasing the ionomer
sulfonic acid side chains in perfluorosulfonic acid (PFSA) and content (>20 wt %) not only deteriorates the electron
make the Nafion clusters diminish,15 which is beneficial for the conduction but also results in ionomer volume surpassing
1412 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

Figure 4. Voltage variation for the as-prepared MEAs with different (a) anode PTL porosity, (b) PTL coated/uncoated with Pt, (c) hot-pressing
time, (d) hot-pressing pressure, and (e) cathode PTL compression rate.

that of pores, which could negatively affect the gas/water facilitates the rapider curing of the CL, forming a more
transport in the CL.33 The Nyquist plot and CV of different homogeneous structure. It should be noted that temperatures
ionomer contents are shown in Figure S3. Therefore, the exceeding 90 °C can result in irreversible aging or degradation
ionomer content must strike a balance among electron and of the proton exchange membrane. To mitigate the risk, the
proton conduction and mass transport. upper limit of temperature operation was set to 90 °C.
3.1.2. MEA Spraying. The temperature on the vacuum The ink flow rate during the spraying process is another
platform could prompt evaporation of the dispersion and factor that affects the performance of the MEAs. Adjusting the
subsequent drying and curing of catalyst particles. Modifying ink flow rate affects the drying rate of the dispersion, which, as
the temperature of the vacuum heating plate impacts the was discussed earlier, results in the structural changes of the
surface characteristics of CL, which could alter the perform- CL.12 Figures 3b and S4b illustrate that the performance of the
ance of the MEAs. Figures 3a and S4a depict the MEAs (No. MEAs (No. 39−42 and 4) improves as the flow rate decreases
35−38 and 4) performance prepared by various heating plates from 4 to 1 mL min−1. Further reduction in the ink flow rate to
(50−90 °C). The MEA performance gradually improves as the 0.5 mL min−1 leads to little performance improvements,
temperature of the heating plate increases. It is reported that suggesting that the dispersion dries rapidly at a rate of 1 mL
reduced heating plate temperatures extend the ink dwell time min−1. Taking the preparation efficiency and optimal perform-
on the proton exchange membrane, resulting in the membrane ance of the as-prepared MEAs into account, we set a flow rate
swelling and catalyst particle agglomeration.12 The inhomoge- of 1 mL min−1 was set.
neous distribution of CL impacts both charge and substance We sequentially investigated the influence of Pt (cathode)
transport efficiency, causing an elevation in ohmic and mass and Ir (anode) loading on the performance of MEAs. We
transfer impedance. On the contrary, higher temperature evaluated two MEAs with Pt loading of 0.40 and 0.13 mgPt
1413 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

Figure 5. Polarization curves of the as-prepared MEAs at different (a) DI-water flow and (b) operating temperature.

cm−2. To maintain a consistent thickness of the CL (Figure 3.1.3. Electrolyzer Assembly. The assembly of the MEA is a
S5), Pt/C catalysts with differing Pt compositions (60 and 20 complex process that requires a comprehensive assessment of
wt %) were employed. As shown in Figure 3c, The electrolysis various factors to optimize both interfacial contact impedance
voltage difference between the two MEAs (No. 43 and 4) is and the mass transport process. The PTL in the 5-in-1 MEA
merely 4 mV at 3 A cm−2. This indicates that due to the faster primarily handles gas−liquid management and facilitates
reaction rate of the HER, the Pt loading on the cathode has electron conduction. Figures 4a and S7a illustrate a
little effect on the device performance. On the other hand, as comparison of performance between two types of Ti felts,
illustrated in Figures 3d and S4d, the anode loading (No. 6−34 one with 77% porosity and the other with 56% porosity (No.
and 4) significantly influences the electrolysis performance. 47 and 4), and the former one exhibits superior performance.
Increasing the Ir loading results in a decrease in electrolysis The catalyst in the CL buried beneath the PTL fibers is
voltage at the low current density region. However, in the high underutilized,17 which may explain the inferior performance of
current density region, the electrolysis voltage slightly rises PTLs with low porosity. Moreover, Ti felts with decreased
when the catalyst loading increases from 2.0 to 3.0 mgIr cm−2. porosity could hinder rapid gas transport due to their reduced
CL grows thicker with more catalyst loading (Figure S6). channel density, increasing the mass transport impedance.
Excessive thickness of CL increases gas diffusion path length, Figures 4b and S7b illustrate that the application of platinum
which elevates the mass transfer impedance. This results in plating on Ti felts leads to a substantial reduction in
performance degradation at high current density regions. Thus, electrolysis voltage (No. 48 and 4). Exposure of Ti leads to
optimizing the performance of the MEA requires striking a the formation of a thin, insulated layer of TiO2, which could
delicate equilibrium between activation polarization and mass elevate the contact resistance of the PTL. Thus, employing Pt
transfer polarization. as a protective layer can significantly enhance the electrolyzer’s
The configuration should also be considered during the performance.
The hot-pressing technique involves the integration of the
preparation of MEAs. The CCM configuration involves
PTL (or PTE) onto the CCM (or proton exchange
directly applying the catalyst onto the surface of a proton
membrane) through both mechanical pressure and heat to
exchange membrane, whereas the PTE configuration distrib-
enhance their contact.34−38 The hot press platform temper-
utes the catalyst onto the PTL. As illustrated in Figure 3e, the
ature has been set at 140 °C, which is above the Nafion glass
performance of the MEA with PTE configuration is inferior to transition temperature (Tg 126 °C) yet well below its viscous
that with CCM configuration at equivalent porosity (No. 44 flow temperature (Tf 180 °C).39 The SEM cross-sections of
and 4). Their performance variance mainly comes from their those MEAs are shown in Figure S8. Results from Figures 4c
different effective contact area between the membrane and the and S7c indicate that prolonged hot-pressing time could
catalyst. For the PTE configuration, some of the catalyst enhance the MEAs performance at low- and medium-current
particles fall into the PTL’s holes and lose contact with the density regions (No. 49−52 and 4). Excessive hot-pressing
proton exchange membrane, resulting in reduced catalyst time (900 s) causes a voltage rise in the high-current density
utilization and increased contact resistance. In contrast, the region due to increased PTL’s compaction, leading to elevated
CCM configuration offers a tighter contact between the gas diffusion resistance. The performance improvement is
catalyst and membrane. Therefore, in this work, the CCM more significant with increased hot-pressing pressure com-
configuration was adopted. pared to hot-pressing time (No. 53−56 and 4), as depicted in
The proton exchange membrane serves as a separator, Figures 4d and S7d. The enhancement in performance is
preventing electric conduction and facilitating proton trans- ascribed to the improved contact between the CCM and PTL,
port. Thinner membranes show lower H+ transport impedance. resulting from increased pressure.
The data presented in Figures 3f and S4f demonstrate that the The compression rate during the assembly of the MEA can
MEA performance worsens as the proton exchange membrane affect the performance of PEMWE. Since the PTFE gaskets are
grows thicker (NR212, N115, and N117, No. 45−46 and 4). highly rigid, the PTFE thickness is adjusted to change the
However, reduced membrane thickness could compromise the compression rate.40 Figures 4e and S7e illustrate that
lifetime and safety of MEAs. Therefore, to balance safety and increasing the cathode PTL compression rate (No. 57−58
performance, Nafion 115 was adopted as the proton exchange and 4) reduces the electrolysis voltage at low- and medium-
membrane in other experiments. current density regions due to the reduced contact impedance
1414 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

of PTL with the flow-field plate and the CL. However, Table 1. Overview of the 16 Input Features and
increasing the compression ratio leads to the collapse of the Corresponding Ranges in the Experimental Section
ideal gas transport pathway within the PTL41,42 and hinders
input feature features value range
mass transfer, which is why at a high-current density region, a
higher compression ratio results in a rise in voltage. Hence, heating plate temperature 50−90 (°C)
achieving a balance between the contact impedance and mass water content 5−60 (wt %)
transfer impedance is crucial when selecting the compression ink flow rate 0.5−4.0 (mL min−1)
ratio. hot pressing time 300−900 (s)
3.1.4. Electrolyzer Operation. The performance and hot pressing pressure 350−550 (kg)
efficiency of PEMWE depend on the operating variables, as cathode compression rate 8−40 (%)
well. The DI water flow rate could influence the transport of anode PTL porosity 56% (0) or 77% (1)
water and gas within the electrolyzer. The water stoichiometry MEA configuration PTE (0) or CCM (1)
ratio (ξ) is defined as the ratio between the circulating water titanium felt with or without platinum plating uncoated Ti felt (0)
volume and the water consumption at 3 A cm−2, which coated Ti felt (1)
evaluates the sufficiency of the water supply.43 The electrolysis membrane thickness 52−178 (μm)
cathode catalyst loading 0.13−0.40 (mgPt cm−2)
voltage remains relatively constant despite varying the inlet DI-
anode catalyst loading 0.5−3.0 (mgIr cm−2)
water flow rate over a wide range of 0.5−20 mL min−1, viz.
anode ionomer content 5−25 (wt %)
4.76 ≤ ξ ≤ 190.47 (No. 4 in the blue background), as
operating temperature 30−90 (°C)
illustrated in Figure 5a. In detail, the electrolysis voltage
deionized water flow rate 0.5−20.0 (mL min−1)
decreased by merely 5 mV when the flow rate dropped from 20
current density 0.05−3.0 (A cm−2)
to 0.5 mL min−1, even at a high current density of 3 A cm−2.
This demonstrates that with enough water supply, changes in
the water flow rate have little effect on the performance. A they fail to adequately capture the key features of the data. The
water flow rate below 0.5 mL min−1 leads to an inadequate present study employed a varied selection of eight widely
water supply, which disrupts proton transport in the recognized ML algorithms, specifically: ANN, XGBoost,
membrane, causes a substantial voltage surge, and even AdaBoost, KNN, RF, LR, SVR, and ENR. Detailed algorithm
damages the MEA.43 However, this study does not center on profiles along with their corresponding code compiled by us
water starvation. can be found in Supporting Note 2 and the GitHub
Figure 5b displays the I−V curves of the electrolyzer at repository.46 The regression model utilized the 16 variables
operating temperatures ranging from 30 to 90 °C (No. 4 in mentioned above as input features, with the voltage serving as
green background). An improvement in the electrolyzer the output target. The database was homogenized and
performance is observed with an increase in temperature, randomly divided into training, validation, and test sets in a
which is a well-known trend. This set of experiments serves ratio of 0.6:0.2:0.2. The best-performing hyperparameters on
mainly to supplement the data set and ensure the accuracy of the validation set were chosen through a grid search and
model establishment in subsequent ML procedures (see fourfold cross-validation to mitigate model underfitting and
Section 3.2). overfitting. The model’s generalization ability to unknown data
3.2. Establishing of the Database. Access to a sufficient was assessed by predicting the test set. Scatter plots showing
amount of data is essential for improving the performance and the model’s predictions of electrolysis voltage on the test and
accuracy of algorithmic models in ML.44 Collecting a training sets are presented in Figure 6 a−h, where the x-axis
significant volume of data in controlled laboratory settings values represent the ML prediction data, and the y-axis values
becomes unfeasible due to the significant expenditures of time represent the real data from the database. Three objective
and resources involved. To augment the data set for fabrication evaluation metrics (Supporting Note 3), viz. RMSE, R-squared
variables, all previously mentioned MEAs were tested at (R2), and MAE, were employed to assess the performance and
temperatures ranging from 30 to 90 °C. The resultant accuracy of these algorithms comprehensively. Examination of
polarization curves are displayed in Figures S9−S23. The Figure 6 indicates that XGBoost outperforms alternative
current density is employed as an input feature due to its direct models, displaying a substantial R2 value of 0.99926, as well
correspondence with the voltage in the polarization curve. In as markedly reduced RMSE and MAE values. XGBoost, an
addition, the variables discussed above are transformed into integrated algorithm, excels in effectively combining predic-
quantitative parameters, which can be computationally tions from numerous weak learners, leading to a substantial
processed. Table 1 displays the pertinent characteristics enhancement in overall prediction accuracy compared with
along with their respective ranges of values. singular models. Consequently, given its victory in this
Finally, we have developed a comprehensive and data-rich competition, XGBoost will be employed for both ranking the
database comprising 16 input features, 441 polarization curves, importance of features and optimizing performance.
and 11,025 data points using only 58 MEAs. With sufficient 3.4. Model Explanation. To enhance our understanding
data collected, the subsequent step involved constructing and of XGBoost, the advanced ensemble learning algorithm often
training the ML model. viewed as a “black box” model, we utilized the SHAP method
3.3. Algorithm Selection. Presently, the field of ML (Supporting Note 4), which clarifies the model’s assessment of
boasts a diverse array of regression algorithms. According to each feature and offers insights into MEA’s performance. The
the no free lunch (NFL) theorem,45 no single algorithm is Global Feature Plot (GFP) is a tool that employs SHAP values
suitable for every problem. Using complex models for simple to rank features by their effect on predictions. It calculates the
tasks can lead to overfitting, where the model overadjusts to SHAP value for each feature, indicating its contribution to the
the training data and loses generalizability. Conversely, for model’s prediction. The features are then ranked based on the
complex tasks, simple models may result in underfitting, as absolute average of their SHAP values. Figure 7a displays the
1415 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

Figure 6. (a−h) Plot of predicted data vs actual data on the test and training set for 8 ML regression algorithms; (i) summary of performance
evaluation from 8 ML regression algorithms.

GFP of the 15 features, excluding the current density. This influence electrolysis voltage. The result aligns strongly with
visualization effectively showcases the combined impact of the prior knowledge discussed in Section 3.1, indicating that
these features on the predictive outcomes of the model. In the the model has a certain extent of reliability. Although the GFP
GFP, the color gradation from blue to red represents a offered a comprehensive view of feature importance, it has
progressive increase in feature values. Positive SHAP values limitations in capturing the specific impact of individual
indicate that a feature’s value has a positive influence on the features on the predictions. To enhance our understanding of
output variable, while negative SHAP values signify an inverse the nonlinear connection between a specific feature and the
relationship. For instance, higher temperature values (red target variable, we also introduce partial dependency plots
color) are associated with a decrease in voltage (negative (PDP). In the PDP, the values of the other features are fixed to
values) while the thicker membranes (red color) increase the focus on the target feature, and the trend of the change of its
voltage (positive values). The study revealed that operation value on the model prediction is observed. Figures 7b−f and
temperature, anode ionomer content, anode catalyst loading, S24 show the PDP of various features. The gradation from
membrane thickness, and hot pressing pressure significantly yellow to blue is purely a visual representation of the increase
1416 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

Figure 7. SHAP explanation for input features: (a) the Global Feature Plot of 15 variables; the partial dependency plot of (b) anode ionomer
content, (c) anode catalyst loading, (d) water content, (e) hot pressing time, and (f) cathode compression rate.

Figure 8. (a) Objective function decreases with the iteration number in GA and corresponding optimized variables to fabricate N115-based MEA;
(b) ML predicted and experimental polarization curves for optimized N115-based MEA.

in SHAP values. This color transition intuitively denotes an then the optimal loading is chosen to be 2.5 mgIr cm−2. This
increase in predictive voltage values, highlighting their discrete-variable-based optimization point selection method
augmented influence on the model’s predictions. Notably, a significantly lacks precision and scientific validity. Conse-
distinct “trough” in SHAP values is observed across five quently, an advanced optimization algorithm is employed to
specific features, viz. anode ionomer content, anode catalyst carry out the enhancement of MEA performance.
loading, water content, hot pressing time, and cathode PTL 3.5. Performance Optimization. In this section, we
compression rate. For instance, in the case of anode catalyst employed five classical algorithms, including the particle swarm
loading (Figure 7c), there must be an optimal point that optimization algorithm, differential evolutionary algorithm,
effectively balances the factors of mass transfer and activation Bayesian optimization algorithm, simulated annealing algo-
polarization. In fact, if the optimization of catalyst loading is rithm, and GA, to optimize the performance of the MEA. The
performed according to the crossover experiment protocol, objective function, decision variables, and constraints are
1417 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

presented in Table S3. The results, including the minimum R2 value of 0.99926, and then it is employed for the predictive
voltage and time cost, obtained from these algorithms are analysis of electrolytic voltage. In the process of minimizing the
presented in Figure S25. Considering the optimization electrolytic voltage, we utilized the GA to optimize the
algorithms’ capacity to identify a global optimal solution and performance of the MEA while considering both time
their efficiency, we have selected the GA as our optimization efficiency and the capability to find optimal solutions. Instead
strategy, achieving a minimized electrolysis voltage of 1.83107 of optimizing all variables, we optimized performance using
V. five key variables identified through SHAP analysis. This
In the previous section, GFP and PDP offer guidance for
approach yielded comparable results to those achieved through
optimizing variables. The five variables, shown in Figure 7b−f,
the optimization of all variables while concurrently improving
viz. anode ionomer content, anode catalyst loading, cathode
compression, hot pressing time, and water content, are more time efficiency by 67.9%. Guided by ML, the MEA we
suitable for optimization compared to others. The remaining fabricated exhibited an electrolytic voltage of 1.828 V at 3 A
variables minimally impact electrolyzer performance and can cm−2, closely aligning with the predicted value, with an R2
be addressed with human brain knowledge, as discussed above. value of 0.998116.
To assess the impact of reducing variable dimensions based on The success of this study extends beyond PEMWE
the SHAP explanation method on the performance and optimization. It presents a novel research approach to tackle
efficiency of GA optimization, we conducted optimization complex engineering optimization issues. We firmly believe
experiments focusing on the aforementioned five key features. that with the ongoing advancement of artificial intelligence,
As depicted in Figure 8a, the objective function exhibits a ML can persistently facilitate significant breakthroughs and
diminishing trend as iterations increase, ultimately stabilizing innovations in the energy sector.
at the minimum voltage. Simultaneously, the GA algorithm
model provides output for the relevant fabrication parameters
(the table in Figure 8a) and the predicted MEA performance
(Figure 8b). Indeed, focusing on five key variables instead of
■
*
ASSOCIATED CONTENT
sı Supporting Information

optimizing across all variables can yield comparable variable The Supporting Information is available free of charge at
outcomes and objective functions, while achieving a significant https://pubs.acs.org/doi/10.1021/acs.iecr.3c03546.
improvement in computational efficiency by 67.9% (Table S3).
It is notable that when dealing with challenges related to high- Description of parameters of the genetic algorithm,
dimensional data, certain model interpretation methods (e.g., advantages and disadvantages of 8 machine learning
SHAP) can be valuable in reducing dimensionality and algorithms, evaluation metrics for machine learning
mitigating computational costs. These methods provide algorithms, working principle of SHAP; membrane
insights into the importance and impact of individual features, electrode assembly, the process of data collection; I−V
enabling the identification of relevant factors and simplifying curves of different water content, anode ionomer
the analysis of complex, high-dimensional data, as discussed content, vacuum heating plate temperature, ink flow
above. rate of the sprayer, cathode catalyst loading, anode
Subsequently, we fabricated the MEA following the GA’s catalyst loading, MEAs configuration, types of proton
outputs (the table in Figure 8a). Analysis of Figure 8b reveals exchange membrane, anode PTL porosity, hot-pressing
an impressive RMSE and MAE, measuring only 0.004983 and time, hot-pressing pressure, cathode PTL compression
0.005128 against the predicted value. An electrolysis voltage of rate, operating temperature; cross-sectional SEM images
only 1.828 V is required to achieve 3 A cm−2 for the optimal of MEA; Nyquist plot, CV, voltammetric charge of
MEA, closely matching the 1.82923 V predicted by the ML different ionomer content; partial dependency plot for
model. To the best of our knowledge, the performance of the heating plate temperature, ink flow rate, cathode catalyst
MEA in this study surpasses that of other MEAs employing loading, membrane thickness, MEA configuration, hot
Nafion 115 as the proton exchange membrane in recent pressing pressure, anode PTL porosity, Ti felt coated/
years8,12,47−62 (as shown in Table S4 and Figure S26). uncoated with Pt, operating temperature, ink flow rate;
Therefore, we believe that ML methods can efficiently analyze performance of optimization algorithms; MEA perform-
data with complex features and the optimization algorithm can ance comparison between our work and others’ work;
converge with a small number of attempts. summary of MEA fabrication variables; objective
functions, decision variables and constraints of opti-
4. CONCLUSIONS mization algorithms; optimization variables of using all
variables or key variables; summary of research work on
The optimization of the MEA for PEMWE has traditionally
performance in PEMWE (PDF)
relied on human scientists exploring limited data and
conducting orthogonal experiments. This study aims to
revolutionize this paradigm by employing ML algorithms for
optimizing MEA variables in high dimensions. In order to
expand the data set in a cost-effective way, we executed MEAs
■ AUTHOR INFORMATION
Corresponding Authors
testing spanning temperatures from 30 to 90 °C. This effort Aidong Tan − Institute of Energy Power Innovation, North
yielded a notable 7-fold expansion in the volume of China Electric Power University, Beijing 102206, P.R. China;
experimental data. Based on a limited quantity of MEAs, we Email: [email protected]
gathered a significant amount of data that serves as the Jianguo Liu − Institute of Energy Power Innovation, North
foundation for accurate predictions using ML algorithms. China Electric Power University, Beijing 102206, P.R.
Compared to other traditional ML algorithms, XGBoost China; orcid.org/0000-0002-9229-4936; Phone: +86
demonstrates superior predictive accuracy with an impressive 010 61771756; Email: [email protected]
1418 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

Authors electrodes for proton exchange membrane water electrolyzers. J.

Yipeng Zhang − Institute of Energy Power Innovation and Mater. Chem. A 2019, 7 (47), 26984−26995.
School of Energy Power and Mechanical Engineering, North (11) Bernt, M.; Siebel, A.; Gasteiger, H. A. Analysis of Voltage
China Electric Power University, Beijing 102206, P.R. China Losses in PEM Water Electrolyzers with Low Platinum Group Metal
Zhuolin Yuan − Institute of Energy Power Innovation, North Loadings. J. Electrochem. Soc. 2018, 165 (5), 305−314.
(12) Alia, S. M.; Reeves, K. S.; Baxter, J. S.; Cullen, D. A. The Impact
China Electric Power University, Beijing 102206, P.R. China
of Ink and Spray Variables on Catalyst Layer Properties, Electrolyzer
Kun Zhao − Institute of Energy Power Innovation, North Performance, and Electrolyzer Durability. J. Electrochem. Soc. 2020,
China Electric Power University, Beijing 102206, P.R. China 167 (14), 144512.
Xiaoyun Shi − Institute of Energy Power Innovation, North (13) Khandavalli, S.; Park, J. H.; Kariuki, N. N.; Zaccarine, S. F.;
China Electric Power University, Beijing 102206, P.R. China Pylypenko, S.; Myers, D. J.; Ulsh, M.; Mauger, S. A. Investigation of
Ping Liu − Institute of Energy Power Innovation, North China the Microstructure and Rheology of Iridium Oxide Catalyst Inks for
Electric Power University, Beijing 102206, P.R. China Low-Temperature Polymer Electrolyte Membrane Water Electro-
Complete contact information is available at: lyzers. ACS Appl. Mater. Interfaces. 2019, 11 (48), 45068−45079.
https://pubs.acs.org/10.1021/acs.iecr.3c03546 (14) Berlinger, S. A.; Dudenas, P. J.; Bird, A.; Chen, X.; Freychet, G.;
McCloskey, B. D.; Kusoglu, A.; Weber, A. Z. Impact of Dispersion
Solvent on Ionomer Thin Films and Membranes. ACS Appl. Polym.
Author Contributions
Mater. 2020, 2 (12), 5824−5834.
Y.Z. and A.T. contributed equally to this work. The manuscript (15) Berlinger, S. A.; McCloskey, B. D.; Weber, A. Z. Inherent
was written through the contributions of all authors. All Acidity of Perfluorosulfonic Acid Ionomer Dispersions and
authors have given approval to the final version of the Implications for Ink Aggregation. J. Phys. Chem. B 2018, 122 (31),
manuscript. 7790−7796.
Notes (16) Ngo, T. T.; Yu, T. L.; Lin, H.-L. Influence of the composition
of isopropyl alcohol/water mixture solvents in catalyst ink solutions
The authors declare no competing financial interest. on proton exchange membrane fuel cell performance. J. Power Sources.

■ ACKNOWLEDGMENTS
This work was supported by the National Key R&D Program
2013, 225, 293−303.
(17) Peng, X.; Satjaritanun, P.; Taie, Z.; Wiles, L.; Keane, A.;
Capuano, C.; Zenyuk, I. V.; Danilovic, N. Insights into Interfacial and
Bulk Transport Phenomena Affecting Proton Exchange Membrane
of China (2021YFB4000100) and the Fundamental Research
Water Electrolyzer Performance at Ultra-Low Iridium Loadings. Adv.
Funds for the Central Universities (2022MS061).
Sci. 2021, 8 (21), No. 2102950.

■ REFERENCES
(1) Buttler, A.; Spliethoff, H. Current status of water electrolysis for
(18) Schuler, T.; Ciccone, J. M.; Krentscher, B.; Marone, F.; Peter,
C.; Schmidt, T. J.; Büchi, F. N. Hierarchically Structured Porous
Transport Layers for Polymer Electrolyte Water Electrolysis. Adv.
energy storage, grid balancing and sector coupling via power-to-gas Energy Mater. 2020, 10 (2), No. 1903216.
and power-to-liquids: A review. Renew. Sustain. Energy Rev. 2018, 82, (19) Kulkarni, D.; Huynh, A.; Satjaritanun, P.; O’Brien, M.;
2440−2454. Shimpalee, S.; Parkinson, D.; Shevchenko, P.; DeCarlo, F.;
(2) Chi, J.; Yu, H. Water electrolysis based on renewable energy for Danilovic, N.; Ayers, K. E.; Capuano, C.; Zenyuk, I. V. Elucidating
hydrogen production. Chin. J. Catal. 2018, 39 (3), 390−394. effects of catalyst loadings and porous transport layer morphologies
(3) Chatenet, M.; Pollet, B. G.; Dekel, D. R.; Dionigi, F.; Deseure, J.; on operation of proton exchange membrane water electrolyzers. Appl.
Millet, P.; Braatz, R. D.; Bazant, M. Z.; Eikerling, M.; Staffell, I.; Catal., B 2022, 308, No. 121213.
Balcombe, P.; Shao-Horn, Y.; Schäfer, H. Water electrolysis: from (20) Ding, R.; Yin, W.; Cheng, G.; Chen, Y.; Wang, J.; Wang, X.;
textbook knowledge to the latest scientific strategies and industrial Han, M.; Zhang, T.; Cao, Y.; Zhao, H.; Wang, S.; Li, J.; Liu, J.
developments. Chem. Soc. Rev. 2022, 51 (11), 4583−4762. Effectively Increasing Pt Utilization Efficiency of the Membrane
(4) Shiva Kumar, S.; Lim, H. An overview of water electrolysis Electrode Assembly in Proton Exchange Membrane Fuel Cells
technologies for green hydrogen production. Energy Rep. 2022, 8, through Multiparameter Optimization Guided by Machine Learning.
13793−13813. ACS Appl. Mater. Interfaces. 2022, 14 (6), 8010−8024.
(5) Corti, H. R. Polymer electrolytes for low and high temperature (21) Ding, R.; Yin, W.; Cheng, G.; Chen, Y.; Wang, J.; Wang, R.;
PEM electrolyzers. Curr. Opin. Electrochem. 2022, 36, No. 101109. Rui, Z.; Li, J.; Liu, J. Boosting the optimization of membrane
(6) Feng, Q.; Yuan, X.; Liu, G.; Wei, B.; Zhang, Z.; Li, H.; Wang, H. electrode assembly in proton exchange membrane fuel cells guided by
A review of proton exchange membrane water electrolysis on explainable artificial intelligence. Energy AI. 2021, 5, No. 100098.
degradation mechanisms and mitigation strategies. J. Power Sources. (22) Zhao, X.; Luo, T.; Jin, H. Predicting Diffusion Coefficients of
2017, 366, 33−55. Binary and Ternary Supercritical Water Mixtures via Machine and
(7) Taie, Z.; Peng, X.; Kulkarni, D.; Zenyuk, I. V.; Weber, A. Z.;
Transfer Learning with Deep Neural Network. Ind. Eng. Chem. Res.
Hagen, C.; Danilovic, N. Pathway to Complete Energy Sector
2022, 61 (24), 8542−8550.
Decarbonization with Available Iridium Resources using Ultralow
(23) Miriyala, S. S.; Chowdhury, S.; Pujari, N. K.; Mitra, K.
Loaded Water Electrolyzers. ACS Appl. Mater. Interfaces. 2020, 12
(47), 52701−52712. Optimally designed Variational Autoencoders for Efficient Wind
(8) Rozain, C.; Mayousse, E.; Guillet, N.; Millet, P. Influence of Characteristics Modelling. IEEE SSCI 2020, 2869−2876.
iridium oxide loadings on the performance of PEM water electrolysis (24) Ebrahimi, A.; Tamnanloo, J.; Mousavi, S. H.; Soroodan
cells: Part I−Pure IrO2-based anodes. Appl. Catal., B 2016, 182, 153− Miandoab, E.; Hosseini, E.; Ghasemi, H.; Mozaffari, S. Discrete-
160. Continuous Genetic Algorithm for Designing a Mixed Refrigerant
(9) Siracusano, S.; Baglio, V.; Van Dijk, N.; Merlo, L.; Aricò, A. S. Cryogenic Process. Ind. Eng. Chem. Res. 2021, 60 (20), 7700−7713.
Enhanced performance and durability of low catalyst loading PEM (25) Ding, R.; Ding, Y.; Zhang, H.; Wang, R.; Xu, Z.; Liu, Y.; Yin,
water electrolyser based on a short-side chain perfluorosulfonic W.; Wang, J.; Li, J.; Liu, J. Applying machine learning to boost the
ionomer. Appl. Energy. 2017, 192, 477−489. development of high-performance membrane electrode assembly for
(10) Bühler, M.; Hegge, F.; Holzapfel, P.; Bierling, M.; Suermann, proton exchange membrane fuel cells. J. Mater. Chem. A 2021, 9 (11),
M.; Vierrath, S.; Thiele, S. Optimization of anodic porous transport 6841−6850.

1419 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

(26) Miriyala, S. S.; Mittal, P.; Majumdar, S.; Mitra, K. Comparative (43) Sun, S.; Xiao, Y.; Liang, D.; Shao, Z.; Yu, H.; Hou, M.; Yi, B.
study of surrogate approaches while optimizing computationally Behaviors of a proton exchange membrane electrolyzer under water
expensive reaction networks. Chem. Eng. Sci. 2016, 140, 44−61. starvation. RSC Adv. 2015, 5 (19), 14506−14513.
(27) Pujari, K. N.; Miriyala, S. S.; Mittal, P.; Mitra, K. Better wind (44) Ding, R.; Zhang, S.; Chen, Y.; Rui, Z.; Hua, K.; Wu, Y.; Li, X.;
forecasting using Evolutionary Neural Architecture search driven Duan, X.; Wang, X.; Li, J.; Liu, J. Application of Machine Learning in
Green Deep Learning. Expert Syst. Appl. 2023, 214, No. 119063. Optimizing Proton Exchange Membrane Fuel Cells: A Review. Energy
(28) Tadepalli, A.; Pujari, K. N.; Mitra, K. A crystallization case AI. 2022, 9, No. 100170.
study toward optimization of expensive to evaluate mathematical (45) Gomez, D.; Rojas, A. An Empirical Overview of the No Free
models using Bayesian approach. Mater. Manuf. Process. 2023, 38 Lunch Theorem and Its Effect on Real-World Machine Learning
(16), 2127−2134. Classification. Neural Comput. 2016, 28 (1), 216−228.
(29) Shi, F.; Lu, S.; Gu, J.; Lin, J.; Zhao, C.; You, X.; Lin, X. (46) https://github.com/MegeDreamer119/MyProject.git.
Modeling and Evaluation of the Permeate Flux in Forward Osmosis (47) Lyu, X.; Foster, J.; Rice, R.; Padgett, E.; Creel, E. B.; Li, J.; Yu,
Process with Machine Learning. Ind. Eng. Chem. Res. 2022, 61 (49), H.; Cullen, D. A.; Kariuki, N. N.; Park, J. H.; Myers, D. J.; Mauger, S.;
18045−18056. Bender, G.; Pylypenko, S.; Serov, A. Aging gracefully? Investigating
(30) Bhosale, H.; Pandya, M. A.; Chatterjee, I.; Mukunth, A.; iridium oxide ink’s impact on microstructure, catalyst/ionomer
Sureshkumar, B.; Rajagopalan, R. S.; Parlikkad, N. R.; Valadi, J. K. interface, and PEMWE performance. J. Power Sources. 2023, 581,
Development of Gradient Boosting Machines for Estimation of Total No. 233503.
and Dynamic Liquid Holdup in Trickle Bed Reactor. Ind. Eng. Chem. (48) Islam, J.; Kim, S.-K.; Rahman, M. M.; Thien, P. T.; Kim, M.-J.;
Res. 2023, 62 (45), 19161−19176. Cho, H.-S.; Lee, C.; Lee, J. H.; Lee, S. The effect of iridium content in
(31) Chen, Q.; Deng, J.; Luo, G. Micromixing Performance and boron carbide-supported iridium catalyst on the activity and stability
Residence Time Distribution in a Miniaturized Magnetic Reactor: of proton exchange membrane water electrolyzer. Mater. Today
Experimental Investigation and Machine Learning Modeling. Ind. Eng. Energy. 2023, 32, No. 101237.
Chem. Res. 2023, 62 (8), 3577−3591. (49) Zhou, H.; Meng, K.; Chen, W.; Chen, B. Exploratory research
(32) Tan, A.; Zhang, Y.; Shi, X.; Ju, C.; Liu, P.; Yang, T.; Liu, J. The on bubbles migration behavior and mass transfer capacity evaluation
poisoning effects of Ti-ion from porous transport layers on the of proton exchange membrane water electrolyzer based on a volume
membrane electrode assembly of proton exchange membrane water of fluid-coupled electrochemical model. Energy Convers. Manag. 2023,
electrolyzers. Chem. Eng. J. 2023, 471, No. 144624. 290, No. 117217.
(33) Bernt, M.; Gasteiger, H. A. Influence of Ionomer Content in (50) Zhao, C.; Yuan, S.; Cheng, X.; An, L.; Li, J.; Shen, S.; Yin, J.;
IrO2/TiO2 Electrodes on PEM Water Electrolyzer Performance. J. Yan, X.; Zhang, J. Effect of perfluorosulfonic acid ionomer in anode
Electrochem. Soc. 2016, 163 (11), 3179−3189. catalyst layer on proton exchange membrane water electrolyzer
(34) Shahgaldi, S.; Alaefour, I.; Li, X. Impact of manufacturing
performance. J. Power Sources. 2023, 580, No. 233413.
processes on proton exchange membrane fuel cell performance. Appl. (51) Kang, Z.; Chen, Y.; Wang, H.; Alia, S. M.; Pivovar, B. S.;
Energy. 2018, 225, 1022−1032.
Bender, G. Discovering and Demonstrating a Novel High-Performing
(35) Najafi Roudbari, M.; Ojani, R.; Raoof, J. B. Investigation of hot
2D-Patterned Electrode for Proton-Exchange Membrane Water
pressing parameters for manufacture of catalyst-coated membrane
Electrolysis Devices. ACS Appl. Mater. Interfaces. 2022, 14 (1),
electrode (CCME) for polymer electrolyte membrane fuel cells by
2335−2342.
response surface method. Energy. 2017, 140, 794−803.
̇ (52) Piñeiro García, A.; Perivoliotis, D.; Wu, X.; Gracia-Espino, E.
(36) Okur, O.; Iyigü n Karadağ, Ç .; Boyacı San, F. G.; Okumuş, E.;
Benchmarking Molybdenum-Based Materials as Cathode Electro-
Behmenyar, G. Optimization of parameters for hot-pressing
manufacture of membrane electrode assembly for PEM (polymer catalysts for Proton Exchange Membrane Water Electrolysis: Can
electrolyte membrane fuel cells) fuel cell. Energy. 2013, 57, 574−580. These Compete with Pt? ACS Sustain. Chem. Eng. 2023, 11 (20),
(37) Hasran, U. A.; Kamarudin, S. K.; Daud, W. R. W.; Majlis, B. Y.; 7641−7654.
Mohamad, A. B.; Kadhum, A. A. H.; Ahmad, M. M. Optimization of (53) Huynh, T. B. N.; Song, J.; Bae, H. E.; Kim, Y.; Dickey, M. D.;
hot pressing parameters in membrane electrode assembly fabrication Sung, Y. E.; Kim, M. J.; Kwon, O. J. Ir−Ru Electrocatalysts Embedded
by response surface method. Int. J. Hydrogen Energy. 2013, 38 (22), in N-Doped Carbon Matrix for Proton Exchange Membrane Water
9484−9493. Electrolysis. Adv. Funct. Mater. 2023, 33 (28), No. 2301999.
(38) Yazdanpour, M.; Esmaeilifar, A.; Rowshanzamir, S. Effects of (54) Torrero, J.; Morawietz, T.; García Sanchez, D.; Galyamin, D.;
hot pressing conditions on the performance of Nafion membranes Retuerto, M.; Martin-Diaconescu, V.; Rojas, S.; Alonso, J. A.; Gago, A.
coated by ink-jet printing of Pt/MWCNTs electrocatalyst for S.; Friedrich, K. A. High Performance and Durable Anode with 10-
PEMFCs. Int. J. Hydrogen Energy. 2012, 37 (15), 11290−11298. Fold Reduction of Iridium Loading for Proton Exchange Membrane
(39) Li, Y.; Wen, Q.; Qin, J.; Zou, S.; Ning, F.; Bai, C.; Pan, S.; Jin, Water Electrolysis. Adv. Energy Mater. 2023, 13 (23), No. 2204169.
H.; Xu, P.; Shen, M.; Song, Y.; Zhou, X. A high-efficient and low- (55) Kang, Z.; Wang, H.; Liu, Y.; Mo, J.; Wang, M.; Li, J.; Tian, X.
consumption nanoimprint method to prepare large-area and high- Exploring and understanding the internal voltage losses through
quality Nafion array for the ordered MEA of fuel cell. Chem. Eng. J. catalyst layers in proton exchange membrane water electrolysis
2023, 451, No. 138722. devices. Appl. Energy. 2022, 317, No. 119213.
(40) Stähler, M.; Stähler, A.; Scheepers, F.; Carmo, M.; Lehnert, W.; (56) Yang, G.; Mo, J.; Kang, Z.; List, F. A.; Green, J. B.; Babu, S. S.;
Stolten, D. Impact of porous transport layer compression on hydrogen Zhang, F.-Y. Additive manufactured bipolar plate for high-efficiency
permeation in PEM water electrolysis. Int. J. Hydrogen Energy. 2020, hydrogen production in proton exchange membrane electrolyzer cells.
45 (7), 4008−4014. Int. J. Hydrogen Energy. 2017, 42 (21), 14734−14740.
(41) Khetabi, E. M.; Bouziane, K.; Zamel, N.; François, X.; Meyer, (57) Hegge, F.; Lombeck, F.; Cruz Ortiz, E.; Bohn, L.; von Holst,
Y.; Candusso, D. Effects of mechanical compression on the M.; Kroschel, M.; Hübner, J.; Breitwieser, M.; Strasser, P.; Vierrath, S.
performance of polymer electrolyte fuel cells and analysis through Efficient and Stable Low Iridium Loaded Anodes for PEM Water
in-situ characterisation techniques - A review. J. Power Sources. 2019, Electrolysis Made Possible by Nanofiber Interlayers. ACS Appl. Energy
424, 8−26. Mater. 2020, 3 (9), 8276−8284.
(42) Martin, A.; Trinke, P.; Stähler, M.; Stähler, A.; Scheepers, F.; (58) Kang, Z.; Yang, G.; Mo, J.; Li, Y.; Yu, S.; Cullen, D. A.; Retterer,
Bensmann, B.; Carmo, M.; Lehnert, W.; Hanke-Rauschenbach, R. S. T.; Toops, T. J.; Bender, G.; Pivovar, B. S.; Green, J. B.; Zhang, F.-
The Effect of Cell Compression and Cathode Pressure on Hydrogen Y. Novel thin/tunable gas diffusion electrodes with ultra-low catalyst
Crossover in PEM Water Electrolysis. J. Electrochem. Soc. 2022, 169 loading for hydrogen evolution reactions in proton exchange
(1), No. 014502. membrane electrolyzer cells. Nano Energy. 2018, 47, 434−441.

1420 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421
Industrial & Engineering Chemistry Research pubs.acs.org/IECR Article

(59) Xie, Z.; Yu, S.; Yang, G.; Li, K.; Ding, L.; Wang, W.; Cullen, D.
A.; Meyer, H. M.; Retterer, S. T.; Wu, Z.; Sun, J.; Gao, P.-X.; Zhang,
F.-Y. Ultrathin platinum nanowire based electrodes for high-efficiency
hydrogen generation in practical electrolyzer cells. Chem. Eng. J. 2021,
410, No. 128333.
(60) Holzapfel, P. K. R.; Buhler, M.; Escalera-Lopez, D.; Bierling,
M.; Speck, F. D.; Mayrhofer, K. J. J.; Cherevko, S.; Pham, C. V.;
Thiele, S. Fabrication of a Robust PEM Water Electrolyzer Based on
Non-Noble Metal Cathode Catalyst: [Mo3S13]2‑ Clusters Anchored to
N-Doped Carbon Nanotubes. Small 2020, 16 (37), No. 2003161.
(61) Jiang, G.; Yu, H.; Hao, J.; Chi, J.; Fan, Z.; Yao, D.; Qin, B.;
Shao, Z. An effective oxygen electrode based on Ir0.6Sn0.4O2 for PEM
water electrolyzers. J. Energy Chem. 2019, 39, 23−28.
(62) Rozain, C.; Mayousse, E.; Guillet, N.; Millet, P. Influence of
iridium oxide loadings on the performance of PEM water electrolysis
cells: Part II − Advanced oxygen electrodes. Appl. Catal., B 2016, 182,
123−131.

1421 https://doi.org/10.1021/acs.iecr.3c03546
Ind. Eng. Chem. Res. 2024, 63, 1409−1421