# Fileset

[Miyamoto_2025.pdf](https://mdr.nims.go.jp/filesets/5903f162-c536-4622-b3ed-4c3a73328cf7/download)

## Creator

Kentaro Miyamoto, Koji Shimizu, [Anh Khoa Augustin Lu](https://orcid.org/0000-0003-4702-0933), Satoshi Watanabe

## Rights

[Creative Commons BY Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/)

## Other metadata

[Construction of Machine Learning Potentials toward the Exploration of Alloy Cluster Catalysts](https://mdr.nims.go.jp/datasets/6195c927-0e0a-43b1-a18f-f505626a81be)

## Fulltext

Construction of Machine Learning Potentials toward the Exploration of Alloy Cluster Catalystse-Journal of Surface Science and Nanotechnology 23, 188–192 (2025)Construction of Machine Learning Potentials towardthe Exploration of Alloy Cluster CatalystsKentaro Miyamoto,a Koji Shimizu,a, b, † Anh Khoa Augustin Lu,a, c, d Satoshi Watanabe a, ‡a Department of Materials Engineering, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japanb Research Center for Computational Design of Advanced Functional Materials, National Institute of Advanced Industrial Science andTechnology (AIST), 1-1-1 Umezono, Tsukuba, Ibaraki 305-8568, Japanc Research Center for Materials Nanoarchitectonics (MANA), National Institute for Materials Science (NIMS), 1-1 Namiki, Tsukuba, Ibaraki305-0044, JapandMathematics for Advanced Materials Open Innovation Laboratory, National Institute of Advanced Industrial Science and Technology (AIST),2-1-1 Katahira, Aoba-ku, Sendai, Miyagi 980-8577, Japan† Corresponding author: koji.shimizu@aist.go.jp‡ Corresponding author: watanabe@cello.t.u-tokyo.ac.jpReceived: 8 February, 2025; Accepted: 27 March, 2025; J-STAGE Advance Publication: 17 May, 2025; Published: 17 May, 2025High entropy alloys (HEAs) are expected to show excellentperformance in various fields, such as catalysts and high-temper-ature structural materials, but the huge number of configurationsmakes it difficult to find the optimal compositions for HEAs. Inthis study, machine learning potentials were developed to accu-rately predict the total and H/CO adsorption energies of multi-element slab models and cluster models of various sizes andshapes, based on density functional theory calculations.Keywords High entropy alloys; Machine learning; Densityfunctional theory; Catalysts; CO2 reduction reactionI. INTRODUCTIONThe electrochemical reduction of CO2 (CO2RR) is gainingincreasing attention as a key technology for achieving asustainable society. However, its practical implementationremains challenging due to the high thermodynamic stabilityof CO2, which requires substantial energy input for its re-duction. Additionally, CO2RR involves multiple reactionpathways, yielding a variety of hydrocarbon compounds,including formic acid, formaldehyde, methanol, methane,acetic acid, and ethanol. Suppression of the hydrogen evo-lution reaction (HER), a major side reaction, is also impor-tant, because it significantly reduces the Faraday efficiency.Effectively controlling these reaction pathways requires thedevelopment of highly active and selective CO2RR catalysts.Copper (Cu) is widely recognized as a promising CO2RRcatalyst [1] due to its balanced adsorption energies for bothhydrogen (H) and carbon monoxide (CO), which are essen-tial for facilitating hydrocarbon formation. Effective catalystsmust weakly adsorb H atoms to suppress HER while main-taining appropriately strong CO adsorption to promote theconversion to highly reduced hydrocarbon products [2].However, the search for more efficient CO2RR catalystsremains crucial, as Cu requires high overpotentials to effec-tively drive CO2RR [3].High-entropy alloy (HEA) catalysts show great promise inaddressing the challenges of CO2RR. The concept of HEAswas introduced by Cantor et al. [4] and Yeh et al. [5],defining them as single-phase solid solutions composed offive or more elements in relatively equal concentrations (5–35 at%). Recent studies have highlighted numerous HEAswith properties superior to conventional alloys [6], increasingthe expectations for their potential to enhance material func-tionality. On the other hand, identifying HEAs designed forspecific applications remains challenging due to the vastnumber of compositions and configurations. To date, only asmall fraction of the potential composition space has beenexplored, suggesting that many more highly functionalizedHEAs are yet to be discovered.Optimizing the shape and size of the catalyst also im-Regular Papere-J. Surf. Sci. Nanotechnol. 23, 188–192 (2025) | DOI: 10.1380/ejssnt.2025-028 188mailto:koji.shimizu@aist.go.jpmailto:watanabe@cello.t.u-tokyo.ac.jphttps://doi.org/10.1380/ejssnt.2025-028proves its activity and selectivity. Nanoparticle catalysts ex-hibit distinct properties from bulk metals due to their highsurface area-to-volume ratios and pronounced quantum sizeeffects [7]. Various mono-elemental nanoparticles, includingAg, Au, Cu, Ni, and Pd, have demonstrated enhancedCO2RR activity [8]. In terms of size, smaller nanoparticlesoften exhibit higher catalytic activity [7], and sub-nanoclus-ters with fewer than ~30 atoms have shown significantlybetter performances [9]. On the other hand, catalytic activitycan diminish below a certain size [10], indicating that acomplex relationship between size and catalytic performance.Regarding shapes, superior catalytic activity is often attrib-uted to low-coordinated surface sites, such as edges, whichare abundant in small nanoparticles [11]. The (111) surfacehas been reported to increase H adsorption energy [12], andAu nanoparticles with an octahedral shape, which consists of(111) facets, have demonstrated high CO2RR activity [7]. Bycarefully tuning their size and shape, nanoparticle catalystscan achieve superior catalytic performance.Experimentally screening potential catalysts one by one isboth time-consuming and costly. In this regard, computa-tional simulations can be used to accelerate the search byboth reducing the cost and accelerating the process. First-principles calculations based on density functional theory(DFT) are a powerful method for accurately predicting phys-ical properties, and in recent years, databases of DFT calcu-lation results have expanded rapidly. However, DFT calcu-lations are computationally intensive, and despite the largeamount of available data, most results are derived from idealcrystal structures consisting of a limited number of elements.In particular, data on HEAs and nanoclusters remains scarce.Given these limitations, developing methods that achieveDFT accuracy with lower computational costs is highlydesirable. In this context, machine learning potentials(MLPs) have emerged as a promising approach, with inter-atomic potentials based on neural networks gaining increas-ing attention in recent years [13, 14].In this study, MLPs are developed to pave the way towardthe design of HEA cluster catalysts for the CO2RR. Since theactivity and selectivity of this reaction are highly correlatedwith the adsorption energies of CO molecules and H atoms,the MLPs were trained to predict these values. By incorpo-rating training datasets of octahedral cluster models withvarious structural features (vertices, edges, faces, etc.), cata-lytic performance on uneven surfaces is also explored.II. COMPUTATINAL DETAILSA. First-principles calculationsDFT calculations of optimized structures, total energies,and adsorption energies were used as training data for MLPs,focusing on nine elements (Ag, Au, Co, Cu, Ni, Ir, Pd, Pt,and Rh) that are promising as catalysts. The datasets used inthis study include not only those calculated in our laboratorybut also those obtained from previous studies [2, 15] and thenovel materials discovery (NOMAD) database [16].Structure optimization calculations were conducted formodels with different atomic configurations in bulk face-centered cubic (FCC) crystals (8 atoms per supercell) con-taining up to 5 of the nine elements, as well as in octahedralclusters (19 and 44 atoms) containing up to 4 of the nineelements. Octahedral clusters have various structural fea-tures, such as vertices, edges, and faces, which were incor-porated into the training data to investigate the unique cata-lytic performance of uneven surfaces. Additionally, structureoptimizations were performed for several unary models withvarying shapes and sizes, as detailed later in Section III.C.For the in-house dataset, the DFT calculations were per-formed using the vienna ab initio simulation package(VASP) [17]. The exchange-correlation functional was basedon the generalized gradient approximation (GGA) with Per-dew-Burke-Ernzerhof (PBE) [18]. The projector-augmented-wave (PAW) method [19] was used for the interaction be-tween inner-shell electrons and nuclei. A cut-off energy of500 eV was used, and k-point samplings of 10 × 10 × 10 and3 × 3 × 3 were used for the bulk and cluster models, respec-tively, based on the Monkhorst-Pack scheme [20]. Theatomic configurations were relaxed until the atomic forceswere reduced to below 0.01 eVÅ−1.DFT calculations for alloy slab models with 1 to 5 ele-ments, including adsorbed H atoms and CO molecules, wereobtained from the dataset of previous studies [2, 15]. In thesecalculations, the exchange-correlation functional was theRevised PBE [21]. Among them, only the slab models con-taining some of nine elements were used as training data.The NOMAD database contains a large number of clustermodel results compared to other open databases. Only mod-els that satisfy the following conditions were downloadedusing the application programming interface:1) Results from structure optimization of cluster modelsusing VASP,2) Calculated with the GGA-PBE exchange-correlationfunctional, and3) Contain Ag, Au, Cu, Co, Ni, Ir, Pd, Pt, and/or Rh asmetals.The total energy after structure optimization for eachmodel was used in this study. The adsorption energies of Hand CO, denoted as �EH and �ECO, were calculated asfollows.�EH ¼ EH� � E� � 12EH2;�ECO ¼ ECO� � E� � ECO;where EH� and ECO� represent the total energies of the modelafter structure optimization with H and CO adsorbates, re-spectively, E� is the total energy of the model withoutadsorbates, and EH2and ECO are the total energies of H2and CO molecules in the gas phase. More negative adsorp-tion energies indicate stronger interactions between the sur-face and the adsorbate.B. MLPsMLPs based on neural networks were created using theRegular Papere-J. Surf. Sci. Nanotechnol. 23, 188–192 (2025) | DOI: 10.1380/ejssnt.2025-028 189https://doi.org/10.1380/ejssnt.2025-028results of DFT calculations. The training data includes theelement type Z, atomic position ~r, and the total energy E.In this study, we used Allegro, a recently developed graphneural network architecture built on the NequIP framework[22, 23]. Allegro demonstrates excellent accuracy and scal-ability, overcoming the drawbacks of conventional messagepassing neural network, such as high computational cost dueto iterative information propagation. Details on hyperpara-meter settings in Allegro adopted in the present work aredescribed in Supplementary Material.III. RESULTS AND DISCUSSIONA. Prediction of adsorption energy for HEAslabsFirst, using the results of DFT calculations of slab models[2, 15], we created MLPs to predict adsorption energies onHEA slabs. For this purpose, the training data includedunary, binary, ternary, and quaternary alloys (17,007 datafor H adsorption and 14,269 data for CO adsorption, with8–80 atoms per supercell), as well as unary, quaternary, andquinary alloy slab models (960 data for H adsorption and 455data for CO adsorption, with 21–22 atoms per supercell). Toassess the prediction accuracy of the trained MLP, slabmodels containing five elements (45 data with 21–22 atomsper supercell) were used as test datasets. As shown in Fig-ure 1(a, b), the accuracy of the created MLPs was high, withmean absolute errors (MAE) of 32.9 and 31.9meV for theadsorption energies of H atoms and CO molecules, respec-tively.These MLPs were used to predict the adsorption energiesfor additional five-element alloy slab models, includingatomic species from the nine elements mentioned earlier.The lattice constant of each alloy was determined as aweighted average of the optimized bulk values of its con-stituent elements, as shown in Table S1 (SupplementaryMaterial), based on the alloy composition. The distances ofH and C from the surface were determined by the modevalues obtained from datasets from the previous study [15],as shown in Figure S1 (Supplementary Material). Adsorptionenergies of H and CO were predicted at three sites (FCC,HCP, and Top), as shown in Figure 1(c), for 100,000 modelswith 20 atoms in a 2 × 2 supercell consisting of 5 layers. Theresults in Figure 1(d–f ) represent the adsorption energies forH and CO. The colors indicate the ratio, with Ag-, Au-, Cu-,Co-, Ir-, Ni-, Pd-, Pt-, and Rh-rich compositions shown inred, pink, orange, gray, light green, blue, cyan, yellow, andgreen, respectively. The trends for the FCC and HCP sites aresimilar, while the Top site shows an overall increase inadsorption energy. Previous studies have reported a positivelinear relationship between the adsorption energies of H andCO [24, 25]. However, our results indicate that this relation-ship does not necessarily hold for HEA surfaces. Notably, thedarker regions in Figure 1(d–f ) highlight areas where Hadsorption is weaker and CO adsorption is stronger than thaton a Cu surface, which is a desirable characteristic forCO2RR catalysts. Many promising candidate materials arelocated within this region. We also demonstrated that high-throughput screening using MLPs enables the efficient iden-tification of HEA models with the optimal adsorption energy(�ECO = −0.67 eV) [24] at a low computational cost.(a) (b) (c)(d) (e) (f)Figure 1: Prediction of (a) H atom adsorption energies and (b) CO molecular adsorption energies for the slab models. (c) Illustration of the threesites considered for the test data. Atoms in the topmost, second and third layers in the slab model are shown in blue, yellow, and light green,respectively. (d–f ) Distribution of H atom and CO molecule adsorption energies at each site of 100,000 5-element slab models with randomlyconfigured elements.Regular Papere-J. Surf. Sci. Nanotechnol. 23, 188–192 (2025) | DOI: 10.1380/ejssnt.2025-028 190https://doi.org/10.1380/ejssnt.2025-028B. Prediction for multi-element modelsThis section describes the investigation of catalysts usingcluster models. First, DFT calculation results for unary-,binary-, and ternary-alloy models (3,969 data for clustermodels and 741 data for bulk models) were used as trainingdata. The trained MLP was then applied to predict the totalenergy of quaternary- and quinary-alloy model. As shown inFigure 2, the MLP accurately predicts the total energy forboth bulk and cluster models (each 200 datasets), even formulti-element compositions not included in the training data-set. This demonstrates that the model effectively learns therelationship between different elements. This result is a sig-nificant milestone in reducing the computational cost oftraining multi-element models for two key reasons. First,generating DFT datasets for multi-element systems is chal-lenging due to convergence difficulties. Second, existingopen databases primarily contain crystal structures with onlya limited number of elements.C. Prediction for models of various shapesand sizesAs mentioned earlier, optimizing the shape and size ofcatalysts can enhance their activity and selectivity. Therefore,creating MLPs capable of predicting properties for clustermodels with various shapes and sizes is crucial. The AlloyCatalysis Automated Toolkit [26] was used to generate unarymodels with octahedral, cubic, icosahedral, and sphericalshapes (Figure 3) across various numbers of atoms, followedby structure optimization for each. MLPs were then trained topredict the energy of these models.Initially, a pre-trained model specialized in catalyst designwas used for predictions. Column I in Table 1 shows theprediction results using the GemNet-OC-S2EFS-OC20+OC22 model [27, 28]. This model was trained by GemNeton a large dataset from the Open Catalyst Project, whichincludes over 1,000,000 structure optimization snapshots.However, this model appears to be less effective when pre-dicting the total energy of clusters with various shapes andsizes, indicating that the models trained on slab structuresmay not be the best fit for cluster models.Column II in Table 1 shows the results of the Allegromodel trained with binary and ternary alloys in octahedralcluster models consisting of 19 atoms (3,804 data), as well aswith binary and ternary alloy bulk crystals (747 data). Exceptfor the smallest icosahedral case, the accuracy of Allegrotrained with clusters (II) is much better than that of theGemNet-OC-S2EFS-OC20+OC22 (I). This indicates that in-cluding certain cluster data in the training dataset lead toimproved prediction performance for clusters with differentsizes and shapes.Column III in Table 1 shows the results of the Allegromodel trained using the data from II, along with randomlyshaped unary clusters consisting of 3–55 atoms obtainedfrom NOMAD (3,228 data). The inclusion of these randomlyshaped clusters significantly improves the accuracy acrossvarious cluster models. As a result, the MLP model in III isnow well-suited for investigating the stability of a broadrange of models.IV. CONCLUSIONSIn this study, MLPs were developed based on limited DFTcalculation datasets to explore a large configurational spaceof multi-element alloy catalysts by predicting the adsorptionenergies of H and CO. The analysis of slab models with theseMLPs highlights the potential of HEAs as catalysts, with(a) (b)Figure 2: Prediction for (a) quaternary and (b) quinary systemsusing the model trained on DFT calculation data for unary to ternarysystems.Figure 3: Type of shapes used in the test data. As regards size, thenumbers of atoms listed in Table 1 were used to create structures.Table 1: Predictions for models of various shapes and sizes using(I) the GemNet-OC-S2EFS-OC20+OC22 pre-trained model, (II) theAllegro trained with binary and ternary alloys in octahedral clustersconsisting of 19 atoms (3,804 data), as well as with binary andternary alloy bulk crystals (747 data), and (III) the Allegro trainedwith the datasets of (II) and randomly shaped unary clusters ob-tained from NOMAD (3,228 data).ShapeNumberof atomsMAE (meVatom−1) Numberof testdatasetsI II IIIOctahedron44 396.4 130.5 17.6 985 446.6 150.9 15.0 9Cubic63 401.4 74.9 32.7 8171 449.7 88.9 26.9 8Icosahedron13 92.5 279.3 56.6 855 420.7 97.6 13.1 7147 469.2 112.2 23.6 8Sphere135 456.4 107.2 19.6 8179 457.8 124.0 17.9 6Regular Papere-J. Surf. Sci. Nanotechnol. 23, 188–192 (2025) | DOI: 10.1380/ejssnt.2025-028 191https://doi.org/10.1380/ejssnt.2025-028adsorption energy values of �EH and �ECO that are highlycorrelated with CO2RR catalytic activity, laying within thedesirable range. Additionally, our MLPs also provides accu-rate predictions of total energies for models containing up tofive elements, as well as cluster models with varying sizesand shapes, despite limited diversity in the training datasets.Given the current limitations of DFT calculations for HEAclusters and the scarcity of available data, this study providesa promising foundation for future advancements in this field.For example, these MLPs serve as powerful tools for high-throughput screening of the optimal alloy compositions andoptimizing the size and shape of cluster catalysts.AcknowledgmentsThis work was supported by Japan Science and TechnologyAgency (JST) as part of SICORP, Grant Number JPMJSC21E2.Part of calculations were performed using the supercomputers at theInformation Technology Center and the Institute for Solid StatePhysics (ISSP), the University of Tokyo.AppendixThe details of hyperparameter settings in Allegro and predictedadditional slab models in Figure 1(d–f ) are available in Supple-mentary Material at https://doi.org/10.1380/ejssnt.2025-028.Note 1The data that support the findings of this study are available fromthe corresponding authors upon reasonable request.Note 2This paper was presented at the 10th International Symposiumon Surface Science, Kitakyushu International Conference Center,Fukuoka, Japan, 20–24 October, 2024.References[1] T. K. Todorova, M. W. Schreiber, and M. Fontecave, ACS Catal.10, 1754 (2020).[2] J. K. Pedersen, T. A. A. Batchelor, A. Bagger, and J. Rossmeisl,ACS Catal. 10, 2169 (2020).[3] S. Ma, M. Sadakiyo, R. Luo, M. Heima, M. Yamauchi, and P. J.A. Kenis, J. Power Sources 301, 219 (2016).[4] B. Cantor, I. T. H. Chang, P. Knight, and A. J. B. Vincent, Mater.Sci. Eng. A 375–377, 213 (2004).[5] J.-W. Yeh, S.-K. Chen, S.-J. Lin, J.-Y. Gan, T.-S. Chin, T.-T.Shun, C.-H. Tsau, and S.-Y. Chang, Adv. Eng. Mater. 6, 299 (2004).[6] M.-H. Tsai and J.-W. Yeh, Mater. Res. Lett. 2, 107 (2014).[7] T. Eom, W. J. Kim, H.-K. Lim, M. H. Han, K. H. Han, E.-K.Lee, S. Lebègue, Y. J. Hwang, B. K. Min, and H. Kim, J. Phys.Chem. C 122, 9245 (2018).[8] H. Tabassum, X. Yang, R. Zou, and G. Wu, Chem Catal. 2, 1561(2022).[9] R. K. Raju, P. Rodriguez, and E. N. Brothers, Phys. Chem.Chem. Phys. 25, 11630 (2023).[10] X. Deng, D. Alfonso, T.-D. Nguyen-Phan, and D. R.Kauffman, ACS Catal. 13, 15301 (2023).[11] H. Mistry, R. Reske, Z. Zeng, Z.-J. Zhao, J. Greeley, P. Strasser,and B. R. Cuenya, J. Am. Chem. Soc. 136, 16473 (2014).[12] L. Gai, Y. K. Shin, M. Raju, A. C. T. van Duin, and S. Raman,J. Phys. Chem. C 120, 9780 (2016).[13] M. Rittiruam, P. Khamloet, A. Ektarawong, C. Atthapak, T.Saelee, P. Khajondetchairit, B. Alling, S. Praserthdam, and P.Praserthdam, Appl. Surf. Sci. 652, 159297 (2024).[14] S. Watanabe, W. Li, W. Jeong, D. Lee, K. Shimizu, E.Mimanitani, Y. Ando, and S. Han, J. Phys. Energy 3, 012003(2021).[15] K. Tran and Z. W. Ulissi, Nat. Catal. 1, 696 (2018).[16] M. Scheidgen, L. Himanen, A. N. Ladines, D. Sikter, M.Nakhaee, Á. Fekete, T. Chang, A. Golparvar, J. A. Márquez, S.Brockhauser, S. Brückner, L. M. Ghiringhelli, F. Dietrich, D.Lehmberg, T. Denell, A. Albino, H. Näsström, S. Shabih, F.Dobener, M. Kühbach, R. Mozumder, J. F. Rudzinski, N. Daelman,J. M. Pizarro, M. Kuban, C. Salazar, P. Ondračka, H.-J. Bungartz,and C. Draxl, J. Open Source Softw. 8, 5388 (2023).[17] G. Kresse and J. Hafner, Phys. Rev. B 47, 558(R) (1993).[18] J. P. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. 77,3865 (1996).[19] P. E. Blöchl, Phys. Rev. B 50, 17953 (1994).[20] H. J. Monkhorst and J. D. Pack, Phys. Rev. B 13, 5188 (1976).[21] B. Hammer, L. B. Hansen, and J. K. Nørskov, Phys. Rev. B 59,7413 (1999).[22] A. Musaelian, S. Batzner, A. Johansson, L. Sun, C. J. Owen,M. Kornbluth, and B. Kozinsky, Nat. Commun. 14, 579 (2023).[23] S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M.Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, Nat.Commun. 13, 2453 (2022).[24] M. Zhong, K. Tran, Y. Min, C. Wang, Z. Wang, C.-T. Dinh, P.De Luna, Z. Yu, A. S. Rasouli, P. Brodersen, S. Sun, O. Voznyy, C.-S. Tan, M. Askerka, F. Che, M. Liu, A. Seifitokaldani, Y. Pang, S.-C. Lo, A. Ip, Z. Ulissi, and E. H. Sargent, Nature 581, 178 (2020).[25] H. Wan, X. Wang, L. Tan, M. Filippi, P. Strasser, J. Rossmeisl,and A. Bagger, ACS Catal. 13, 1926 (2023).[26] S. Han, G. Barcaro, A. Fortunelli, S. Lysgaard, T. Vegge, andH. A. Hansen, npj Comput. Mater. 8, 121 (2022).[27] R. Tran, J. Lan, M. Shuaibi, B. M. Wood, S. Goyal, A. Das, J.Heras-Domingo, A. Kolluru, A. Rizvi, N. Shoghi, A. Sriram, F.Therrien, J. Abed, O. Voznyy, E. H. Sargent, Z. Ulissi, and C.Lawrence Zitnick, ACS Catal. 13, 3066 (2023).[28] L. Chanussot, A. Das, S. Goyal, T. Lavril, M. Shuaibi, M.Riviere, K. Tran, J. Heras-Domingo, C. Ho, W. Hu, A. Palizhati, A.Sriram, B. Wood, J. Yoon, D. Parikh, C. L. Zitnick, and Z. Ulissi,ACS Catal. 11, 6059 (2021).All articles published on e-J. Surf. Sci. Nanotechnol. are licensedunder the Creative Commons Attribution 4.0 International (CC BY4.0). You are free to copy and redistribute articles in any medium orformat and also free to remix, transform, and build upon articles forany purpose (including a commercial use) as long as you giveappropriate credit to the original source and provide a link to theCreative Commons (CC) license. If you modify the material, youmust indicate changes in a proper way.Copyright: ©2025 The author(s)Published by The Japan Society of Vacuum and Surface ScienceRegular Papere-J. Surf. Sci. Nanotechnol. 23, 188–192 (2025) | DOI: 10.1380/ejssnt.2025-028 192https://doi.org/10.1380/ejssnt.2025-028https://doi.org/10.1021/acscatal.9b04746https://doi.org/10.1021/acscatal.9b04746https://doi.org/10.1021/acscatal.9b04343https://doi.org/10.1016/j.jpowsour.2015.09.124https://doi.org/10.1016/j.msea.2003.10.257https://doi.org/10.1016/j.msea.2003.10.257https://doi.org/10.1002/adem.200300567https://doi.org/10.1080/21663831.2014.912690https://doi.org/10.1021/acs.jpcc.8b02886https://doi.org/10.1021/acs.jpcc.8b02886https://doi.org/10.1016/j.checat.2022.04.012https://doi.org/10.1016/j.checat.2022.04.012https://doi.org/10.1039/D3CP00739Ahttps://doi.org/10.1039/D3CP00739Ahttps://doi.org/10.1021/acscatal.3c03446https://doi.org/10.1021/ja508879jhttps://doi.org/10.1021/acs.jpcc.6b01064https://doi.org/10.1016/j.apsusc.2024.159297https://doi.org/10.1088/2515-7655/abc7f3https://doi.org/10.1088/2515-7655/abc7f3https://doi.org/10.1038/s41929-018-0142-1https://doi.org/10.21105/joss.05388https://doi.org/10.1103/PhysRevB.47.558https://doi.org/10.1103/PhysRevLett.77.3865https://doi.org/10.1103/PhysRevLett.77.3865https://doi.org/10.1103/PhysRevB.50.17953https://doi.org/10.1103/PhysRevB.13.5188https://doi.org/10.1103/PhysRevB.59.7413https://doi.org/10.1103/PhysRevB.59.7413https://doi.org/10.1038/s41467-023-36329-yhttps://doi.org/10.1038/s41467-022-29939-5https://doi.org/10.1038/s41467-022-29939-5https://doi.org/10.1038/s41586-020-2242-8https://doi.org/10.1021/acscatal.2c05315https://doi.org/10.1038/s41524-022-00807-6https://doi.org/10.1021/acscatal.2c05426https://doi.org/10.1021/acscatal.0c04525https://doi.org/10.1380/ejssnt.2025-028