Machine-learning-accelerated design of high-performance platinum intermetallic nanoparticle fuel cell catalysts

Carbon supported PtCo intermetallic alloys are known to be one of the most promising candidates as low-platinum oxygen reduction reaction electrocatalysts for proton-exchange-membrane fuel cells. Nevertheless, the intrinsic trade-off between particle size and ordering degree of PtCo makes it challenging to simultaneously achieve a high specific activity and a large active surface area. Here, by machine-learning-accelerated screenings from the immense configuration space, we are able to statistically quantify the impact of chemical ordering on thermodynamic stability. We find that introducing of Cu/Ni into PtCo can provide additional stabilization energy by inducing Co-Cu/Ni disorder, thus facilitating the ordering process and achieveing an improved tradeoff between specific activity and active surface area. Guided by the theoretical prediction, the small sized and highly ordered ternary Pt2CoCu and Pt2CoNi catalysts are experimentally prepared, showing a large electrochemically active surface area of ~90 m2 gPt‒1 and a high specific activity of ~3.5 mA cm‒2.

The method can be beneficial.After the below comments are addressed, I recommend for publication in the Journal: 1.) For machine learning, it usually requires big data during training to obtain reliable outcomes.Can they authors clarify this size of training data in their work?2.) "corresponding non-noble metal salts on a commercial carbon support,"：the authors need to give details regarding commercial carbon support, materials and band.3.) Can the authors look at this paper: "Hu, B., Yuan, J., Zhang, J., Shu, Q., Guan, D., Yang, G., ... & Shao, Z. (2021).High activity and durability of a Pt-Cu-Co ternary alloy electrocatalyst and its large-scale preparation for practical proton exchange membrane fuel cells.Composites Part B: Engineering, 222, 109082."Which also shows outstanding performance of PtCoCu?4.) It is very impressive that the machine learning tool can identify the catalyst materials that perform superiorly, both performance and durability.The authors may need to find similar literature work (maybe in general field) that supports this kind of identification.5.) Can the authors look at this paper: "Wang, Y., Seo, B., Wang, B., Zamel, N., Jiao, K., & Adroher, X. C. (2020).Fundamentals, materials, and machine learning of polymer electrolyte membrane fuel cell technology.Energy and AI, 1, 100014."Which summarizes a few general machine learning tools for material/chemistry screening?6.) Can the authors provide details regarding their fuel cell configuration and operating condition such as GDL, CL thickness, RH,….?

Reviewer #3 (Remarks to the Author):
In this contribution, authors describe machine learning application to down select fuel cell catalyst in the Pt2CoM family.Based on the active learning, informed by DFT, the authors prepared Pt2 CoCu and Pt2CoNi catalysts and performed experimental analysis.Overall, the paper is publishable in JPCC and Chemistry of Materials.This reviewer doesn't recognize this manuscript is offering any new significant insights of PEMF catalysts or machine learning to satisfy nature communication criteria.Some of the general comments are provided below.
The active learning (based on GPR) and DFT (PBE calculations) associated with the paper are standard practices used in ML and literature.No new computational or chemical insights were provided.What's new here ?Does ML is needed to identify M= Ni or Copper?Importantly, the identification of Cu or Nickel is not surprising ( there are many in the literature, see https://doi.org/10.1039/D0RA05468B, & Sapkota et al, 2022).
The advantage or need and configurational materials space of the catalysts were not described.No quantitative metrics of how machine learning accelerated the materials discovery is explained.
No GitHub details were provided.
There are statements such as PtCoCu or PtCoNi show high ORR activity.No explanations or valid hypothesis were mentioned.Detailed DFT studies are needed to show the binding sites and demonstrate reaction pathways.
Overall, this reviewer does not recommend this paper for Nature Communications Reviewer #1 (Remarks to the Author): This paper reports a machine-learning (ML) approach to intermetallic nanoparticle catalyst design, synthesis and test.ML-guided prediction gives them Pt2CoM (M =Cu or Co) to be the best catalyst for ORR among 15 different M's the authors screened.
Then the authors made Pt2CoM nanoparticles and converted them into the desired intermetallic structure.As predicted, these intermetallic nanoparticles showed much enhanced catalysis for ORR in both H-cell and MEA tests.The work is of great interest for current searching of active and robust catalysts for fuel cell applications.The paper is generally well-written, but some important issues need to be clarified before the work can be considered for publication in this journal.

Response:
We sincerely appreciate the reviewer for the positive comments.
1) The authors started off PtCo and alloy PtCo with M to search for the best Pt2CoM as catalysts.The authors did not mention what could happen of PtM are directly made and studied as catalysts for ORR.Intermetallic PtCo is known to be active and robust, and intermetallic PtCoNi has also been reported to show even higher activity towards ORR.The activity reported in this paper is similar to what has been reported.The authors should comment these more specifically, as these prior arts are more related to what they reported here in this paper.
Response: Many thanks for the reviewer's comments on this issue.
PtCo represents the most promising low-Pt catalyst for practical PEMFCs applications.
For example, Toyota has launched several fuel-cell vehicles based on the binary PtCo alloy catalysts.Therefore, in this work, we performed our theoretical and experimental studies by focusing on the binary PtCo and ternary PtCoM systems and we did not extended our studies to other binary systems, such as PtCu or PtNi.
We agree with the reviewer that binary PtCo and even ternary PtCoM (M=Cu, Ni, etc) ORR catalysts have been widely reported.Despite these prior works match our computational screening results, there are notable differences that distinguish our research from reported ones.We focus more on the formation process of IMC, aiming to maintain the intrinsic high strain levels of PtCo without sacrificing the active surface area.We therefore resorted to machine learning to expedite the screening process for identifying combinations with high thermodynamic driving forces.This enables us to achieve a balance between intrinsic activity and ECSA, ultimately leading to high mass activity that is meaningful for practical applications.As suggested by the reviewer, we have already cited and discussed the related works reported by others (PtCoNi, J. Am.Chem. Soc. 2020, 142, 45;PtCoCu, Small, 2023, 19, 2300112).

[Revision to manuscript] (Page1)：
Carbon supported platinum-based intermetallic alloys are promising candidates as low-platinum oxygen reduction reaction electrocatalysts for proton-exchangemembrane fuel cells.Intermetallic PtCo is known to be one of the most promising intermetallic fuel cell catalyst.
[Revision to manuscript] (Page7)： …, and thus were finally selected for the experimental validation.There has been prior research reporting PtCoNi and PtCoCu ternary system as excellent ORR catalyst due to their near-optimum strain levels for higher ORR activity 33,34 , aligning well with our computational screening results.More importantly, our study provides a novel design perspective that the introduction of Cu/Ni leverages the thermodynamic driving force for the disordered-to-order transition, resulting in a more favorable tradeoff between specific activity (SA) and electrochemically active surface area (ECSA).
2) It is not clear to me what makes Pt2CoCu and Pt2CoNi to be the optimum combination as the ORR catalyst.How can Cu or Ni make the PtCo more active?Is it possible to characterize CoCu or CuNi position in the alloy structure?
Response: We would like to thank the reviewer for the above constructive suggestions.
(a) "How can Cu or Ni make the PtCo more active?"PtCo demonstrates remarkable efficiency as an ORR catalyst, however the formation of highly ordered intermetallic alloys often necessitates elevated annealing temperatures and inevitably results in particle sintering issues, leading to a reduction in the effective electrochemical surface area.Our work aims to enhance the ordering of Pt-Co alloy by improving thermodynamic driving forces.Computationally, we show that the screened element Cu and Ni can effectively facilitate the Pt-Co ordering.Thus, a catalyst that combines high intrinsic activity (high SA) and small size (high ECSA) can be achieved.As shown in the Figure R1, the improved tradeoff between ordering (higher activity) and size (larger ECSA) leads to the optimized ORR performance.Based on experimental and theoretical results, the introduction of Cu/Ni does not necessarily enhance the intrinsic activity of PtCo.Instead, it plays an essential role in augmenting the thermodynamic driving force for the disordered-to-order transition process, consequently promoting the production of highly-ordered intermetallic compounds (IMCs).The calculated adsorption energies of OH* and O* were plotted as the activity descriptor to evaluate the ORR activity of the Pt2CoM (M= Ga, Ni, Cu, and Ti) using Pt-shell slab models with ordered Pt2CoM-core (Fig. R2).For comparison, we also marked the calculated activity of fully ordered and randomly ordered PtCo with various arrangements of subsurface PtCo core.Generally, fully ordered PtCo showed higher activity than randomly ordered one, and ternary Pt2CoCu and Pt2CoNi show comparable ORR activity with fully ordered PtCo.Experimentally, the RDE results also demonstrated that the specific activity of Pt2CoCu and Pt2CoNi is close to that of the highly-ordered PtCo (Fig. R1).
According to the reviewers' suggestions, we have added the following sentences in the main text to highlight the above discussion:

[Revision to manuscript] (Page7)：
For ternary alloys, Pt2CoCu and Pt2CoNi show comparable ORR activity with fully ordered PtCo, and were finally selected for the experimental validation.… More importantly, our study provides a novel design perspective that the introduction of Cu/Ni leverages the thermodynamic driving force for the disordered-to-order transition, resulting in a more favorable tradeoff between specific activity (SA) and electrochemically active surface area (ECSA).
(b) "Is it possible to characterize CoCu or CuNi position in the alloy structure?" It was difficult to distinguish on the basis of the contrast due to the similar atomic radius and similar atomic weight.A previously published work (J.Am.Chem. Soc. 2020, 142, 45) has also tried to identify this issue, but failed.However, our theoretical prediction found the positions of CoCu or CuNi have a significant impact on the energy of ternary   Soc. 2020, 142, 45).
3) I do not understand Figure 1b.Is it true that MCo interacts more strongly with Pt than Co does?Is there any experimental evidence to support this calculated result?
Response: Many thanks for the reviewer's comments on this issue.The original Figure 1b was a schematic plot, illustrating the presence of Pt2CoM combination with potentially higher thermodynamic driving force of disordered-to-ordered transition.The ordering energy ordering measures the thermodynamic driving force for the disorder-to-order transition, which is defined as the energy difference between ordered configurations (SRO of Pt-Co/M:  The dilemma of the PtCo synthesis often revolves around a lower ordering degree, which potentially inhibits its specific activity and durability.Thus, high temperature has to be used to promote the IMC nucleation with the expense of losing ECSA (e.g.high-ordered/ large-sized PtCo*).Therefore, we aim at the de novo design of the element composition to increase the thermodynamic driving force for the disorderedto-order transition and thus promote the ordering degree.By comparing similar particle size of PtCo # and PtCoCu/Ni, without sacrificing ECSA, it is observed that PtCoCu/Ni (>50% ordering degree) show a higher ordering degree than that of PtCo # (~5%).known that an IMCs@Pt core-shell structure would be formed in acid electrolytes and the Pt-shell could effectively stabilize M against leaching to guarantee the durability under harsh voltage conditions 26, 29".This is not the proof that their nanoparticles have the same structure.
Response: Many thanks for the reviewer's comments on this issue.We have revised these vague expressions.During the ADT test in RDE, the Pt2CoCu/Pt2CoNi catalysts showed a drop of 17.1% and 19.2% in the MA, along with a decrease of 10.2% and 22.2% in the SA.For the ADT test in MEA, the Pt2CoCu catalysts showed a 25.3% loss of MA.The MEA performance degradation was due to the non-noble metal dissolution and particle sintering.In the XRD analysis of Pt2CoCu-MEA after ADT (Fig. R7), we observed that the super-lattice peaks became weaker, indicating the dissolution of CoCu and consequent degradation in performance.

Page 10：
…, thus leading to a large MA of ~3.0A mgPt −1 .Moreover, after 30K accelerated durability test (ADT) by cycling the potential between 0.6 and 0.95 V in RDE, the Pt2CoCu/Pt2CoNi catalysts showed a slight drop of 17.1% and 19.2% in the MA, along with a decrease of 10.2% and 22.2% in the SA (Fig. 5c,d and Fig. S12).

Page 10：
Prior to PEMFCs tests, the pristine IMCs catalysts were subjected to acid leaching and low-temperature H2-annealing to form active and stable Pt-IMCs@Pt core-shell structures 8,11 .EDS elemental mapping indicated the successful formation of core-shell structure with a Pt-rich shell (Fig. 6a).Atomic resolution HAADF-STEM and corresponding intensity profiles clearly verified an L10 intermetallic core surrounded by a Pt shell with three atomic layers (Fig. 6b).
Reviewer #2 (Remarks to the Author): This paper presents a machine learning study to screen catalyst for PEM fuel cells.They identified two potential ternary candidates, which were verified by their experiment.
The ordered structures were characterized using various imaging methods.The fuel cell performance using the catalyst seems outstanding, superior to state-of-the-art in terms of both Pt loading and performance.The paper is well prepared, and the results are exciting.The method can be beneficial.After the below comments are addressed, I recommend for publication in the Journal: Response: We sincerely appreciate the reviewer for the positive comments.
1) For machine learning, it usually requires big data during training to obtain reliable outcomes.Can they authors clarify this size of training data in their work?
Response: Many thanks for the reviewer's comments on this issue.The sizes of each DFT-calculated configuration are summarized in Table R1.As the review said, for machine learning, a significant amount of data is typically required to ensure reliable results.To reduce the data requirements in DFT, active learning is employed in our work.Active learning begins with a small initial dataset and strategically incorporates new data points into the dataset and retrain the model iteratively, which significantly reduces the computation intensity.This approach effectively mitigates issues of uneven or insufficient data points, consequently reducing data size demands.
In Table 2R, we compiled a list of studies employed machine learning method in predicting alloy configuration energies.While these studies vary in the problems studied, dataset construction, and selected machine learning models, it is evident that, for the specific task of predicting PtCoM configuration energies in this study, the employed method achieves a comparable level of accuracy while requiring less data than reported in the literature.

[Revision to manuscript] (Page7)：
The Pt2CoCu and Pt2CoNi catalysts were synthesized by the wet-impregnation of H2PtCl6 and corresponding non-noble metal salts on a carbon support Black Pearl 2000, followed by a high-temperature annealing at 1000 °C.
3) Can the authors look at this paper: "Hu, B., Yuan, J., Zhang, J., Shu, Q., Guan, D., Response: We would like to thank the reviewer for bringing this remarkable work to our attention.The author reported PtCoCu as an efficient and durable fuel cell oxygen reduction catalyst, mainly attributing the improved performance to significant compressive strain.Our work offers a novel perspective by optimizing the chemical ordering.According to the reviewer's suggestion, we have cited this paper in the revised manuscript.
4) It is very impressive that the machine learning tool can identify the catalyst materials that perform superiorly, both performance and durability.The authors may need to find similar literature work (maybe in general field) that supports this kind of identification.

Response:
We appreciate the reviewer for the above kind suggestions.We added the following citations on machine learning for material discovery.Response: We are grateful to the reviewer for introducing this outstanding work to our notice.As per the reviewer's recommendation, we have carefully added this paper to our citation.

[Revision to manuscript] (Page3) ：
Recently, machine learning methods have demonstrated significant potential in accelerating material discovery by efficiently navigating design spaces and predicting properties, thereby substantially reducing the cost of identifying and optimizing novel materials. 26-29.
6.) Can the authors provide details regarding their fuel cell configuration and operating condition such as GDL, CL thickness, RH,….?
Response: Many thanks for the reviewer's comments on this issue.The fuel cell configuration and operating condition are presented in the experimental section.A gas diffusion layer (GDL) was used Freudenberg (H24CX483, 235um) including a microporous layer.The catalysts layer (CL) thickness of the cathode and anode are about 5um and 2um.Two GDLs, two gaskets, and the prepared CCM constitute the membrane electrode assembly (MEA) with a 34% compression.Single fuel cell test station was Scribner 850e.The seven channel serpentine flow field was applied for the all single-cell tests (designed by Hubert Gasteiger and co-workers), where the pressure drop between the inlet and outlet of the flow filed was less than 10 kPa.The H2-air performance of single cell was conducted at 80 ºC, 100% relative humidity, 150 kPaabs, outlet H2-air at 0.5/2 L min -1 flow rate.
Reviewer #3 (Remarks to the Author): In this contribution, authors describe machine learning application to down select fuel cell catalyst in the Pt2CoM family.Based on the active learning, informed by DFT, the authors prepared Pt2CoCu and Pt2CoNi catalysts and performed experimental analysis.
Overall, the paper is publishable in JPCC and Chemistry of Materials.This reviewer doesn't recognize this manuscript is offering any new significant insights of PEMF catalysts or machine learning to satisfy nature communication criteria.Some of the general comments are provided below.

Response:
We are sorry that the novelty of our work may have not been fully conveyed to the reviewer.Here, we would like to reemphasize the new insights of our work offers, specifically the correlation between chemical ordering and performance.We hope this clarification will help highlight its significance.Detailed point-to-point responses are provided below.To the best of our knowledge, our work is among the earliest endeavors, if not the first attempt, to statistically quantify the impact of chemical ordering on thermodynamic stability with the aid of machine learning method.Additionally, the distinct effects of Pt-Co/M ordering and Co-Cu/Ni ordering are reported for the first time.The welldefined ordered structures of intermetallic compounds (IMC) at the atomic level ensure a more accurate alignment between experimental and machine learning simulations.
This alignment facilitates the identification of potential IMC combinations with a high thermodynamic driving force for the disorder-to-order transition within the vast design space.
Based on the reviewer's feedback, we have made the following changes to the main text to better convey our novelty.The number of data points of DFT calculated configuration and ML predicted are 120 (1 configurations per data point) and 3800 (average of 20 configurations for each data point), respectively.

[Revision to manuscript]
1. Page3 ： …, which would significantly limit the nucleation rate of IMC phase.
Recently, machine learning methods have demonstrated significant potential in accelerating material discovery by efficiently navigating design spaces and predicting properties, thereby substantially reducing the cost of identifying and optimizing novel materials. 26-29.

Page 6 ：
A higher SRO value represents a higher disordered degree.Triangles and circles represent data points computed by DFT (1 configurations per data point) and ML prediction model (average of 20 configurations for each data point), respectively.

Page5 ：
The trained machine learning model was applied to predict the formation energies of 300 ordered ( = −1/3) and 300 random ( as well as 3800 structures with varying ordering degrees (Fig. S6).

Page5 ：
…, but the introduction of Mn and Fe would significantly suppress the thermodynamic driving force for the disorder-to-order transition of Pt2CoM (Fig. 2c and Table S1).In the realm of computational efficiency, training and predicting with machine learning models take negligible time compared to DFT calculations, leading to a significant reduction in the time required to establish the correlation between chemical ordering and stability (Fig. S9).

Response:
We thank the reviewer for suggesting these literatures.As noted in the review, there are studies in the literature employing Pt2CoCu and Pt2CoNi as oxygen reduction reaction (ORR) electrocatalysts (J.Am.Chem.Soc. 2020, 142, 45;Small, 2023, 19, 2300112).However, these investigations mostly emphasize that the enhanced activity of PtCoM stems from their near-optimum strain.To some extent, these findings also validate the accuracy of our theoretical prediction.
On the other hand, our work provides a novel design perspective by enhancing the ordering of PtCo through element doping.With the aid of DFT and machine learning, we identified Cu and Ni for their capability to increase the thermodynamic driving force for intermetallic formation while preserving the superior intrinsic activity from candidate elements.Experimentally, PtCoCu/Ni (>50% ordering degree) can be achieved at lower annealing temperature and smaller particle size.The prepared ternary alloy PtCoCu/Ni exhibits higher SA compared to PtCo# (small size, low order) and a larger ECSA than PtCo* (large size, high order).Additionally, enhanced chemical ordering can also effectively contributes to the durability of the catalysts.
We have made the following edit to the text to address the reviewer's concern, and properly cited the corresponding literatures.
[Revision to manuscript] (Page7)： …, and thus were finally selected for the experimental validation.There has been prior research reporting PtCoNi and PtCoCu ternary system as excellent ORR catalyst due to their near-optimum strain levels for higher ORR activity 33,34 , aligning well with our computational screening results.More importantly, our study provides a novel design perspective that the introduction of Cu/Ni leverages the thermodynamic driving force for the disordered-to-order transition, resulting in a more favorable tradeoff between specfic activity (SA) and electrochemically active surface area (ECSA).
3. The advantage or need and configurational materials space of the catalysts were not described.No quantitative metrics of how machine learning accelerated the materials discovery is explained.

Response:
We appreciate the reviewer for bringing this to our attention.A detailed response to this question has been provided in the response to the prior question "Does ML is needed to identify M= Ni or Copper?" 4. No GitHub details were provided.

Fig. R1 .
Fig. R1.The introduction of Cu/Ni break the trade-off between intrinsic activity and active surface area in the synthesis of PtCo.

Fig. R2 .
Fig. R2.ORR activity volcano plot as a function of OH and O adsorption energy relative to Pt of four ordered Pt2CoM, fully-ordered PtCo and five randomly-ordered PtCo.
systems.The disordering of Co-Cu or Cu-Ni can provide additional thermodynamic stabilization energy (FigureR3), suggesting that the Co with Cu/Ni tend to have random site occupation in the ternary Pt2CoCu/Pt2CoNi alloy structure.

Fig. R3 .
Fig. R3.The effect of CoCu (a) or CuNi (b) position for the energy of ternary systems.
As shown in Fig 2C, various metals exhibits distincteffects on the ordering energy ordering .Among them, the screened metals, Cu and Ni, prove to be effective in leveraging the ordering energy, and consequently were selected for experimental validation.The correlation between the ordering energy trend and the Pt-Co-M pair interactions is an interesting topic, but falls beyond the scope of the current study.We are working on this issue and will present our findings in future work Taking into account the reviewer's feedback, we have revised this schematic illustration to more accurately represent the catalyst design guideline of this work, specifically emphasizing the incorporation of a third metal to enhance the thermodynamic driving forces.(FigureR4).
Fig. R4.Schematic illustration showing that the thermodynamic driving force of disordered-to-ordered transition could be enhanced by forming ternary IMC structure.

Fig
Fig R5.The compassion of ordering degree between PtCoCu/Ni and PtCo under the similar size.[Revision to manuscript] (Fig.1b) ：

Fig
Fig R6.XRD and RDE test of CoCu/C and CuNi/C catalysts.

Fig
Fig R7.XRD pattern of Pt2CoCu-MEA after ADT."This is not the proof that their nanoparticles have the same structure."Assuggested by the reviewer, we further performed energy dispersive spectroscopy (EDS) elemental mapping and aberration-corrected HAADF-STEM to verify IMCs@Pt core-shell structure.EDS elemental mapping indicated the successful formation of core-shell structure with a Pt-rich shell (Fig.R8).Atomic resolution HAADF-STEM and corresponding intensity profiles clearly verified an L10 intermetallic core surrounded by three atomic layers of a Pt shell (Fig.R9).
Yang, G. ... & Shao, Z. (2021).High activity and durability of a Pt-Cu-Co ternary alloy electrocatalyst and its large-scale preparation for practical proton exchange membrane fuel cells.Composites Part B: Engineering, 222, 109082."Which also shows outstanding performance of PtCoCu?

[
Revision to manuscript] (Page3) ： Recently, machine learning methods have demonstrated significant potential in accelerating material discovery by efficiently navigating design spaces and predicting properties, thereby substantially reducing the cost of identifying and optimizing novel materials 26-29 .5.) Can the authors look at this paper: "Wang, Y., Seo, B., Wang, B., Zamel, N., Jiao, K., & Adroher, X. C. (2020).Fundamentals, materials, and machine learning of polymer electrolyte membrane fuel cell technology.Energy and AI, 1, 100014."Which summarizes a few general machine learning tools for material/chemistry screening?

1.
The active learning (based on GPR) and DFT (PBE calculations) associated with the paper are standard practices used in ML and literature.No new computational or chemical insights were provided.What's new here?Does ML is needed to identify M= Ni or Copper?"No new computational or chemical insights were provided.What's new here?"Response: Thanks for the candid feedback from the reviewer.Our primary focus is on enhancing the ordering of Pt-Co alloys by introducing a third element, thereby reducing the annealing temperature to achieve a highly ordered alloy.Computational screening reveals that the addition of Cu and Ni effectively promotes ordering in Pt-Co without compromising the alloy's oxygen reduction reaction (ORR) activity.Utilizing machine learning techniques, we statistically correlate chemical ordering and relative stability, enabling a quantitative assessment of the thermodynamic driving force behind intermetallic alloy formation.Furthermore, our findings highlight that the ordering of Pt-Co contributes to alloy stability, and the introduction of disorder between Co-Cu/Ni provides additional stabilization energy.All of the aforementioned insights are novel findings resulting from our computational work.
Pt2CoM (M represents base metal element) combinations with high thermodynamic driving force for the disorder-to-order transition.With the aid of machine learning method, we are able to statistically quantify the impact of chemical ordering on thermodynamic stability.In particular, the alloying of Cu or Ni, inducing Co-Cu or Co-Ni disorder, provides additional stabilization energy, facilitating the ordering process and an improved tradeoff between electrochemically active surface area (ECSA) and specific activity (SA).Guided by the theoretical prediction, the small sized and highly ordered Pt2CoCu and Pt2Co Ni catalysts are experimentally prepared, showing a large ECSA of ~90 m 2 gPt -1 and a high SA of ~3.5 mA cm -2 .makes it feasible to quickly discover potential IMC combinations with high thermodynamic driving force for the disorder-to order transition from the enormous design space.Computationally, we found that alloying Cu or Ni promotes the formation of IMC due to the addtional stabilization energy introduced by the Co-Cu or Co-Ni disordering."Does ML is needed to identify M= Ni or Copper?"Response: We thank the reviewer for raising this question, and our answer is affirmative.Machine learning (ML) plays a crucial role in computing the ordering energy and comprehending the stabilization factor for Ni and Cu doping.Quantitatively evaluating the thermodynamic driving force of Pt2CoM transitions from disorder to order requires evaluating a significant number of configurations at different chemical orderings in Pt2CoM.Given the diversity in the atomic arrangement of Pt, Co, and Melements, this constitutes an extensive space, making it impractical to conduct density functional theory (DFT) calculations for all configurations.To address this challenge, machine learning method was employed, utilizing a limited amount of DFT data (~100 per Pt2CoM system) as the training dataset, to predict the energies of additional Pt2CoM configurations with various chemical orderings.This approach substantially reduces both DFT computational costs and time expenditures.As our super bulk model of Pt2CoM contains 108 atoms (54 Pt atoms, 27 Co atoms and 27 M atoms), if we neglect the symmetry, the total number of isomers for this 108-atom Pt2CoM model can be a rather extensive number.We employ active learning in a batchwise manner to choose just over 100 representative configurations for DFT calculations.This allows for a comprehensive and unbiased representation of the entire configuration space.In particular, one DFT geometry optimization task for Pt2CoM bulk consumes approximately 2 hours when executed on two 2.1 GHz 28-core CPU nodes.In contrast, training a GPR model and utilizing it for energy prediction can be accomplished within just one minute on a personal computer, this time is negligible when compared with DFT computations, the majority of the time spent in machine learning is on generating the DFT training dataset.Specifically, as shown in FigureR10, through 120 DFT data points, only a rough correlation between structural stability and chemical ordering can be discerned.However, by utilizing machine learning to quickly analyze 190 varying ordering degrees (20 configurations for each data point, in total 3800 configurations), a clearer structure-performance relationship as a function of Pt-Co/M and Co-M ordering can be established, saving nearly 32 times.It is essential to note that the time savings achieved through machine learning become even more significant when applied to a larger size of supercell or a greater number of structures.

Table R1 .
The size of DFT calculated configurations for the different Pt2CoM.The minor difference in data size is attributed to the removal of duplicate configurations.