Interpretable design of Ir-free trimetallic electrocatalysts for ammonia oxidation with graph neural networks

Pillai, Hemanth Somarajan; Li, Yi; Wang, Shih-Han; Omidvar, Noushin; Mu, Qingmin; Achenie, Luke E. K.; Abild-Pedersen, Frank; Yang, Juan; Wu, Gang; **n, Hongliang

doi:10.1038/s41467-023-36322-5

Interpretable design of Ir-free trimetallic electrocatalysts for ammonia oxidation with graph neural networks

Article
Open access
Published: 11 February 2023

Volume 14, article number 792, (2023)
Cite this article

Download PDF

You have full access to this open access article

From

View current issue

Interpretable design of Ir-free trimetallic electrocatalysts for ammonia oxidation with graph neural networks

Download PDF

9327 Accesses
22 Citations
17 Altmetric
1 Mention
Explore all metrics

An Author Correction to this article was published on 20 February 2023

This article has been updated

Abstract

The electrochemical ammonia oxidation to dinitrogen as a means for energy and environmental applications is a key technology toward the realization of a sustainable nitrogen cycle. The state-of-the-art metal catalysts including Pt and its bimetallics with Ir show promising activity, albeit suffering from high overpotentials for appreciable current densities and the soaring price of precious metals. Herein, the immense design space of ternary Pt alloy nanostructures is explored by graph neural networks trained on ab initio data for concurrently predicting site reactivity, surface stability, and catalyst synthesizability descriptors. Among a few Ir-free candidates that emerge from the active learning workflow, Pt₃Ru-M (M: Fe, Co, or Ni) alloys were successfully synthesized and experimentally verified to be more active toward ammonia oxidation than Pt, Pt₃Ir, and Pt₃Ru. More importantly, feature attribution analyses using the machine-learned representation of site motifs provide fundamental insights into chemical bonding at metal surfaces and shed light on design strategies for high-performance catalytic systems beyond the d-band center metric of binding sites.

Adsorbate chemical environment-based machine learning framework for heterogeneous catalysis

Article Open access 02 October 2022

Ab Initio to Activity: Machine Learning-Assisted Optimization of High-Entropy Alloy Catalytic Activity

Article 03 October 2022

Investigating Periodic Table Interpolation for the Rational Design of Nanoalloy Catalysts for Green Hydrogen Production from Ammonia Decomposition

Article Open access 07 September 2023

Introduction

Ammonia (NH₃), an essential molecule of the planetary nitrogen cycle, has been widely used in modern industries ranging from medicine, food, and agriculture to the manufacturing of household chemicals. Particularly promising is its application for energy generation and storage technologies because of its easy liquefaction and high volumetric energy density, appealing characteristics for co** with the variability of renewable energy sources, e.g., wind, solar, and hydro¹. In this aspect, design of materials for efficiently catalyzing the ammonia oxidation reaction (AOR) at the anode of direct ammonia fuel cells (DAFCs) holds the key to enable the transition away from fossil feedstocks to a sustainable energy economy. Platinum (Pt) is known as the best elemental metal catalyst for NH₃ oxidation to benign dinitrogen (N₂) in alkaline media, albeit suffering from large overpotentials (~0.5 V)^{2,Full size image}

To gain insights into the local structure of the alloy nanoparticles, we conducted high-angle annular dark-field-scanning transmission electron microscopy (HAADF-STEM) and EDS measurements. The STEM images showed the cubic nanoparticle with a particle diameter of 10.0 ± 0.2 nm (Fig. 3b), and the enlarged image (in red circle, Fig. 3c) displays the atomic structure of the surface region of Pt₃Ru_1/2Co_1/2 with the lattice fringes of 0.212 ± 0.01 nm at the lateral direction and 0.312 ± 0.02 nm at the angular direction, corresponding to (111) and (220) planes of Pt₃Ru_1/2Co_1/2 alloys, respectively. These lattice fringes are slightly smaller than that of the standard pure Pt (JCPDS: 04-0802), with the corresponding lattice fringes of 0.221 nm and 0.360 nm, respectively. Figure 3d shows the overlay of elemental map** images, where the atom columns of Pt, Ru and Co are also clearly resolved in Supplementary Figs. 14, 15. Energy-dispersive X-ray spectroscopy line scanning further demonstrates the uniform distribution of Pt, Ru, and Co atoms in nanostructures (EDS, Supplementary Fig. 16).

The X-ray diffraction (XRD) patterns in Fig. 3e show broad diffraction peaks at ca. 23.8°, assigning to the (002) facet of carbon from the reduced graphene oxide³⁷. The diffraction peaks of the pure Pt catalyst can be identified at ca. 39.8, 46.2, 67.5, 81.3 and 85.7°, corresponding to the (111), (200), (220), (311) and (222) facets of face centered cubic crystalline Pt (JCPDS: 04-0802). Compared to the pure Pt catalyst, these diffraction peaks in Pt₃Ir and Pt₃Ru_1/2Co_1/2 catalysts show slightly positive shift because Pt atoms were partially replaced by smaller Ir or Ru/Co atoms, leading to a lattice compression of the nanoparticles^13,14. These results further confirm the formation of Pt₃Ir and Pt₃Ru_1/2Co_1/2 alloy nanoparticles. X-ray photoelectron spectroscopy (XPS) was conducted to identify the near surface elemental compound and chemical state. High-resolution XPS of Pt 4 f in rGO-supported Pt₃Ir and Pt₃Ru_1/2Co_1/2 (Fig. 3f) shows that the binding energy of Pt 4 f shifts to a higher value relative to rGO-supported Pt, which can be attributed to the strong electronic interaction within the Pt₃Ir and Pt₃Ru_1/2Co_1/2³⁸.

Catalytic performance of the Pt, Pt₃Ir, Pt₃Ru and Pt₃Ru_1/2Co_1/2 nanocubes were tested via cyclic voltammetry (CV) in an Ar-saturated 1.0 M KOH + 0.1 M NH₃ aqueous electrolyte (Pt₃Ru was included as a reference system). Figure 4a shows that Pt₃Ir has the lowest onset potential of 0.35 V vs. RHE, while Pt₃Ru and Pt₃Ru_1/2Co_1/2 have onset potentials of 0.46 and 0.42 V, all of which are lower than pure Pt of 0.48 V. Nevertheless, Pt₃Ru_1/2Co_1/2 has the highest peak mass current activity of 174.0 A g^–1_Pt, higher than those of pure Pt (78.6 A g^–1_Pt), Pt₃Ir (33.1 A g^–1_Pt) and Pt₃Ru (80.2 A g^–1_Pt) catalysts. Further CO strip** measurements in CO-saturated 0.1 M HClO₄ aqueous solution were conducted to estimate the electrochemical activate area (ECSA) of those Pt-based catalysts (see Supplementary Fig. 17). The ECSA values for Pt, Pt₃Ir, Pt₃Ru and Pt₃Ru_1/2Co_1/2 were estimated to be 27.8, 37.7, 19.6 and 23.7 m² g^–1, respectively, which were summarized in Supplementary Fig. 18. Accordingly, the AOR specific activities for those Pt-based catalysts were plotted in Fig. 4c, showing a similar activity trend as the mass activity. The temperature dependent AOR activity of the rGO-supported Pt₃Ru_1/2Co_1/2 catalyst was further measured, as shown in Fig. 4d. With increasing temperature from 25 to 80 °C, the AOR current significantly increased, resulting in lowered onset potentials and six-fold enhancement of the peak current density. In addition, the synthesized rGO-supported Pt₃Ru_1/2Ni_1/2 and Pt₃Ru_1/2Fe_1/2 nanocubes (Supplementary Fig. 19, 20) also show more active AOR performance than Pt, Pt₃Ir, and Pt₃Ru (Supplementary Fig. 21). For example, the highest peak current densities for Pt₃Ru_1/2Ni_1/2 and Pt₃Ru_1/2Fe_1/2 are 123.0 and 181.4 A g^–1_Pt, respectively. These activity trends correspond well with the prediction trend from the screening shown in Fig. 2g, further validating the design approach. It is important to note that the adsorption energies of both *N bridge and hollow species on Pt₃Ru_1/2Ni_1/2, Pt₃Ru_1/2Fe_1/2 and Pt₃Ru_1/2Co_1/2 are within 0.1 eV, which is similar to the uncertainty of the microkinetic models introduced by linear adsorption-energy scaling relations. The short-term AOR stability of the Pt, Pt₃Ir, Pt₃Ru and Pt₃Ru_1/2Co_1/2 nanocubes was studied via chronoamperometry (CA) at 0.55 V vs. RHE in an Ar-saturated 1.0 M KOH + 0.1 M NH₃ aqueous electrolyte. As shown in Fig. 4b, a rapid decline in current density is noticed during the CA running. The initial rapid decline is derived from the removal of non-faradaic double-layer currents, while the subsequent decay is due to the deactivation of surface sites^37,39,40. This is likely attributed to the formation of poisoning species such as *NO that blocks the active sites^40,41. Compared to pure Pt, Ir alloying results in an enhanced AOR activity (Fig. 4a), while Ru does not improve the activity and shows a faster activity loss during the CA running (Fig. 4b). Encouragingly, the introduction of Co tends to stabilize surface sites in Pt₃Ru_1/2Co_1/2, and showed ~4 times higher remained current than Pt₃Ru, further indicating the critical role of both Ru and Co atoms in the ternary alloy system. Specifically, the Pt₃Ru_1/2Co_1/2 retained the highest current density (6.7 A g^–1_Pt) after 1,800 s of CA running when compared with Pt₃Ir (5.6 A g^–1_Pt), Pt₃Ru (1.7 A g^–1_Pt) and Pt (4.2 A g^–1_Pt). Furthermore, their electrochemical surface area changes before and after 1800 s of CA testing was studied and shown in Supplementary Fig. 22, which further confirmed the superiority of the ternary Pt₃Ru_1/2Co_1/2 toward AOR electrocatalysis. Most importantly, those catalysts are Ir-free, which is an important step toward potentially large-scale commercial applications of direct ammonia fuel cells.

**Fig. 4: NH₃ electrooxidation performance.**

Advancing theory of chemisorption with interpretable deep learning

To gain physical insights into the superior AOR activity of Pt₃Ru–M (M: Fe, Co, or Ni) electrocatalysts, we have inspected the causal relationships between the machine-learned features of a site motif and reactivity descriptors using post hoc interpretation techniques. This type of fundamental interrogation of deep learning models is enabled by the interpretable nature of the TinNet framework that embeds a physically transparent, data-derived map** from neural state variables to output targets. We used the Shapley Additive exPlanations (SHAP)^42,43,44 that are built upon cooperative game theory for a local interpretation of reactivity origin of metal sites, by computing the SHAP value of each machine-learned feature, i.e., feature contribution to a prediction relative to a baseline or reference (see Methods). As shown in Fig. 5a, a local interpretation of Pt₃Ru relative to Pt and Pt₃Ru_1/2Co_1/2 to Pt₃Ru, taking *N_h as an example, suggests that the adsorbate resonance energies of the frontier N_2p orbitals are the most important factors in governing site reactivity instead of the commonly employed d-band center metric. Even though the d-band center of the Ru atom 2 and 3 in the Pt-Ru ensemble relative to pure Pt shifts up by +1.07 and +1.12 eV, respectively, it only leads to a –0.55 eV energy stabilization. In contrast, the adsorbate resonance energy of the *N 2p_z (2p_x and 2p_y) orbital shifts up by +0.96 (+0.56 and +0.62) eV at Pt₃Ru relative to that at Pt, which contributes –0.47 (–0.30 and –0.29) eV stabilization to N_h adsorption. Interestingly, the 3^rd metal Co at the atom site 3 (most stable position) weakens N_h adsorption by slightly shifting down the resonance energies of the N_2p orbitals, giving near-optimal reactivity descriptors. Local interpretable model-agnostic explanations (LIME) analysis was also performed, showing consistent interpretation where the variations in binding strengths can be largely explained due to the changes in the adsorbate resonance energies of frontier orbitals (Supplementary Fig. 23). Nevertheless, the factors that govern the surface reactivity, i.e., adsorption energies, are system dependent. In the case of Pt, Pt₃Ru, and Pt₃Ru_1/2Co_1/2, the variation of N_h adsorption energies can be explained by the resonance energies of adsorbate frontier orbitals. However, there also exist many systems in which the electronic structure of an active site, e.g., d-band center, dominates. This can be illustrated by SHAP analysis of Pt-terminated Pt₃Ru_1/2Ir_1/2 and Pt₃Ag surfaces relative to pure Pt, as shown in Supplementary Fig. 24. The effect of adsorbate resonance energies on chemical bonding can be illustrated through a graphic solution to the Newns–Anderson model in Fig. 5b. The upper intersect of an adsorbate line with the Hilbert-transformed chemisorption function represents the adsorbate-metal antibonding state. In accordance with the DFT-calculated density of states projected onto the N 2p_z orbital, an upshift (downshift) in the adsorbate resonance energy results in higher (lower) lying antibonding states pinning across the Fermi level, thus influencing the occupancy of those states and adsorption energies. Consequently, the Co plays an important role in tailoring the adsorbate resonance energies of frontier orbitals, pushing the site motif toward the top of the activity volcano. A global SHAP analysis for N_b adsorption models in Fig. 5c further consolidates a pivotal role of adsorbate resonance energies in chemical bonding across a diverse range of ternary Pt alloy surfaces.

**Fig. 5: Physicochemical insights from interpretable deep learning.**

Discussion

Theory integration into deep learning provides an opportunity to further advance the theory of chemisorption at metal surfaces by learning from data. Instead of assuming invariant adsorbate resonance energies of frontier orbitals upon a perturbation of a d-metal site, often realized through strain and/or ligand engineering, feature extraction algorithms by convolutional operations automatically learn the tunability of such a resonance state in multimetallic systems. The physical insights into chemical bonding at metal surfaces from interpretable deep learning have important consequences in rational design of improved catalytic systems, by manipulating a broad range of site attributes, e.g., electronegativity, number of d-electrons, d-orbital radius, and atomic size. With adsorbate resonance energies as tunable factors, many strategies can be envisioned, e.g., alkali metal promoters⁴⁵ and electron tunneling by forming a p-n junction diode⁴⁶. In fact, alkali promoters are often used in heterogeneous catalysis for modifying the catalytic activity of metal nanoparticles. In addition to the electrostatic and direct chemical bonding effects proposed previously⁶³ at an operating potential of 0.3 V_RHE with the ammonia activity approximated to be 0.1 (c_NH3 = 0.1 M). Adsorbate-adsorbate interactions were included for *NH₂ using a first order interaction model where the adsorbate interaction function was modeled via a smooth piecewise linear function. The *NH₂ adsorbate interaction parameter was calculated on Pt(100) by plotting the *NH₂ formation free energies as a function of coverage as shown in Supplementary Figure 3a. Within the microkinetic model, we allow for the reactions to occur on either the bridge or hollow sites. The deprotonation barriers of *HNNH and *NNH were set to a value of 0.25 eV since it was expected that these steps would never be rate limiting. For the activity map two descriptors were used which were the nitrogen bridge and hollow binding energies (${\varDelta E}_{N}$), which were both calculated with respect to a gas phase nitrogen atom in a box. Scaling relations between the formation free energies (at U = 0.3 V_RHE) of the adsorbates or transition state and these two descriptors are shown in Supplementary Figure 2.

Graph neural networks

All machine learning models in this work were based on graph neural networks (GNN). The GNN consists of two parts: convolutional network (CNN) and fully-connected neural network (FCNN). The convolutional network extracted high-level features from the graph feature representation of adsorbate-substrate systems, and the fully-connected neural network mapped high-level features to properties of interest or governing factors in the theory module. Five different graph neural networks were trained including two classification GNNs to predict the likelihood of surface reconstruction for bridge and hollow *N structures, and three regression models. One of the regression models was built to predict the formation energies of the *OH covered alloy surfaces, and two more regression models were built to predict the hollow and bridge *N binding energies. For the binding energy predictions, we used the TinNet that integrates GNNs with a theory module to allow for model interpretability. TinNet models used in this study adopted the Newns-Anderson-type model Hamiltonians in the theory module.

GNNs have 5 hyperparameters that require optimization, including the number of layers and neurons within the layer of convolutional network and fully-connected neural network and the learning rate. The Bayesian optimization algorithm in the Ray Tune package was used for hyperparameter optimization, which is a Python library for hyperparameter tuning at any scale^64,65. The regular 10-fold cross-validation method was adopted in the hyperparameter optimization and a total of 10 models were trained for each hyperparameters set. For each training, we used an 81/9/10 training/validation/test split. The validation set was used for early stop** to avoid overfitting the training set. The test set was used to evaluate the performance of the trained models. The hyperparameters set with the smallest average loss is the optimal hyperparameters set, which will be used in subsequent model training.

The nested 10-fold cross-validation was performed to avoid overfitting. The database was divided into 10 folds, of which one fold was used for validation, 1 fold was used for test, and the rest was used for training, where the validation and test can be the same fold and a total of 100 models were trained. The 90 models whose test set was not equal to the validation set were used to study the in-sample prediction performance of the models. Other 10 final models were used for the design space of ternary Pt alloys to demonstrate the out-of-sample predicting performance of models and to suggest candidates worthy of performing DFT calculations. Since this study focused on Pt-based catalysts, the pure Pt system was always included in the training set. The AdamW⁶⁶ was used as the optimizer to update weights and biases of models. To avoid getting stuck in local minima, the training set was split into multiple mini-batches where the size of each mini-batch was 64. The negative log likelihood (NLL) and the mean squared error (MSE) were used as loss functions for classification and regression problems, respectively. For TinNet models, the loss function contained not only the adsorption energy, but also the low-order d-band moments of the active site and projected density of states onto N2p orbitals.

SHAP analysis

Shapley values were computed by a model-independent kernel as implemented in the Shapley additive explanations (SHAP) (version 0.36.0) package to explain the governing electronic factors extracted by TinNet. The Shapley value concept was first established to measure a player’s importance within a team⁶⁷. With this idea, players would get a share of the overall benefit or reward based on how important was the role they played in the game’s result. Within the context of TinNet predicted binding energies, Shapley values can be used to find how the electronic factors learnt from TinNet, govern the binding energy for any given system. Shapley values estimate the feature importance, in terms of how much it impacts the final output (magnitude of SHAP values) as well as the direction in which it can influence the final output (sign of SHAP value). Specifically, for binding energy predictions, positive values indicate that the electronic feature contributed to increasing the adsorption energy positively, i.e., weakening of the bond whereas negative sign shows a contribution to strengthening the bond. Shapley values are computed as:

$${{{{{{\rm{\varphi }}}}}}}_{j}({{{{{{\rm{val}}}}}}})=\mathop{\sum}\limits_{S\subset \{1,\ldots,p\}{{\backslash }}\{j\}}\frac{\left|S\right|!\left(p-\left|S\right|-1\right)!}{p!}({{{{{{\rm{val}}}}}}}\left(S\cup \{j\}\right)-{{{{{{\rm{val}}}}}}}\left(S\right))$$

(1)

The function val(S) estimates the overall contribution of a given set of features (val(M) = f(x_i)), and the sum is across all potential subsets where feature i is not included.

Catalysts synthesis

GO was synthesized according to an improved Hummers method⁶⁸. A shape-directing agent-assisted reduction method was used to prepare rGO-supported Pt, Pt₃M or Pt₃Ru_1/2M_1/2 (M = Fe, Co or Ni) nanocubes, in which the metal mass percentage in all samples was controlled to be 20 wt%. Herein, the synthesis of rGO-supported Pt₃Ru_1/2Co_1/2 nanocubes was used as an example to simply summarize the synthesis details. Typically, 0.05 mmol of potassium tetrachloroplatinate (II) [K₂PtCl₄] (≥ 45% Pt, Aladdin), ruthenium chloride (III) [RuCl₃] (45-55% Ru, Aladdin) and cobalt chloride (II) [CoCl₂] (97% Co, Aladdin), as well as 34 mg of GO powder were added into a round-bottom flask that contains 20 mL of ethylene glycol (EG, Sinopharm Chemicals). The solution was then stirred for 30 min at room temperature to form a homogeneous mixture. Afterward, 0.75 mmol of tetramethylammonium bromide (C₄H₁₂BrN, Aladdin) was subjected into the mixture and sonicated for 5 min, then the flask was transferred into an oil bath. Thereafter, the reaction mixture was heated to 180 °C with stirring and condensing reflux for 30 min. After cooling down to room temperature, the final products were washed with deionized water at least three times and then freeze-dried at –50 °C for 24 h. Additionally, the iron and nickel precursors used are iron chloride (III) [FeCl₃] (≥ 99.9% Fe, Aladdin) and nickel chloride (II) [NiCl₂] (99.9% Ni, Aladdin), respectively.

Physical characterization

The overall particle size and distribution of the catalysts’ morphologies were characterized by scanning electron microscopy on JSM-7800 F microscope at a working voltage of 20 kV. The crystal phases in catalyst samples were analyzed by using powder XRD on a Philips 1730 diffractometer with Cu K_α radiation (λ = 1.5406 Å). XPS was conducted using a Thermo ESCALAB 250 spectrophotometer equipped with an Al K_α line source. Atomic-resolution HAADF images of atomically dispersed Pt sites in the ternary alloy nanocubes were captured in a FEI Theims Z STEM operated at 300 kV. TEM and HR-TEM images were performed on Tecnai G2-F30 equipped with an EDS. The ICP-MS was tested on NexION 300X Spectrometer to determine the metal content in the catalyst. Ultraviolet photoemission spectroscopy (UPS) were performed on an ESCALAB 250** electron spectrometer (Thermo Fisher Scientific) using a monochromatic Al Kα source (300 W).

Electrochemical measurements

For the ink preparation, the catalyst powder (2 mg, e.g., Pt₃Ru_1/2Co_1/2/rGO) was mixed with isopropanol (994 μL, Sinopharm Chemicals) and 5 wt% Nafion solution (6 μL, Sigma-Aldrich) under ultrasonication for about 30 min, and then the final ink was ready to use. Afterward, a 10 μL of the prepared ink was drop-cast on the glassy carbon electrode and dried in air to yield a thin-film electrode, which was ready for the electrochemical test. For all the studied catalysts, the Pt loading was fixed at about 8 μg cm⁻² for each tested electrode, estimated based on the overall Pt ratio within the catalyst (determined by ICP-MS). The AOR performance of the catalysts was measured by the rotating disk electrode (RDE, 5.0 mm in diameter) techniques with a CHI Electrochemical Station (Model 760E) in a three-electrode electrochemical cell. KCl-saturated Ag/AgCl electrode and a graphite rod were used as the reference and counter electrodes, respectively. Cyclic voltammetry measurements were firstly conducted in an Ar-saturated 1.0 M KOH between 0.05 and 1.1 V vs RHE at a scan rate of 100 mV s⁻¹ to activate the catalysts. The activated electrode was then measured in an Ar-saturated 1.0 M KOH + 0.1 M NH₃ solution at room temperature and an RDE rotation rate of 900 rpm in the same potential window. Note that the NH₃ solutions were prepared by mixing NH₃ solution (25 − 28%, RHAWN) with DI water and KOH. Chronoamperometry (CA) measurements were also conducted at 0.55 vs RHE for 1,800 s and a rotation rate of 900 rpm to further evaluate the AOR activity changes. During each AOR measurement, a fresh electrode was prepared and used to standardize the testing procedure.

Data availability

Currently, all source data for microkinetic modeling and machine learning that were used in this study are available from the GitHub repository: https://github.com/hlxin/aor.

Change history

20 February 2023
A Correction to this paper has been published: https://doi.org/10.1038/s41467-023-36746-z

References

Sinsel, S. R. Challenges and solution technologies for the integration of variable renewable energy sources—a review. Renew. Energy 145, 2271–2285 (2020).
Katsounaros, I. et al. On the mechanism of the electrochemical conversion of ammonia to dinitrogen on Pt(1 0 0) in alkaline environment. J. Catal. 359, 82–91 (2018).
Article CAS Google Scholar
Pillai, H.; **n, H. New insights into electrochemical ammonia oxidation on Pt(100) from first principles. Ind. Eng. Chem. Res. https://doi.org/10.1021/acs.iecr.9b01471 (2019).
Vidal-Iglesias, F. J., Solla-Gullón, J., Feliu, J. M., Baltruschat, H. & Aldaz, A. DEMS study of ammonia oxidation on platinum basal planes. J. Electroanal. Chem. 588, 331–338 (2006).
Article CAS Google Scholar
Rosca, V. & Koper, M. T. M. Electrocatalytic oxidation of ammonia on Pt(111) and Pt(100) surfaces. Phys. Chem. Chem. Phys. 8, 2513–2524 (2006).
Article CAS PubMed Google Scholar
Gerischer, H. & Mauerer, A. Untersuchungen zur anodischen oxidation von ammoniak an platin-elektroden. J. Electroanal. Chem. Interfacial Electrochem. 25, 421–433 (1970).
Article CAS Google Scholar
Matsui, T. et al. In situ attenuated total reflection infrared spectroscopy on electrochemical ammonia oxidation over Pt electrode in alkaline aqueous solutions. Langmuir 31, 11717–11723 (2015).
Article CAS PubMed Google Scholar
Siddharth, K. et al. Surface engineering in improving activity of Pt nanocubes for ammonia electrooxidation reaction. Appl. Catal. B 269, 118821 (2020).
Article CAS Google Scholar
Sacré, N. et al. Tuning Pt–Ir interactions for NH3 electrocatalysis. ACS Catal. 8, 2508–2518 (2018).
Article Google Scholar
Song, L. et al. Temperature-dependent kinetics and reaction mechanism of ammonia oxidation on Pt, Ir, and PtIr alloy catalysts. J. Electrochem. Soc. 165, J3095–J3100 (2018).
Article CAS Google Scholar
Hung, C.-M. Electrochemical properties of PtPdRh alloy catalysts for ammonia electrocatalytic oxidation. Int. J. Hydrog. Energy 37, 13815–13821 (2012).
Article CAS Google Scholar
Lin, X. et al. Hyperbranched concave octahedron of PtIrCu nanocrystals with high-index facets for efficiently electrochemical ammonia oxidation reaction. J. Colloid Interface Sci. 601, 1–11 (2021).
Article ADS CAS PubMed Google Scholar
Li, Y. et al. High-performance ammonia oxidation catalysts for anion-exchange membrane direct ammonia fuel cells. Energy Environ. Sci. 14, 1449–1460 (2021).
Article CAS Google Scholar
Li, Y. et al. Ternary PtIrNi catalysts for efficient electrochemical ammonia oxidation. ACS Catal. 10, 3945–3957 (2020).
Article CAS Google Scholar
Medford, A. J. et al. From the Sabatier principle to a predictive theory of transition-metal heterogeneous catalysis. J. Catal. 328, 36–42 (2015).
Article CAS Google Scholar
Nørskov, J. K., Abild-Pedersen, F., Studt, F. & Bligaard, T. Density functional theory in surface chemistry and catalysis. Proc. Natl Acad. Sci. USA 108, 937–943 (2011).
Article ADS PubMed PubMed Central Google Scholar
Zhao, Z.-J. et al. Theory-guided design of catalytic materials using scaling relationships and reactivity descriptors. Nat. Rev. Mater. 4, 792–804 (2019).
Article ADS Google Scholar
Greeley, J. et al. Alloys of platinum and early transition metals as oxygen reduction electrocatalysts. Nat. Chem. 1, 552–556 (2009).
Article CAS PubMed Google Scholar
Wang, S., Pillai, H. S. & **n, H. Bayesian learning of chemisorption for bridging the complexity of electronic descriptors. Nat. Commun. 11, 6132 (2020).
Article ADS CAS PubMed PubMed Central Google Scholar
Palkovits, R. & Palkovits, S. Using artificial intelligence to forecast water oxidation catalysts. ACS Catal. 9, 8383–8387 (2019).
Jiang, X., Wang, Y., Jia, B., Qu, X. & Qin, M. Prediction of oxygen evolution activity for NiCoFe oxide catalysts via machine learning. ACS Omega 7, 14160–14164 (2022).
Article CAS PubMed PubMed Central Google Scholar
Zhong, M. et al. Accelerated discovery of CO₂ electrocatalysts using active machine learning. Nature 581, 178–183 (2020).
Article ADS CAS PubMed Google Scholar
Gao, Q. et al. Breaking adsorption-energy scaling limitations of electrocatalytic nitrate reduction on intermetallic CuPd nanocubes by machine-learned insights. Nat. Commun. 13, 1–12 (2022).
ADS Google Scholar
Gu, G. H. et al. Autobifunctional mechanism of jagged Pt nanowires for hydrogen evolution kinetics via end-to-end simulation. J. Am. Chem. Soc. 143, 5355–5363 (2021).
Article CAS PubMed Google Scholar
Lee, J. & **nouchi, R. Machine learning-based screening of highly stable and active ternary Pt alloys for oxygen reduction reaction. J. Phys. Chem. C https://doi.org/10.1021/acs.jpcc.1c02890 (2021).
Li, Z., Achenie, L. E. K. & **n, H. An adaptive machine learning strategy for accelerating discovery of perovskite electrocatalysts. ACS Catal. 10, 4377–4384 (2020).
Article CAS Google Scholar
Tran, K. & Ulissi, Z. W. Active learning across intermetallics to guide discovery of electrocatalysts for CO₂ reduction and H₂ evolution. Nat. Catal. 1, 696–703 (2018).
Article CAS Google Scholar
**e, T. & Grossman, J. C. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Phys. Rev. Lett. 120, 145301 (2018).
Article ADS CAS PubMed Google Scholar
Wang, S.-H., Pillai, H. S., Wang, S., Achenie, L. E. K. & **n, H. Infusing theory into deep learning for interpretable reactivity prediction. Nat. Commun. 12, 5288 (2021).
Article ADS CAS PubMed PubMed Central Google Scholar
Akhade, S. A., Bernstein, N. J., Esopi, M. R., Regula, M. J. & Janik, M. J. A simple method to approximate electrode potential-dependent activation energies using density functional theory. Catal. Today 288, 63–73 (2017).
Article CAS Google Scholar
Janik, M. J., McCrum, I. T. & Koper, M. T. M. On the presence of surface bound hydroxyl species on polycrystalline Pt electrodes in the “hydrogen potential region” (0–0.4 V-RHE). J. Catal. 367, 332–337 (2018).
Article CAS Google Scholar
Wallace, S. W., McCrum, I. T. & Janik, M. J. Ammonia electro-oxidation mechanism on the platinum (100) surface. Catal. Today https://doi.org/10.1016/j.cattod.2020.09.024 (2020).
Ye, J.-Y. et al. Ammonia electrooxidation on dendritic Pt nanostructures in alkaline solutions investigated by in-situ FTIR spectroscopy and online electrochemical mass spectroscopy. J. Electroanal. Chem. 819, 495–501 (2018).
Article CAS Google Scholar
Elnabawy, A. O., Herron, J. A., Karraker, S. & Mavrikakis, M. Structure sensitivity of ammonia electro-oxidation on transition metal surfaces: a first-principles study. J. Catal. https://doi.org/10.1016/j.jcat.2021.03.010 (2021).
Jain, A. et al. Commentary: The Materials Project: a materials genome approach to accelerating materials innovation. APL Mater. 1, 011002 (2013).
Article ADS Google Scholar
Newns, D. M. Self-consistent model of hydrogen chemisorption. Phys. Rev. 178, 1123–1135 (1969).
Article ADS CAS Google Scholar
Liu, Z. et al. Surface structure engineering of PtPd nanoparticles for boosting ammonia oxidation electrocatalysis. ACS Appl. Mater. Interfaces https://doi.org/10.1021/acsami.2c04711 (2022).
Article PubMed PubMed Central Google Scholar
Lu, B.-A. et al. Structurally disordered phosphorus-doped Pt as a highly active electrocatalyst for an oxygen reduction reaction. ACS Catal. 11, 355–363 (2021).
Article CAS Google Scholar
Li, Z.-F., Wang, Y. & Botte, G. G. Revisiting the electrochemical oxidation of ammonia on carbon-supported metal nanoparticle catalysts. Electrochim. Acta 228, 351–360 (2017).
Finkelstein, D. A., Bertin, E., Garbarino, S. & Guay, D. Mechanistic similarity in catalytic N₂ production from NH₃ and NO_2– at Pt(100) thin films: toward a universal catalytic pathway for simple N-containing species, and its application to in situ removal of NH₃ poisons. J. Phys. Chem. C: Nanomater. Interfaces 119, 9860–9878 (2015).
Article CAS Google Scholar
Kim, H., Chung, M. W. & Choi, C. H. NOx-induced deactivation of Pt electrocatalysis towards the ammonia oxidation reaction. Electrochem. Commun. 94, 31–35 (2018).
Article CAS Google Scholar
Lundberg, Scott. n.d. Shap: A Game Theoretic Approach to Explain the Output of Any Machine Learning Model. Github https://github.com/slundberg/shap (2022).
Lundberg, S. M. & Lee, S.-I. A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems; NIPS’17, 4768–4777 (Curran Associates Inc., 2017).
Muscat, J. P. & Newns, D. M. Chemisorption on metals. Prog. Surf. Sci. 9, 1–43 (1978).
Article ADS CAS Google Scholar
Resasco, J. et al. Promoter effects of alkali metal cations on the electrochemical reduction of carbon dioxide. J. Am. Chem. Soc. 139, 11277–11287 (2017).
Article CAS PubMed Google Scholar
Winkler, J. R. & Gray, H. B. Long-range electron tunneling. J. Am. Chem. Soc. 136, 2930–2939 (2014).
Article CAS PubMed PubMed Central Google Scholar
**n, H. & Linic, S. Analyzing relationships between surface perturbations and local chemical reactivity of metal sites: alkali promotion of O₂ dissociation on Ag(111). J. Chem. Phys. 144, 234704 (2016).
Article ADS PubMed Google Scholar
Liu, Z.-P. & Hu, P. An insight into alkali promotion: a density functional theory study of CO dissociation on K/Rh(111). J. Am. Chem. Soc. 123, 12596–12604 (2001).
Article CAS PubMed Google Scholar
Li, J., Zhang, Y. & Kornienko, N. Heterogeneous electrocatalytic reduction of CO₂ promoted by secondary coordination sphere effects. N. J. Chem. 44, 4246–4252 (2020).
Article Google Scholar
Wu, X. et al. Engineering the coordination sphere of isolated active sites to explore the intrinsic activity in single-atom catalysts. Nanomicro Lett. 13, 136 (2021).
ADS CAS PubMed PubMed Central Google Scholar
Vijay, S., Kastlunger, G., Chan, K. & Nørskov, J. K. Limits to scaling relations between adsorption energies? J. Chem. Phys. 156, 231102 (2022).
Article ADS CAS PubMed Google Scholar
Saini, S., Halldin Stenlid, J. & Abild-Pedersen, F. Electronic structure factors and the importance of adsorbate effects in chemisorption on surface alloys. npj Comput. Mater. 8, 1–12 (2022).
Article Google Scholar
Vojvodic, A. & Nørskov, J. K. New design paradigm for heterogeneous catalysts. Natl Sci. Rev. 2, 140–143 (2015).
Article CAS Google Scholar
Pérez-Ramírez, J. & López, N. Strategies to break linear scaling relationships. Nat. Catal. 2, 971–976 (2019).
Article Google Scholar
Kresse, G. & Furthmüller, J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B: Condens. Matter 54, 11169–11186 (1996).
Article ADS CAS PubMed Google Scholar
Hammer, B., Hansen, L. B. & Nørskov, J. K. Improved adsorption energetics within density-functional theory using revised Perdew-Burke-Ernzerhof functionals. Phys. Rev. B: Condens. Matter 59, 7413–7421 (1999).
Article ADS Google Scholar
Mathew, K., Sundararaman, R., Letchworth-Weaver, K., Arias, T. A. & Hennig, R. G. Implicit solvation model for density-functional study of nanocrystal surfaces and reaction pathways. J. Chem. Phys. 140, 084106 (2014).
Article ADS PubMed Google Scholar
Mathew, K., Kolluru, V. S. C., Mula, S., Steinmann, S. N. & Hennig, R. G. Implicit self-consistent electrolyte model in plane-wave density-functional theory. J. Chem. Phys. 151, 234101 (2019).
Henkelman, G., Uberuaga, B. P. & Jonsson, H. A climbing image nudged elastic band method for finding saddle points and minimum energy paths. J. Chem. Phys. 113, 9901–9904 (2000).
Article ADS CAS Google Scholar
Nie, X., Esopi, M. R., Janik, M. J. & Asthagiri, A. Selectivity of CO₂ reduction on copper electrodes: the role of the kinetics of elementary steps. Angew. Chem. Int. Ed. 52, 2459–2462 (2013).
Article CAS Google Scholar
Nie, X., Esopi, M. R., Janik, M. J. & Asthagiri, A. Selectivity of CO₂ reduction on copper electrodes: the role of the kinetics of elementary steps. Angew. Chem. Int. Ed. 52, 2459–2462 (2013).
Article CAS Google Scholar
Nørskov, J. K. et al. Origin of the overpotential for oxygen reduction at a fuel-cell cathode. J. Phys. Chem. B 108, 17886–17892 (2004).
Article Google Scholar
Medford, A. J. et al. CatMAP: a software package for descriptor-based microkinetic map** of catalytic trends. Catal. Lett. 145, 794–807 (2015).
Article CAS Google Scholar
Liaw, R. et al. Tune: a research platform for distributed model selection and training. Preprint at https://arxiv.org/abs/1807.05118 (2018).
Fernando Nogueira. n.d. BayesianOptimization: A Python Implementation of Global Optimization with Gaussian Processes. Github https://github.com/fmfn/BayesianOptimization (2022).
Loshchilov, I. & Hutter, F. Decoupled Weight Decay Regularization. Preprint at http://arxiv.org/abs/1711.05101 (2017).
Shapley, L. in Classics in Game Theory (ed. Kuhn, H. W.) 69–79 (Princeton University Press, 2020).
Marcano, D. C. et al. Improved synthesis of graphene oxide. ACS Nano 4, 4806–4814 (2010).
Article CAS PubMed Google Scholar

Download references

Acknowledgements

H.S.P., S.H.W., N.O., Q.M., L.E.K.A., and H.X. acknowledge the financial support from the NSF CAREER program (CBET-1845531). H.S.P. specially thanks the Non-Academic Research Internships for Graduate Students (INTERN) program for supporting part the of work in SUNCAT catalysis center under the guidance of F.A.-P. The computational resource used in this work is provided by the advanced research computing at Virginia Polytechnic Institute and State University. Y.L. thanks the financial support from the Natural Science Foundation of Jiangsu Province (SBK20210769) and China Postdoctoral Science Foundation (2022M711543). G.W. thanks the support from the University at Buffalo, SUNY. F.A-P. would like to acknowledge support from the U.S. DOE, Office of Science, Office of Basic Energy Sciences, Chemical Sciences, Geosciences, and Biosciences Division, Catalysis Science Program to the SUNCAT Center for Interface Science and Catalysis.

Author information

Authors and Affiliations

Department of Chemical Engineering, Virginia Polytechnic Institute and State University, Blacksburg, VA, USA
Hemanth Somarajan Pillai, Shih-Han Wang, Noushin Omidvar, Qingmin Mu, Luke E. K. Achenie & Hongliang **n
School of Materials Science and Engineering, Jiangsu University, Zhenjiang, Jiangsu, China
Yi Li & Juan Yang
SUNCAT Center for Interface Science and Catalysis, SLAC National Accelerator Laboratory, Menlo Park, CA, USA
Frank Abild-Pedersen
Department of Chemical and Biological Engineering, University at Buffalo, The State University of New York, Buffalo, NY, USA
Gang Wu

Authors

Hemanth Somarajan Pillai
View author publications
You can also search for this author in PubMed Google Scholar
Yi Li
View author publications
You can also search for this author in PubMed Google Scholar
Shih-Han Wang
View author publications
You can also search for this author in PubMed Google Scholar
Noushin Omidvar
View author publications
You can also search for this author in PubMed Google Scholar
Qingmin Mu
View author publications
You can also search for this author in PubMed Google Scholar
Luke E. K. Achenie
View author publications
You can also search for this author in PubMed Google Scholar
Frank Abild-Pedersen
View author publications
You can also search for this author in PubMed Google Scholar
Juan Yang
View author publications
You can also search for this author in PubMed Google Scholar
Gang Wu
View author publications
You can also search for this author in PubMed Google Scholar
Hongliang **n
View author publications
You can also search for this author in PubMed Google Scholar

Contributions

H.S.P., S.H.W., L.E.K.A., F.A.-P., and H.X. conceived the idea and designed the computational approach. H.S.P., S.H.W., Q.M., and N.O. conducted DFT calculations and machine learning model development. H.S.P., S.H.W., and N.O. performed the detailed model interpretation. Y.L., J.Y., and G.W. performed all experiments and carried out data analyses for the validation of model predictions. All authors contributed to the writing and editing of this manuscript.

Corresponding authors

Correspondence to Yi Li, Gang Wu or Hongliang **n.

Ethics declarations

Competing interests

The authors declare no competing interests.

Peer review

Peer review information

Nature Communications thanks Byungchan Han, Hector Prats and the other anonymous reviewer(s) for their contribution to the peer review of this work. Peer review reports are available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary Information

Peer review file

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Cite this article

Pillai, H.S., Li, Y., Wang, SH. et al. Interpretable design of Ir-free trimetallic electrocatalysts for ammonia oxidation with graph neural networks. Nat Commun 14, 792 (2023). https://doi.org/10.1038/s41467-023-36322-5

Download citation

Received: 14 December 2022
Accepted: 24 January 2023
Published: 11 February 2023
DOI: https://doi.org/10.1038/s41467-023-36322-5
Springer Nature Limited

Interpretable design of Ir-free trimetallic electrocatalysts for ammonia oxidation with graph neural networks

From

Abstract

Similar content being viewed by others

Adsorbate chemical environment-based machine learning framework for heterogeneous catalysis

Ab Initio to Activity: Machine Learning-Assisted Optimization of High-Entropy Alloy Catalytic Activity

Investigating Periodic Table Interpolation for the Rational Design of Nanoalloy Catalysts for Green Hydrogen Production from Ammonia Decomposition

Introduction

Advancing theory of chemisorption with interpretable deep learning

Discussion

Graph neural networks

SHAP analysis

Catalysts synthesis

Physical characterization

Electrochemical measurements

Data availability

Change history

20 February 2023

References

Acknowledgements

Author information

Authors and Affiliations

Contributions

Corresponding authors

Ethics declarations

Competing interests

Peer review

Peer review information

Additional information

Supplementary information

Supplementary Information

Peer review file

Rights and permissions

About this article

Cite this article

Navigation

Interpretable design of Ir-free trimetallic electrocatalysts for ammonia oxidation with graph neural networks

From

Abstract

Similar content being viewed by others

Adsorbate chemical environment-based machine learning framework for heterogeneous catalysis

Ab Initio to Activity: Machine Learning-Assisted Optimization of High-Entropy Alloy Catalytic Activity

Investigating Periodic Table Interpolation for the Rational Design of Nanoalloy Catalysts for Green Hydrogen Production from Ammonia Decomposition

Introduction

Advancing theory of chemisorption with interpretable deep learning

Discussion

Graph neural networks

SHAP analysis

Catalysts synthesis

Physical characterization

Electrochemical measurements

Data availability

Change history

20 February 2023

References

Acknowledgements

Author information

Authors and Affiliations

Contributions

Corresponding authors

Ethics declarations

Competing interests

Peer review

Peer review information

Additional information

Supplementary information

Supplementary Information

Peer review file

Rights and permissions

About this article

Cite this article

Share this article

Search

Navigation