Abstract
Biatom catalysts (BACs) are well-known as promising candidates for the oxygen reduction reaction (ORR). However, the vast design space of the BACs causes a high cost for experimental and computational screening. Herein, by employing density functional theory (DFT) calculations and machine-learning (ML) algorithms, we propose a data-driven high-throughput design approach for the rapid screening of ORR catalysts consisting of 3d, 4d, and 5d transition metal dimers (TM2) inserted in pristine and heteroatom (S, C, B, P)-doped g-CN support with excellent activity. According to the DFT calculations and ML predictions, we conclude that the BACs of Ag2/Pd2-C-I, Pt2-B-I, Pd2-P-I, and Pd2/Ag2-C-II can serve as promising candidates for ORR catalysts. The ML analysis reveals that the intrinsic properties of the metal active center play a dominant role in the catalytic activity of BACs, accounting for 71.6% of the total feature importance, while the coordination environment contributes to a lesser extent of 28.4%. Thus, based on the suitable metal active center, doping C/B/P heteroatoms on the g-CN support is an effective way to further improve the catalytic activity. A bridge from intrinsic properties to catalytic activity has been established in this work, which is conducive to unveiling the origin of the ORR activity and accelerating the screening process for high-activity ORR catalysts.


1. Introduction
To solve the energy crisis and environmental pollution caused by the tremendous consumption of nonrenewable energy, exploring technologies of clean energy conversion and storage has become a hot research topic. , These energy conversion devices depend heavily on electrocatalysts with high catalytic activity toward the oxygen reduction reaction (ORR). , Therefore, extensive efforts have been dedicated to developing efficient catalysts in order to break the limitation of the sluggish kinetics for ORR on the cathode. , Since the proposal of single-atom catalysts (SACs), which have attracted widespread attention and are regarded as a promising alternative to traditional noble-based catalysts because of their low cost and high activity, − although numerous SACs have been reported experimentally as high-performance electrocatalysts toward the ORR, − SACs still confront severe challenges. First, SACs may suffer an unsatisfactory performance due to the simple active sites and low metal loading for complex reactions; − second, achieving both high catalytic activity and high selectivity on the isolated site is rather difficult.
Recently, biatom catalysts (BACs), formed by introducing an additional metal atom into SACs, have been considered as a desirable approach to enhance the catalytic activity and selectivity of SACs. − For instance, the BAC with Co-Ni dual sites in N-doped carbon was reported to exhibit higher activity for the oxygen evolution reaction (OER) and ORR than SACs. Zhang et al. synthesized a locally distributed atomic Pt-Co N-carbon-based catalyst, which not only possesses outstanding activity and durability for ORR but also shows a high selectivity for the 4e– pathway. Furthermore, codoping of heteroatoms (such as N, S, C, B, P, etc.) to regulate the coordination environment of the metal atom is also an effective method to promote the catalytic performance. − The Fe-N3P SAC formed by Fe inserted in P, N codoped carbon support offers superior electrochemical performance relative to the Fe-N4 and Pt/C catalysts. A novel bifunctional catalyst of FeCo alloy embedded in an N, S dual-doped carbon matrix (FeCo/NSC) was prepared by Chang et al., showing high ORR and OER activities.
Benefiting from the high conductivity, large specific surface area, and low cost, carbon-based 2D materials have been used as promising substrates for many reactions. − Especially, carbon materials with inherent hole sites provide abundant N/C coordination atoms to capture foreign metal atoms such as C2N, g-CN, g-C3N4, and graphdiyne, ,− arousing a great deal of research enthusiasm. However, there are also some restrictions on the application of carbon-based catalysts. The various edge types, active sites, and support types lead to a vast space of catalysts; − the trial-and-error approach commonly used in experiments is expensive and less efficient. Developing an efficient and low-cost design strategy is crucial to boosting the screening process of carbon-based catalysts. On the other hand, an in-depth understanding of the interactions between the metal atom and the coordination environment on the substrate is also significant for exploring novel catalysts with high activity and selectivity. In this regard, machine learning (ML) is an efficient method to accelerate the screening of electrocatalytic materials and illustrate the intrinsic relationships between the material intrinsic properties and catalytic activities. For example, using an ML model, the hydrogen evolution reaction (HER) and ORR performance on carbon-based SACs can be quickly predicted from the electronic, geometric, and thermodynamic descriptors. ,,
Herein, we constructed a total of 224 BACs on both pristine g-CN and S-, B-, C-, and P-doped g-CN and investigated their ORR performances using density functional theory (DFT) calculations and ML. The thermodynamic stability of the BACs is assessed by using the formation energy. Moreover, feature importance analysis clarifies the contributions of the metal active center and environmental atoms to the catalytic activity. Electronic structure calculations reveal that the doping of heteroatoms can regulate the occupancy of bonding and antibonding states by increasing the energy level of the d-band center for catalysts, thereby enhancing the adsorption strength of ORR intermediates and consequently improving the ORR activity.
2. Computational Details
The calculations of structure optimization and catalytic performance used the spin-polarized density functional theory (DFT) method, as implemented in the Vienna ab initio Simulation Package (VASP). , The projector augmented wave (PAW) pseudopotentials were adopted to represent the ion–electron interaction. The cutoff energy of the plane-wave basis set was chosen to be 500 eV. The Perdew–Burke–Ernzerhof (PBE) functional of the generalized gradient approximation (GGA) was employed to treat the exchange–correlation interaction. ,
In the process of structure optimization, a 2 × 2 × 1 g-CN supercell was constructed, and the positions of the atoms, as well as the shape and size of the structure, were fully relaxed using the conjugate gradient method. To avoid the periodic interactions, a vacuum layer of 20 Å along the z-axis was applied. The 2 × 2 × 1 Monkhorst–Pack meshes were used for sampling the Brillouin zones. The results of K-point convergence tests were listed in Table S1. The convergence criteria for the maximum force on each atom were set to 0.02 eV Å–1. When the energy change between the two iterated steps is less than 10–5 eV, the convergence of the total energy was achieved. The empirical correction in Grimme’s scheme (DFT-D3) was selected to describe the van der Waals (vdW) interaction. The dipole correction was applied to account for the dipole moment induced by asymmetric doping. The implicit solvation model implemented in VASPsol was used to treat the solvent effect, and a dielectric constant of 80 was set to simulate the water environment.
The ORR proceeds along 4e– pathway in acidic media (pH = 0) can be expressed as
| 2.1 |
The four elementary reaction steps are given as follows:
| 2.2 |
| 2.3 |
| 2.4 |
| 2.5 |
in which * denotes the support, (g) and (l) are the gas and liquid phases, respectively.
The Gibbs free energy changes (ΔG) for each elementary step were calculated with the computational hydrogen electrode (CHE) model proposed by Nørskov et al. The free energy changes were obtained by the following equation:
| 2.6 |
where ΔE is the adsorption energy of the intermediates of ORR based on the DFT calculations, ΔZPE represents the zero-point energy correction; T and ΔS are the temperature (T = 298.15 K) and the entropy change; ΔG U and ΔG pH denote the applied electrode potential and the correction of Gibbs free energy by pH, respectively.
The overpotentials for the ORR (ηORR) can be calculated using the eqs :
| 2.7 |
where ΔG 1, ΔG 2, ΔG 3, and ΔG 4 are the reaction free energies of the four elementary reaction steps as eqs –, respectively. Among them, the elementary reaction step with maximum reaction free energy was denoted as the potential-determining step (PDS).
3. Results and Discussion
3.1. Geometric Structures and Stability
As shown in Figure a, the 2 × 2 × 1 g-CN supercell possesses a native N-edge hexagonal cavity with a circumcircle diameter of 5.47 Å (the distance between the two farthest N atoms in the cavity), which is consistent with previous work. , For all 3d, 4d, and 5d TM dimers (denoted as TM2 and excluding Tc2, Cd2, and Hg2) embedded in the cavity of the pristine g-CN, two adsorption configurations involving each metal atom coordinated to N2 and N3 were considered (TM2-N2 and TM2-N3). It is worth noting that the optimized TM2-N3 systems exhibit two distinct configurations with respect to the three-N coordination environment (Figure S1). Among them, the systems of Co2-N3, Cu2-N3, Fe2-N3, Rh2-N3, Ru2-N3, and Pd2-N3 adopt the Type-II configuration after structural optimization. Almost all TM2 embedded in g-CN are more stable with the TM2-N3 than the TM2-N2 configuration, except for the Zn2-dimer (where the Zn atom prefers to bind with two N atoms). Owing to their large atomic radii, the doped Sc2, Zn2, Y2, Zr2, Mo2, Ag2, W2, Re2, Os2, Ir2, and Au2 atoms tend to protrude from the g-CN plane after structural optimization, resulting in pronounced structural undulation of the substrate, whereas the remaining TM2-N3 BACs retain a planar configuration. The distance between the two TM atoms (TM-TM) and the lengths of the TM-N coordination bonds in TM2-N3 are summarized in Table S2, where the TM-TM bond lengths vary from 1.82 to 3.04 Å, and the length of the coordination bonds for the two TM atoms is consistent (TM1-N1 = TM2-N4, TM1-N2 = TM2-N5, TM1-N3 = TM2-N6). In addition, the insertion of TM2 induces the variation in the size and shape of the cavity in g-CN. The geometric information on the altered cavities and the lattice constants of TM2-N3 are shown in Table S3, and the size of the cavity is characterized by the distance between the two farthest N atoms in the cavity (denoted as D). For TM2-N3, the lattice constants are in the range of 14.19–14.33 Å (a = b), and the D of the cavities (D = 5.57–6.65 Å) is much larger than that in the pristine g-CN.
1.

Model of (a) metal atoms coordinated with two nitrogen atoms (TM2-N2) and three nitrogen atoms (TM2-N3) in a g-CN sheet. (b) TM2 embedded in S/C/B/P-doped g-CN sheet with different configurations. (c) The stability of TM2-N3 and TM2- S/C/B/P-I BACs.
Based on the TM2-N3 configuration, doping with S, C, P, and B atoms is explored. We only replaced one N atom with a heteroatom, because embedding more heteroatoms with large atomic radii may lead to distortion and instability of the structures. As shown in Figure b, there are two configurations for one heteroatom doping (TM2-S/C/B/P-I and TM2-S/C/B/P-II). All doping configurations exhibit positive substitution formation energies (Table S4), indicating that heteroatom substitution is endothermic relative to the pristine g-CN lattice and requires external energy input. This is consistent with the fact that these doped g-CN structures are experimentally synthesized via high-temperature thermal polymerization/copolymerization routes, which supply the thermal energy to overcome the energy barriers. The structural parameters, including bond lengths, D, and lattice constants of TM2-S/C/B/P-I and TM2-S/C/B/P-II are listed in Tables S5–S20. Different from the lattice constants of a = b in TM2-N3, there is a significant difference between the lattice constants a and b in TM2-S/C/B/P-I and TM2-S/C/B/P-II systems (e.g., Sc2–S-I with a = 14.34 Å and b = 14.13 Å) due to their disrupted structural symmetry caused by heteroatom doping. Meanwhile, the difference between the lattice constants a and b in TM2-S/C/B/P-I is more apparent than that in TM2-S/C/B/P-II due to the distinct doping positions of the heteroatoms. As for the Zn2 embedded in g-CN and heteroatom-doped g-CN, the most stable Zn2-N2 and Zn2–S/C/B/P-I (Zn2–S/C/B/P-II) structures are shown in Figure S2.
The diffusion and aggregation of the metal atoms on the substrate will significantly reduce the catalytic activity and durability of materials. Thus, we evaluated the stability of TM2-N3 and TM2-S/C/B/P-I by calculating the formation energies (E f) of TM2 on different supports. The E f can be defined as follows:
| 3.1 |
in which and E sub are the energies of the substrate with and without TM2 doping, and E M is the energy of the metal atom in its most stable phase. Obviously, the clustering of TM atoms on the surface is unfavorable, and the system is stable when E f is negative. The calculated E f of TM2-N3, TM2-S-I and TM2-C-I, TM2-B-I and TM2-P-I is plotted in Figure c. It can be found that among all the systems, the stabilities of TM2-C-I systems are better than those of the others. Depending on the doped metal, the stability of these systems is in the order 3d > 4d > 5d. With an increase in the atomic number of the TM atom, the stability of all BACs initially decreases and then increases in each period. Thus, the early-TM dimers and late-TM dimers such as Sc2, Ti2, V2, Zn2, Y2, Zr2, Nb2, Pd2, Hf2, and Pt2 can be tightly embedded in all supports. In particular, for TM2-C-I systems, all the 3d and 4d TMs embedded in the substrate are thermodynamically stable, while the positive values of E f for 5d TMs of Ta2, W2, Re2, Os2, and Au2 indicate that these TMs are prone to aggregation on the C-doped g-CN support. The TM2-N3 and TM2-S-I systems exhibit comparable stability. The 3d TMs, including Sc2, Ti2, V2, Cr2, Mn2, Ni2, Cu2, and Zn2 on g-CN, and Sc2, Ti2, V2, Co2, Ni2, and Zn2 on S-doped g-CN, are identified to be stable. Similarly, the 4d TMs such as Y2, Zr2, Nb2 and Pd2 on both g-CN and S-doped g-CN, and 5d TMs, specifically Pt2 on g-CN (as well as Hf2 and Pt2 on S-doped g-CN), are also found to be stable.
3.2. ORR Catalytic Activity
In this section, the ORR catalytic performance of TM2-N3 and TM2-S/C/B/P-I systems has been investigated. Following the research of Nørskov et al., the adsorption free energies of oxygenated intermediates OH*, O*, and OOH* (ΔG OH*, ΔG O*, and ΔG OOH*) determine the reaction free energies of the four elementary steps in the complete ORR process. Herein, adsorption free energies of all intermediates have been calculated, and the Gibbs free energy change of each step, along with the corresponding overpotential for ORR (ηORR) on TM2-N3 and TM2-S/C/B/P-I systems has been calculated and listed in Tables S21–S25. We first focus on the potential-determining step (PDS) for each system, which is a critical factor affecting the ORR overpotential. The last step, OH* → H2O, is the PDS for most TM2-N3 and TM2-S/C/B/P-I systems. In addition, the PDS for Pd2/Os2-N3, Pd2/Pt2-S-I, Pd2/Au2-C-I, Au2-B-I, and Pd2/Ag2/Pt2-P-I systems is the first step (O2 → OOH*), while for Ti2/Zr2/Ag2-N3, Mn2-B-I, and Nb2/Mo2/Ir2-P-I systems, it is the second step (OOH* → O*). For Sc2/Zn2/Pt2-N3, Os2-S-I, Pt2-C-I, Ir2-B-I, and Mn2-P-I systems, the PDS becomes the third step (O* → OH*).
A lower value of ηORR indicates higher catalytic activity of the BAC. Among all studied BACs, those with late-TM active centers exhibit better catalytic performance than those with early-TM centers. Several catalysts demonstrating superior performance in each system were selected and illustrated in Figure a. For TM2-N3 systems, Pt2-N3 shows the best performance with an ηORR of 0.75 V, which is much higher than that of Pt (0.45 V). This indicates that the pristine g-CN is not a suitable support for ORR BACs. As for the systems with heteroatom doping, some BACs stand out with ηORR comparable to that of the ideal oxygen electrocatalyst Pt, such as Ag2/Au2/Pd2-C-I, Ag2/Pt2-B-I, and Pd2-P-I. Therefore, doping with heteroatoms such as C, B, and P in TM2-N3 systems has proved to be an effective strategy to further enhance the catalytic activity of metal active centers. In contrast, S doping has a negligible effect on the catalytic activity. The ORR free energy diagrams of the BACs with excellent catalytic performance are plotted in Figure b–d to further clarify the reaction pathway. Obviously, the free energy diagrams for the elementary reduction steps are downhill at U = 0 V, indicating spontaneous ORR processes. Upon applying an electrode potential of U = 1.23 V to Au2/Pd2-C-I and Pd2-P-I BACs, the transition from O2 to OOH* results in a maximal Gibbs free energy increase of 0.40, 0.34, and 0.50 eV, respectively, corresponding to ORR overpotentials of 0.40, 0.34, and 0.50 V for these BACs. At the same electrode potential for Ag2-C-I and Ag2/Pt2-B-I BACs, the last step yields the maximum Gibbs free energy change, leading to ηORR of 0.43, 0.44, and 0.49 V for Ag2-C-I and Ag2/Pt2-B-I, respectively. Given their stability, Ag2/Pd2-C-I, Pt2-B-I, and Pd2-P-I BACs are identified as promising candidates for ORR catalysts with high activity and stability.
2.

(a) The BACs with outstanding catalytic performance in each system. The dashed lines represent the ηORR of Pt (111). The free energy diagrams of ORR on (b) TM2-C-I (TM = Ag, Au, Pd), (c) TM2-B-I (TM = Ag and Pt), and (d) TM2-P-I (TM = Pd). The values in the yellow square are the ηORR of the BACs.
3.3. Machine Learning-Assisted Prediction of Catalytic Performance
3.3.1. Origin of Catalytic Activity
ML can serve as a powerful tool in exploring the origins of catalytic activity and accelerating the screening process for catalysts with outstanding performance. Through the DFT calculations from the previous section, ηORR data from 130 samples were utilized as the labels for the ML model. A feature pool consisting of 59 features was initially employed, which can be divided into two classes: the elemental features of the metal active center and the features of the coordination environment. All the descriptors and abbreviations of these features are listed in Table S26. It is important to note that the structural features, such as TM-TM, TM1-Nn, TM2-Nn etc., are derived from structures optimized by DFT calculations. To more accurately characterize the structure information surrounding the metal active center, the Coulomb matrix (cm ij ) for the first coordination spheres is calculated as input features with the following equation:
| 3.2 |
where Z i and Z j denote the nuclear charges, r i and r j represent the respective Cartesian coordinates. Given that the Coulomb matrix is symmetric, only the upper triangular portion is utilized in our analysis (36 elements).
The data set is randomly divided into a training set and a test set in a ratio of 85% to 15%, respectively. A large number of input features can negatively affect model performance, mainly due to redundant and irrelevant features that increase the risk of overfitting. Therefore, five features with low variance (variance = 0) were removed based on the training set. Feature normalization was performed using the mean and standard deviation derived from the training set, and the same scaler was then applied to the test set. Subsequently, the Pearson correlation coefficient (PCC) was calculated based on the training set to identify linear correlations between features, as illustrated in the heatmap of the PCC matrix in Figure S3. Feature pairs with an absolute PCC value exceeding 0.95 were considered highly correlated, and among them, the feature showing a stronger correlation with ηORR was retained. After PCC analysis, a total of 11 features were removed.
Six different ML algorithms, including K-nearest neighbor (KNN), Support Vector Regression (SVR), Random Forest (RF), Gradient Boosting Decision Tree (GBDT), AdaBoost, and Bagging, are compared to identify the most suitable algorithm for further study. Details of the hyperparameters for these ML models are available in Table S27. To reduce overfitting and eliminate sampling error, 5-fold cross-validation (5 CV) is used for each ML model on the training set, and the coefficient of determination (R 2) is applied to evaluate the predictive accuracy of the model. Figure a illustrates the comparison between the six algorithms in terms of ηORR prediction performance for 5 CV, leading to the selection of GBDT for further ηORR predictions due to its superior accuracy on 5 CV with R 2 of 0.78.
3.

(a) Comparison of R 2 for ηORR prediction by six ML algorithms using 5 CV. (b) Recursive feature elimination process in the GBDT model. (c) Comparison of the ηORR obtained from DFT calculations with those predicted by the ML model. (d) The feature importance determined by the GBDT model.
In order to improve the performance of the GBDT model, recursive feature elimination was employed to identify key atomic properties within the training set. Employing the GBDT model, the importance of each feature is initially evaluated, followed by the gradual removal of features with low importance values and the subsequent retraining of the model. Finally, a subset of nine features is determined by comparing the 5 CV scores of the GBDT model, as shown in Figure b.
We retrained the GBDT model with these nine features, employing grid search and cross-validation for hyperparameter selection. The model performance is evaluated by R2 and root mean squared error (RMSE). Figure c depicts the comparison of DFT-computed ηORR with those predicted by the well-trained GBDT model. The scatter points distributed around the y = x line indicate that the ML-predicted ηORR is in good agreement with the DFT results. The R 2 and RMSE for the training set are 0.96 and 0.22 V; for the test set, they are 0.93 and 0.32 V. The relatively high R 2 of the test set demonstrates that by learning the hidden information in the data set, the GBDT model can achieve impressive precision for predicting the ηORR on g-CN-based BACs. The 5 CV scores shown in Figure b were used to determine the optimal number of features and were calculated on the training set. The final model performance was evaluated on an independent test set that was not involved in the RFE or cross-validation procedure.
Nine features (including the d-band center of the corresponding pure metal surface of TM (εd), Pauling electronegativity (χ), d orbital electrons number (e d) and the second ionization potential (I2) of the TM atom, D, the bond length of TM-TM and TM1-N2 (TM-TM, TM1-N2), and features cm8 and cm28) are ranked in terms of feature importance, as shown in Figure d. It is clear that the intrinsic properties of the metal active center play a dominant role in the catalytic activity on BACs, accounting for 71.6% of the total feature weight. Conversely, the coordination environment contributes to a lesser extent in regulating the catalytic activity, with a feature weight of 28.4%. Given that noble metals exhibit superior catalytic performance compared to other metals in TM2-N3 systems, the predominant role of the TM atom well explains why the metal active center of the high-performing TM2-C/B/P BACs is mostly a noble metal. Although the metal active center is a crucial factor in the catalytic activity of BACs, the role of the coordination atoms is equally vital. For catalysts with a suitable metal active center, regulating the coordination atoms can further improve the catalytic activity.
To clarify the contribution of the two highest-ranked features (εd and χ) to ηORR, we employed the Shapley Additive Explanations (SHAP) analysis framework. The SHAP values of εd and χ are presented in Figure S4. It can be found that the feature values of εd show a positive correlation with the SHAP values, while χ exhibits a negative correlation. Therefore, a BAC with a low value of εd and a high value of χ may present a superior catalytic activity.
3.3.2. Catalytic Activity of the TM2-S/C/B/P-II BACs
In this section, the ORR performance of 104 TM2-S/C/B/P-II BACs is predicted using the well-trained GBDT model, applying the same parameters as in the previous training process. Subsequently, the ORR catalytic performance of these BACs has been validated by DFT calculations to obtain precise theoretical ηORR values and to assess the predictive accuracy of the GBDT model. The comparison between ML-predicted ηORR and DFT-calculated ηORR is depicted in Figure a. It is worth noting that only the systems with ηORR< 0.9 V are compared. The ηORR of the ML-predicted systems is close to the DFT-calculated values, with an RMSE of 0.22 V, revealing the high accuracy of the GBDT model. In terms of catalytic performance, the g-CN-based BACs with a noble metal as the active center generally present desirable catalytic activity. Among all the BACs in Figure a, Pd2-C-II and Ag2-C-II are the promising ORR electrocatalysts with low overpotentials of only 0.49 and 0.44 V. The ORR free energy diagrams at U = 0 V for these catalysts are illustrated in Figure b,c. It is worth noting that for Pd2-C-II, the last reaction step of OH* to H2O is the PDS, with the free energy change of 0.74 eV, leading to a ηORR of 0.49 V, for Ag2-C-II BAC, the PDS is the second step from OOH* to O*. Moreover, the E f of Pd2-C-II and Ag2-C-II are computed, revealing values of −0.72 eV and −0.12 eV, respectively. These negative E f values indicate the excellent stability of these BACs. Meanwhile, the thermodynamic stability at room temperature and electrochemical stability of the promising candidates were further evaluated by ab initio molecular dynamics (AIMD) simulations and dissolution potential (U diss). The total energy of all systems fluctuates within a small range, and no significant structural distortion is observed in the final configurations. These results indicate that the studied structures possess excellent thermal stability at room temperature (Figure S5). All candidates exhibited U diss > 0, confirming their electrochemical stability (Table S28).
4.

(a) Comparison of ML-predicted ηORR with DFT-calculated ηORR on TM2-C/B/P-II. The free energy diagrams of ORR on (b) Pd2-C-II and (c) Ag2-C-II. The steps highlighted in red are the PDS for each BAC. The structural diagrams correspond to the optimized structures of OOH*, O* and OH* intermediates on Pd2-C-II and Ag2-C-II.
3.4. Effect of Heteroatom Doping on Catalytic Performance
To further clarify the reason for the enhanced catalytic activity resulting from heteroatom doping, we calculated the electronic structures of Pd2-N3/Pt2-N3 and Pd2-C-I/Pt2-B-I. In Figure a–b, the charge distributions of Pd2-N3 and Pt2-N3 are apparently symmetric. This symmetry is broken by the doped heteroatom, accompanied by a charge accumulation region between the metal and the C or B atoms, which increases the adsorption strength of the intermediates. The projected density of states of Pd2-N3/Pt2-N3 and Pd2-C-I/Pt2-B-I in Figure c,d reveals that Pd2-C-I/Pt2-B-I possesses a higher density states at the Fermi level than Pd2-N3/Pt2-N3, resulting in higher electronic conductivity and benefiting charge transfer in electrochemical processes. Moreover, the d-band center of Pd2-C-I/Pt2-B-I shifts to a higher energy level and is closer to the Fermi level compared to that of Pd2-N3/Pt2-N3. This leads to a higher energy of the antibonding state, resulting in less occupancy of antibonding states in the interaction with oxygen intermediates and thereby enhancing the adsorption strength.
5.

Charge density difference diagram of (a) Pd2-N3/Pd2-C-I and (b) Pt2-N3/Pt2-B-I. The isosurface value is 0.008 e/Å–3. Cyan and yellow stand for the charge depletion and accumulation, respectively. The calculated PDOS of (c) Pd2-N3/Pd2-C-I and (d) Pt2-N3/Pt2-B-I. The crystal orbital Hamilton population (COHP) between the metal active center and the O atom of OH* adsorbed on (e) Pd2-N3, (f) Pt2-N3, (g) Pd2-C-I, and (h) Pt2-B-I. The dashed lines represent the Fermi level. Red stands for antibonding states and blue stands for bonding states.
In order to confirm this conclusion, we calculated the crystal orbital Hamilton population (COHP) between TM and O atoms in the systems with OH* adsorbed on Pd2-N3/Pt2-N3 and Pd2-C-I/Pt2-B-I in Figure e–h. The COHP is displayed following the universal method of plotting negative contributions (bonding) on the right side and positive contributions (antibonding) on the left. It is worth noting that for OH* adsorbed on Pd2-N3 and Pt2-N3, the two Pd or Pt atoms have the same interaction with O; thus they have identical COHP diagrams. Compared to Pd2-N3 and Pt2-N3, the number of unoccupied antibonding orbitals (above the Fermi level) increases in Pd2-C-I and Pt2-B-I, corresponding to an increased adsorption strength of OH*. Subsequently, the Integrated Crystal Orbital Hamilton Population (ICOHP) values are calculated to provide more quantitative information. In Figure e–h, the ICOHP values of Pd2-C-I and Pt2-B-I are more negative than those of Pd2-N3 and Pt2-N3, suggesting a significantly stronger binding strength of OH* to the former systems and thus guaranteeing better ORR activity (Figure S6).
4. Conclusions
In conclusion, we first calculated the stability and ORR performance for 26 TM2 dimers embedded in pristine g-CN and S, C, B, P-doped g-CN with TM2-S/C/B/P-I configuration by DFT calculations. Then, using ηORR of these BACs as labels, six different supervised ML algorithms were selected to train the model, among which GBDT emerged as the best algorithm to characterize the ORR activity. Based on the well-trained GBDT model, we predicted the catalytic activities of TM2-S/C/B/P-II BACs. Finally, BACs such as Ag2/Pd2-C-I, Pt2-B-I, Pd2-P-I, and Pd2/Ag2-C-II are identified as promising candidates for ORR catalysts with high activity and stability. ML analysis reveals that regulating the coordination atoms is an efficient way to further enhance the catalytic activity of BACs. Additionally, electronic structure calculations reveal that the doped heteroatom can increase the d-band center of the catalysts, thus enhancing the adsorption strength of ORR intermediates and improving the ORR activity. This study provides fundamental insights into ORR and a cost-effective ML approach for developing new catalysts for ORR, as demonstrated by our theoretical predictions.
Supplementary Material
Acknowledgments
This work was financially supported by the National Natural Science Foundation of China under Grant No. 51371017. This research was supported by the high-performance computing (HPC) resources at Beihang University.
The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.6c08354.
Supporting Information contains the two coordination configurations and optimized Zn2 structures in pristine and S/C/B/P-doped g-CN; the Pearson correlation heatmap of 23 input features; the SHAP analysis of the top-two features εd and χ; the AIMD energy evolution at 300 K for 5000 fs, with final structures; and the ORR overpotential volcano plot; It also contains the K-point convergence energies of g-CN, substitution formation energies, and TM-TM/TM-N/TM-S/C/B/P bond lengths with cavity dimensions and lattice constants for all TM2-embedded structures; Adsorption free energies of oxygenated intermediates and ORR overpotentials; the initial 59 machine-learning features, ML hyperparameters, dissolution potentials of the six catalysts, and hydrogen adsorption free energies (PDF)
The authors declare no competing financial interest.
References
- Aralekallu S., Singh V.. M-N-C-Based Non-Precious Metal Catalyst Materials for Electrocatalytic ORR Applications. Fuel. 2026;404:136163. doi: 10.1016/j.fuel.2025.136163. [DOI] [Google Scholar]
- Chu S., Majumdar A.. Opportunities and Challenges for a Sustainable Energy Future. Nature. 2012;488:294–303. doi: 10.1038/nature11475. [DOI] [PubMed] [Google Scholar]
- Li L., Huang B., Tang X., Hong Y., Zhai W., Hu T., Yuan K., Chen Y.. Recent Developments of Microenvironment Engineering of Single-Atom Catalysts for Oxygen Reduction toward Desired Activity and Selectivity. Adv. Funct. Mater. 2021;31:2103857. doi: 10.1002/adfm.202103857. [DOI] [Google Scholar]
- Zaman S., Huang L., Douka A. I., Yang H., You B., Xia B. Y.. Oxygen Reduction Electrocatalysts toward Practical Fuel Cells: Progress and Perspectives. Angew. Chem., Int. Ed. 2021;60:17832–17852. doi: 10.1002/anie.202016977. [DOI] [PubMed] [Google Scholar]
- Chen K., Wang X., Yang N., Zhang C., Wang Z., Zhen C., Cui Y., Wang H., Chu L., Gu M. D.. et al. Dual-Channel σ-Donation/π-Backdonation-Driven -p-p- Coupling in Calcium Dual-Atom Catalysts Enables Accelerated ORR. Adv. Energy Mater. 2026;16:e04855. doi: 10.1002/aenm.202504855. [DOI] [Google Scholar]
- Chen Z., Yang T. C., Yang C. M., Bao C., Jiang Z., Deng S., Lai N.. Symmetry Distortion Engineering Single-Atom Catalysts: Exceptional ORR Activity and Ultra-Stable Zn-Air Batteries. Chem. Eng. J. 2025;516:164156. doi: 10.1016/j.cej.2025.164156. [DOI] [Google Scholar]
- Qiao B., Wang A., Yang X., Allard L. F., Jiang Z., Cui Y., Liu J., Li J., Zhang T.. Single-Atom Catalysis of CO Oxidation Using Pt1/FeOx . Nat. Chem. 2011;3:634–641. doi: 10.1038/nchem.1095. [DOI] [PubMed] [Google Scholar]
- Liu T., Jing Y., Li Y.. First-Principles Insights into the Selectivity of CO2 Electroreduction over Heterogeneous Single-Atom Catalysts. J. Phys. Chem. Lett. 2024;15:6216–6221. doi: 10.1021/acs.jpclett.4c01096. [DOI] [PubMed] [Google Scholar]
- Hu X., Chen S., Chen L., Tian Y., Yao S., Lu Z., Zhang X., Zhou Z.. What is the Real Origin of the Activity of Fe–N–C Electrocatalysts in the O 2 Reduction Reaction? Critical Roles of Coordinating Pyrrolic N and Axially Adsorbing Species. J. Am. Chem. Soc. 2022;144:18144–18152. doi: 10.1021/jacs.2c08743. [DOI] [PubMed] [Google Scholar]
- Wang Y., Hu R., Li Y., Wang F., Shang J., Shui J.. High-Throughput Screening of Carbon-Supported Single Metal Atom Catalysts for Oxygen Reduction Reaction. Nano Res. 2022;15:1054–1060. doi: 10.1007/s12274-021-3598-2. [DOI] [Google Scholar]
- Yang Z., Zhang H., Dai Q., Wei B., Zhou G., Shang J.. Machine Learning-Assisted Screening and Prediction of Bifunctional OER/ORR Single-Atom Catalysts for Non-Metal-Doped C2N. Appl. Surf. Sci. 2026;724:165762. doi: 10.1016/j.apsusc.2025.165762. [DOI] [Google Scholar]
- Yu L., Su Z., Ma N., Luo S., Yeung Y. Y., Si Y., Ren Y., Fan J., Wei Z.. Designing High-Performance Ni/co Single-Atom Catalysts for ORR/OER through Spin-Coordination Engineering. J. Colloid Interface Sci. 2026;706:139596. doi: 10.1016/j.jcis.2025.139596. [DOI] [PubMed] [Google Scholar]
- Li W., Mo Y., Kang L., Li C., Zheng J., Qiu C.. Active Center Inheritance: A Design Principle for Constructing Atomically Dispersed Bimetallic OER/ORR Catalysts Derived from Single-Atom Catalysts. ACS Catal. 2025;15:13872–13885. doi: 10.1021/acscatal.5c02722. [DOI] [Google Scholar]
- Tian H., Song A., Zhang P., Sun K., Wang J., Sun B., Fan Q., Shao G., Chen C., Liu H.. et al. High Durability of Fe–N–C Single-Atom Catalysts with Carbon Vacancies toward the Oxygen Reduction Reaction in Alkaline Media. Adv. Mater. 2023;35:e2210714. doi: 10.1002/adma.202210714. [DOI] [PubMed] [Google Scholar]
- Zhang X., Xu X., Yao S., Hao C., Pan C., Xiang X., Tian Z. Q., Shen P. K., Shao Z., Jiang S. P.. Boosting Electrocatalytic Activity of Single Atom Catalysts Supported on Nitrogen-Doped Carbon through N Coordination Environment Engineering. Small. 2022;18:e2105329. doi: 10.1002/smll.202105329. [DOI] [PubMed] [Google Scholar]
- Kumar A., Bishnoi P., Parmar A., Khalid M., Ubaidullah M., Al-Enizi A. M.. Accelerating Single-Atom ORR Catalyst Discovery through Theory and Machine Learning: A Critical Review. Coord. Chem. Rev. 2026;560:217870. doi: 10.1016/j.ccr.2026.217870. [DOI] [Google Scholar]
- Xia H., Sun H., Yang D., Zhao J., Gao G., Wu L., Huang L., Jiang X.. Curvature-Engineered Steering of Oxygen Electroreduction Pathways on Single-Atom Catalysts. Angew. Chem., Int. Ed. 2026;65:e3924995. doi: 10.1002/anie.3924995. [DOI] [PubMed] [Google Scholar]
- Hu R., Li Y., Wang F., Shang J.. Rational Prediction of Multifunctional Bilayer Single Atom Catalysts for the Hydrogen Evolution, Oxygen Evolution and Oxygen Reduction Reactions. Nanoscale. 2020;12:20413–20424. doi: 10.1039/D0NR05202G. [DOI] [PubMed] [Google Scholar]
- Hu B., Huang A., Zhang X., Chen Z., Tu R., Zhu W., Zhuang Z., Chen C., Peng Q., Li Y.. Atomic Co/Ni Dual Sites with N/P-Coordination as Bifunctional Oxygen Electrocatalyst for Rechargeable Zinc-Air Batteries. Nano Res. 2021;14:3482–3488. doi: 10.1007/s12274-021-3535-4. [DOI] [Google Scholar]
- Xue M., Gao Y., Wang E., Zhou J., Sun Z.. First-Principles Study of the Pt@(Hf2–xZrx)B2O2 Bifunctional ORR/OER Catalysts. J. Mater. Chem. A. 2026;14:6560–6571. doi: 10.1039/D5TA06919J. [DOI] [Google Scholar]
- Jasin Arachchige L., Xu Y., Dai Z., Zhang X., Wang F., Sun C.. Theoretical Investigation of Single and Double Transition Metals Anchored on Graphyne Monolayer for Nitrogen Reduction Reaction. J. Phys. Chem. C. 2020;124:15295–15301. doi: 10.1021/acs.jpcc.0c03899. [DOI] [Google Scholar]
- Zhang L., Fischer J., Jia Y., Yan X., Xu W., Wang X., Chen J., Yang D., Liu H., Zhuang L., Hankel M., Searles D. J., Huang K., Feng S., Brown C. L., Yao X.. Coordination of Atomic Co-Pt Coupling Species at Carbon Defects as Active Sites for Oxygen Reduction Reaction. J. Am. Chem. Soc. 2018;140:10757–10763. doi: 10.1021/jacs.8b04647. [DOI] [PubMed] [Google Scholar]
- Chen J., Lu Y., Zhang F., Xiao X., Xu Q.. Dual-Atom Catalysts for Oxygen Reduction Reaction. Matter. 2026;9:102532. doi: 10.1016/j.matt.2025.102532. [DOI] [Google Scholar]
- Wang Z., Jin X., Zhu C., Liu Y., Tan H., Ku R., Zhang Y., Zhou L., Liu Z., Hwang S. J.. et al. Atomically Dispersed Co 2 −N 6 and Fe–N 4 Costructures Boost Oxygen Reduction Reaction in Both Alkaline and Acidic Media. Adv. Mater. 2021;33:e2104718. doi: 10.1002/adma.202104718. [DOI] [PubMed] [Google Scholar]
- Liu J., Xu H., Zhu J., Cheng D.. Understanding the Pathway Switch of the Oxygen Reduction Reaction from Single- to Double-/Triple-Atom Catalysts: A Dual Channel for Electron Acceptance–Backdonation. JACS Au. 2023;3:3031–3044. doi: 10.1021/jacsau.3c00432. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Wang J., Xu R., Sun Y., Liu Q., Xia M., Li Y., Gao F., Zhao Y., Tse J. S.. Identifying the Zn–Co Binary as a Robust Bifunctional Electrocatalyst in Oxygen Reduction and Evolution Reactions via Shifting the Apexes of the Volcano Plot. J. Energy Chem. 2021;55:162–168. doi: 10.1016/j.jechem.2020.07.010. [DOI] [Google Scholar]
- Zhang L., Guo X., Zhang S., Huang S.. Building Up the “Genome” of Bi-Atom Catalysts toward Efficient HER/OER/ORR. J. Mater. Chem. A. 2022;10:11600–11612. doi: 10.1039/D2TA02050E. [DOI] [Google Scholar]
- Kong F., Si R., Chen N., Wang Q., Li J., Yin G., Gu M., Wang J., Liu L.-M., Sun X.. Origin of Hetero-Nuclear Au-Co Dual Atoms for Efficient Acidic Oxygen Reduction. Appl. Catal., B. 2022;301:120782. doi: 10.1016/j.apcatb.2021.120782. [DOI] [Google Scholar]
- Wang J., Liu C., Li S., Li Y., Zhang Q., Peng Q., Tse J. S., Wu Z.. Advanced Electrocatalysts with Dual-Metal Doped Carbon Materials: Achievements and Challenges. Chem. Eng. J. 2022;428:132558. doi: 10.1016/j.cej.2021.132558. [DOI] [Google Scholar]
- Li Z., He H., Cao H., Sun S., Diao W., Gao D., Lu P., Zhang S., Guo Z., Li M., Liu R., Ren D., Liu C., Zhang Y., Yang Z., Jiang J., Zhang G.. Atomic Co/Ni Dual Sites and Co/Ni Alloy Nanoparticles in N-Doped Porous Janus-like Carbon Frameworks for Bifunctional Oxygen Electrocatalysis. Appl. Catal., B. 2019;240:112–121. doi: 10.1016/j.apcatb.2018.08.074. [DOI] [Google Scholar]
- Yin L., Zhang S., Sun M., Wang S., Huang B., Du Y.. Heteroatom-Driven Coordination Fields Altering Single Cerium Atom Sites for Efficient Oxygen Reduction Reaction. Adv. Mater. 2023;35:e2302485. doi: 10.1002/adma.202302485. [DOI] [PubMed] [Google Scholar]
- Zhao R., Li Q., Chen Z., Jose V., Jiang X., Fu G., Lee J.-M., Huang S. B.. N-Doped Ultrathin Carbon Nanosheet Superstructure for High-Performance Oxygen Reduction Reaction in Rechargeable Zinc-Air Battery. Carbon. 2020;164:398–406. doi: 10.1016/j.carbon.2020.04.019. [DOI] [Google Scholar]
- Liu J., Xiao J., Luo B., Tian E., Waterhouse G. I. N.. Central Metal and Ligand Effects on Oxygen Electrocatalysis over 3d Transition Metal Single-Atom Catalysts: A Theoretical Investigation. Chem. Eng. J. 2022;427:132038. doi: 10.1016/j.cej.2021.132038. [DOI] [Google Scholar]
- Li G., Tang Y., Fu T., Xiang Y., Xiong Z., Si Y., Guo C., Jiang Z. S.. N Co-Doped Carbon Nanotubes Coupled with CoFe Nanoparticles as an Efficient Bifunctional ORR/OER Electrocatalyst for Rechargeable Zn-Air Batteries. Chem. Eng. J. 2022;429:132174. doi: 10.1016/j.cej.2021.132174. [DOI] [Google Scholar]
- Yuan K., Lutzenkirchen-Hecht D., Li L., Shuai L., Li Y., Cao R., Qiu M., Zhuang X., Leung M. K. H., Chen Y., Scherf U.. Boosting Oxygen Reduction of Single Iron Active Sites via Geometric and Electronic Engineering: Nitrogen and Phosphorus Dual Coordination. J. Am. Chem. Soc. 2020;142:2404–2412. doi: 10.1021/jacs.9b11852. [DOI] [PubMed] [Google Scholar]
- Chang S., Zhang H., Zhang Z.. FeCo Alloy/NS Dual-Doped Carbon Composite as a High-Performance Bifunctional Catalyst in an Advanced Rechargeable Zinc-Air Battery. J. Energy Chem. 2021;56:64–71. doi: 10.1016/j.jechem.2020.07.047. [DOI] [Google Scholar]
- Zhong W., Qiu Y., Shen H., Wang X., Yuan J., Jia C., Bi S., Jiang J.. Electronic Spin Moment as a Catalytic Descriptor for Fe Single-Atom Catalysts Supported on C2N. J. Am. Chem. Soc. 2021;143:4405–4413. doi: 10.1021/jacs.1c00889. [DOI] [PubMed] [Google Scholar]
- Wang Y., Huang X., Fu H., Shang J.. Theoretically Revealing the Activity Origin of the Hydrogen Evolution Reaction on Carbon-Based Single-Atom Catalysts and Finding Ideal Catalysts for Water Splitting. J. Mater. Chem. A. 2022;10:24362–24372. doi: 10.1039/D2TA07167C. [DOI] [Google Scholar]
- Zhu D., Wu H. W., Fong W. K., Tabor R. F., Zhang J.. The Ratio of sp2 and sp3 Hybridized Carbon Determines the Performance of Carbon-Based Catalysts in H2O2 Electrosynthesis from O2 . Angew. Chem., Int. Ed. 2025;64:e202500145. doi: 10.1002/anie.202500145. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Amoo C. C., Ge Q., Ordomsky V., Sun J.. Synthesizing Liquid Fuels Over Carbon-Based Catalysts via CO2 Conversion. Adv. Sci. 2025;12:2410280. doi: 10.1002/advs.202410280. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Guo L., Li R., Jiang J., Zou J.-J., Mi W.. Electrocatalytic Performance of Mn-Adsorbed g-C3N4: A First-Principles Study. J. Mater. Chem. A. 2021;9:26266–26276. doi: 10.1039/D1TA07286B. [DOI] [Google Scholar]
- Han S., Wei X., Huang Y., Zhang J., Yang J., Wang Z.. Tuning the Activity and Selectivity of Nitrogen Reduction Reaction on Double-Atom Catalysts by B Doping: A Density Functional Theory Study. Nano Energy. 2022;99:107363. doi: 10.1016/j.nanoen.2022.107363. [DOI] [Google Scholar]
- Barrio J., Pedersen A., Feng J., Sarma S. C., Wang M., Li A. Y., Yadegari H., Luo H., Ryan M. P., Titirici M.-M., Stephens I. E. L.. Metal Coordination in C2N-Like Materials towards Dual Atom Catalysts for Oxygen Reduction. J. Mater. Chem. A. 2022;10:6023–6030. doi: 10.1039/D1TA09560A. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Gao Y., Cai Z., Wu X., Lv Z., Wu P., Cai C.. Graphdiyne-Supported Single-Atom-Sized Fe Catalysts for the Oxygen Reduction Reaction: DFT Predictions and Experimental Validations. ACS Catal. 2018;8:10364–10374. doi: 10.1021/acscatal.8b02360. [DOI] [Google Scholar]
- Singh S. K., Takeyasu K., Nakamura J.. Active Sites and Mechanism of Oxygen Reduction Reaction Electrocatalysis on Nitrogen-Doped Carbon Materials. Adv. Mater. 2019;31:e1804297. doi: 10.1002/adma.201804297. [DOI] [PubMed] [Google Scholar]
- Yang Q., Xiao Z., Kong D., Zhang T., Duan X., Zhou S., Niu Y., Shen Y., Sun H., Wang S., Zhi L.. New Insight to the Role of Edges and Heteroatoms in Nanocarbons for Oxygen Reduction Reaction. Nano Energy. 2019;66:104096. doi: 10.1016/j.nanoen.2019.104096. [DOI] [Google Scholar]
- Quílez-Bermejo J., Pérez-Rodríguez S., Canevesi R., Torres D., Morallón E., Cazorla-Amorós D., Celzard A., Fierro V.. Easy Enrichment of Graphitic Nitrogen to Prepare Highly Catalytic Carbons for Oxygen Reduction Reaction. Carbon. 2022;196:708–717. doi: 10.1016/j.carbon.2022.05.032. [DOI] [Google Scholar]
- Pu P., Song X., Ding H., Deng Y., Feng H., Zhang X.. Fast-Track to Catalyst Stability: Machine Learning Optimized Predictions for M1/M2-N6-Gra Catalysts. J. Phys. Chem. Lett. 2025;16:4253–4261. doi: 10.1021/acs.jpclett.5c00097. [DOI] [PubMed] [Google Scholar]
- Umer M., Umer S., Zafari M., Ha M., Anand R., Hajibabaei A., Abbas A., Lee G., Kim K. S.. Machine Learning Assisted High-Throughput Screening of Transition Metal Single Atom Based Superb Hydrogen Evolution Electrocatalysts. J. Mater. Chem. A. 2022;10:6679–6689. doi: 10.1039/D1TA09878K. [DOI] [Google Scholar]
- Kresse G., Furthmüller J.. Efficient Iterative Schemes for Ab Initio Total-Energy Calculations Using a Plane-Wave Basis Set. Phys. Rev. B. 1996;54:11169. doi: 10.1103/PhysRevB.54.11169. [DOI] [PubMed] [Google Scholar]
- Kresse G., Furthmüller J.. Efficiency of Ab-Initio Total Energy Calculations for Metals and Semiconductors Using a Plane-Wave Basis Set. Comput. Mater. Sci. 1996;6:15–50. doi: 10.1016/0927-0256(96)00008-0. [DOI] [PubMed] [Google Scholar]
- Blochl P. E.. Projector Augmented-Wave Method. Phys. Rev. B. 1994;50:17953–17979. doi: 10.1103/PhysRevB.50.17953. [DOI] [PubMed] [Google Scholar]
- Ernzerhof M., Scuseria G. E.. Assessment of the Perdew–Burke–Ernzerhof Exchange-Correlation Functional. J. Chem. Phys. 1999;110:5029–5036. doi: 10.1063/1.478401. [DOI] [PubMed] [Google Scholar]
- Perdew J. P., Burke K., Ernzerhof M.. Generalized Gradient Approximation Made Simple. Phys. Rev. Lett. 1996;77:3865–3868. doi: 10.1103/PhysRevLett.77.3865. [DOI] [PubMed] [Google Scholar]
- Monkhorst H. J., Pack J. D.. Special Points for Brillouin-Zone Integrations. Phys. Rev. B. 1976;13:5188–5192. doi: 10.1103/PhysRevB.13.5188. [DOI] [Google Scholar]
- Grimme S., Antony J., Ehrlich S., Krieg H.. A Consistent and Accurate Ab Initio Parametrization of Density Functional Dispersion Correction (DFT-D) for the 94 Elements H-Pu. J. Chem. Phys. 2010;132:154104. doi: 10.1063/1.3382344. [DOI] [PubMed] [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. 2014;140:084106. doi: 10.1063/1.4865107. [DOI] [PubMed] [Google Scholar]
- Nørskov J. K., Rossmeisl J., Logadottir A., Lindqvist L., Kitchin J. R., Bligaard T., Jónsson H.. Origin of the Overpotential for Oxygen Reduction at a Fuel-Cell Cathode. J. Phys. Chem. B. 2004;108:17886–17892. doi: 10.1021/jp047349j. [DOI] [PubMed] [Google Scholar]
- Li J., Cao C., Hao J., Qiu H., Xu Y., Zhu H.. Self-Assembled One-Dimensional Carbon Nitride Architectures. Diamond Relat. Mater. 2006;15:1593–1600. doi: 10.1016/j.diamond.2006.01.013. [DOI] [Google Scholar]
- Huang B., Wu Y., Chen B., Qian Y., Zhou N., Li N.. Transition-Metal-Atom-Pairs Deposited on g-CN Monolayer for Nitrogen Reduction Reaction: Density Functional Theory Calculations. Chin. J. Catal. 2021;42:1160–1167. doi: 10.1016/S1872-2067(20)63745-7. [DOI] [Google Scholar]
- Kulkarni A., Siahrostami S., Patel A., Norskov J. K.. Understanding Catalytic Activity Trends in the Oxygen Reduction Reaction. Chem. Rev. 2018;118:2302–2312. doi: 10.1021/acs.chemrev.7b00488. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
