Page 39 - Read Online

Page 12 of 16 Lu et al. J Mater Inf 2024;4:31 https://dx.doi.org/10.20517/jmi.2024.65

Figure 9. (A) Feature importance and (B) SHAP value distribution in XGBR model. SHAP: Shapley additive explanations; XGBR: Extreme
gradient boosting regression.

properties of the metal center play a key role in determining the catalytic performance of the catalysts. To
further understand the relationships between C-PDS and the key features, SHAP analysis was employed. As
shown in Figure 9B, the most important feature S has a clear bipolar distribution, indicating that the last
6
protonation step generally has a lower ΔG than the first step. Similarly, S , ranked fourth, reveals that
1E
end-on adsorption tends to result in a higher energy barrier for the first protonation step. N , the second
m
most important feature, showed a more complex SHAP distribution, reflecting its multidimensional nature.
As an ordering system for chemical elements, N incorporates various elemental properties such as electron
m
configuration, atomic radius, and ionization energy, leading to a mixture of SHAP values. Nevertheless, N
m
plays a significant role in influencing catalytic activity, consistent with previous findings on
TM@graphene-NCM. The third most important feature, N , also showed a clear correlation with C-PDS.
d
Higher N values tend to lower the C-PDS, suggesting that a greater N enhances the interaction between
d
d
the TM center and N , promoting N activation and reducing the reaction barrier. These findings provide
2
2
valuable insights into the factors influencing NRR activity in TM@C N-NCM systems and offer guidance
2
for the design and optimization of carbene-coordinated SACs based on the identified universal principles.
CONCLUSIONS
In this study, a combination of DFT calculations and ML methods was employed to comprehensively
evaluate the NRR catalytic performance of 28 TM@C N-NCM catalysts by embedding TMs into a
2
two-dimensional C N-NCM substrate. Through multi-step screening, full-pathway validation, and
2
competition analysis with HER, eight catalysts (TM = Nb, Fe, Mn, W, V, Ta, Zr, Ti) with high activity and
selectivity were identified. Among them, Nb@C N-NCM exhibited the best performance, with U of -0.29 V
L
2
. Remarkably, all identified catalysts showed lower U values than their TM@graphene-NCM counterparts,
L
revealing the crucial role of the C N substrate in enhancing catalytic performance. ML analysis using the
2
2
XGBR achieved excellent predictive accuracy for C-PDS free energies, with an R of 0.91 and MAE of 0.19.
Feature importance analysis identified S , N , and N as the most influential factors affecting NRR catalytic
6
d
m
performance. SHAP analysis further validated the significant roles of these features, demonstrating that
tuning these intrinsic properties can effectively modulate the catalytic activity of TM@C N-NCM catalysts.
2