3.3 High throughput screening strategy and full reaction
pathways for selected homonuclear systems
Based on the conclusion discussed above, we then use a three-step
screening to filtrate high catalytic candidates for eNRR from the stable
TM anchored systems: (1) we first screen out the adsorption energy of
N2 molecule on
TM2@sitex(x= 1, 2, 3, 4) by calculating the Gibbs free energy change of the
adsorbent (ΔG*N2), where ΔG*N2< 0 means that the N2 molecule can
spontaneously adsorb on the active site; (2) Based on the results from
the first screening and the priority of the first hydrogenation site as
discussed former, the Gibbs free energy changes of the first
hydrogenation (ΔG*N2→*N2H) process is smaller than 0.50
eV could be the new screening strategy for following selection. (3)
After ensuring the candidates from the second filtration, we further
screen out the catalyst with the Gibbs free energy changes of the last
hydrogenation process (ΔG*NH2→*NH3) on the
TM2@sitex (x= 1, 2, 3, 4) less than 0.50
eV. Meanwhile, we also implemented state-of-the-art machine learning to
predict certain descriptors for explaining the screen relationship using
a gradient-boosted regression (GBR) algorithm. It should be worth noting
that the intrinsic properties, such as d-band center, charge transfer,
spin moment, etc., are not considered, which can advantageously minimize
the intrinsic properties’ impact and reprove the influence of the atomic
structural effect. The details of the machine learning strategies are
noted in Support Information Note 2 .
For the eNRR process, the spontaneous adsorption of the
N2 molecule on TM2@sitex(x = 1, 2, 3, 4) is the fundamental step for the later activation. As
plotted in Figure 3(a) and Table S7 , among all the
adsorbed systems, most of them show negative ΔGN2 (-2.71
eV to -0.02 eV), demonstrating the efficient capture of the
N2 molecule. On the contrary, the rest systems, such as
Pd, Ag, W(side-on), Pt and Au in site1, Co(side-on),
Ni(side-on), Cu, Pd, Ag, Os(side-on), Pt in site2,
Pd(side-on), Ag(end-on), Pt(side-on) and Au(end-on) in
site3 as well as Ag(end-on) and Au(side-on) in
site4 show positive ΔGN2 of 0.01 to 0.91
eV, would be excluded for further exploration. For the machine learning
prediction part, end-on and side-on configurations are predicted with
nine input factors. The machine learning results are plotted inFigure 3 (d, e). For end-on configuration, the outermost
electron number of TM atoms (Ne), as well as the average
bond length between TM atoms and neighboring coordinated atoms
(ra), has the most proportion in the bar chart with
feature importance of 41%, 16.3% to predict ΔG*N2. The
high coefficient of 0.97/0.91 and a low RMSE of 0.09/0.16
(Figure S7 ) demonstrate that the built GBR model is reliable.
The Ne could be a good descriptor for end-on
N2 adsorption. The main reason may originate from the
most active electron of them. For the side-on configuration part,
ΔG*N2 also successfully predicted by the machine
learning method within a high coefficient of 0.99/0.97 and low RMSE of
0.08/0.13 (Figure S7 ), further revealing the GBR model’s
reliability. At the same time, d-electron count (θd) and
TM bond length (rTM) become the main features with
proportions of 47.9% and 21.7%, respectively.
According to recent studies69, 70, the limiting
potential steps of the eNRR are highly related to the first protonation
step (*N2 → *N2H) and the last
protonation step (*NH2 → *NH3) among the
six protonation steps. Thus, the ΔG*N2→*N2H and
ΔG*NH2→*NH3 are comprehensively calculated in the second
and third screening, and the built machine learning model is also
carried out for descriptors finding. As depicted in Figure
3(b) , we first screen potential candidates and then employ the GBR
model to predict descriptors. For the screening process, there are 15
kinds of N2-TM2@sitex (x
= 1, 2, 3, 4) configurations with ΔG*N2→*N2H less than
0.50 eV, including Cr (end-on), Ru(side-on) in site1,
Co(end-on), W(side-on) in site2, Cr(end-on/side-on),
Mo(end-on/side-on), W(side-on), Os(side-on) in site3,
and V(side-on), Cr(side-on), Ni(end-on), Mo(side-on), W(end-on/side-on),
Re(end-on/side-on) in site4, so those systems are
selected for the last filtration. Meanwhile, machine learning predicts
that rTM, as well as the radius of the TM atom
(rd), are good factors that affect the first catalytic
step, with the features of 49.2% and 31.2%, along with a high
coefficient of 0.98/0.91 and low RMSE of 0.21 and 0.44, respectively
(Figure S7 ). Successively, the values of
ΔG*NH2→*NH3 are further investigated to make deep
filtration, as concluded in Figure 3(c) , 10 kinds of candidates
are selected eventually for the ΔG*NH2→*NH3 is less than
0.50 eV, including Mn and Ru in site2, Co, W in
site2, Os in site3, V, Cr, Ni, W, Re in
site4, as for the rest systems whose
ΔG*NH2→*NH3 more than 0.50 eV are eliminated for further
exploration. Therefore, after a series of screenings, Mn and Ru in
site2, Co, W in site2, Os in
site3, and V, Cr, Ni, W, Re in site4have high activity and may be promising candidates for full eNRR
reaction. Besides, though there are few data after the second screening
(15 kinds of screened systems), machine learning is also carried out to
search for excellent descriptors, as depicted in Figure 3(g) ,
the coefficient also reaches up to 0.96/0.95 for the
ΔG*NH2→*NH3 with low RMSE of 0.18/0.12. Simultaneously,
rTM as well as ra between TM and the
neighbor modified atoms, are the major contributors (42.7%, 37.3%),
proving they are good descriptors for the last hydrogenation step.