Rational Design of Atomic Site Catalysts for Electrocatalytic Nitrogen Reduction Reaction: One Step Closer to Optimum Activity and Selectivity

The electrocatalytic nitrogen reduction reaction (NRR) has been one of the most intriguing catalytic reactions in recent years, providing an energy-saving and environmentally friendly alternative to the conventional Haber–Bosch process for ammonia production. However, the activity and selectivity issues originating from the activation barrier of the NRR intermediates and the competing hydrogen evolution reaction result in the unsatisfactory NH3 yield rate and Faradaic efficiency of current NRR catalysts. Atomic site catalysts (ASCs), an emerging group of heterogeneous catalysts with a high atomic utilization rate, selectivity, and stability, may provide a solution. This article undertakes an exploration and systematic review of a highly significant research area: the principles of designing ASCs for the NRR. Both the theoretical and experimental progress and state-of-the-art techniques in the rational design of ASCs for the NRR are summarized, and the topic is extended to double-atom catalysts and boron-based metal-free ASCs. This review provides guidelines for the rational design of ASCs for the optimum activity and selectivity for the electrocatalytic NRR. Rational design of atomic site catalysts (ASCs) for nitrogen reduction reaction (NRR) has both scientific and industrial significance. In this review, the recent experimental and theoretical breakthroughs in the design principles of transition metal ASCs for NRR are comprehensively discussed, and the topic is also extended to double-atom catalysts and boron-based metal-free ASCs.


Introduction
Ammonia (NH 3 ) is an important chemical in both industry and agriculture because it can serve as a precursor to various foods, fertilizers, pharmaceuticals, and detergents [1,2]. In addition, ammonia is also a clean, carbon-free energy carrier since its combustion only produces water and dinitrogen (N 2 ), both of which are environmentally friendly [3]. Therefore, converting dinitrogen-the most abundant gas in Earth's atmosphere (over 78%)-into ammonia, one of the nitrogen fixation processes, is a topic with vital significance in Earth's nitrogen cycle [2]. According to Mineral Commodity Summaries from the US Geological Survey, approximately 170 million tons of ammonia in total were produced globally in 2018 [4]. A major challenge in nitrogen fixation is breaking the chemically inert N≡N triple bond, which has a high bond energy (940.95 kJ mol −1 ) [5,6]. Traditionally, ammonia production is conducted through the Haber-Bosch (H-B) process, which demands high temperature (~ 800 K) and high pressure [200-300 atm (1 atm = 101 325 Pa)] to proceed. The H-B process accounts for approximately 1.4% of the global energy consumption, among which ~ 75% is used in steam reforming instead of ammonia synthesis ( Fig. 1) [7][8][9][10]. Moreover, the H-B process may worsen environmental issues, such as the emission of greenhouse gases, as it accounts for approximately 1% of global emissions [11,12].
The electrocatalytic or photo(electro)catalytic nitrogen reduction reaction (NRR) is viewed as an energy-saving alternative to the H-B process for ammonia production because it can proceed under ambient conditions [11,[13][14][15][16][17]. Other advantages of the NRR include the elimination of carbon fuel consumption and carbon dioxide emissions, high energy efficiency, and smaller plant infrastructure

NRR Mechanisms
The prevailing opinions suggest that NRR is a six protoncoupled electron transfer (PCET) process, and possible mechanisms include dissociative and associative pathways [7,13,53,54]. For ASCs, the direct dissociation of the N≡N triple bond in the first step is unfavorable because it has a large energy barrier, so usually only the associative NRR pathway is considered [53]. Associative mechanisms include distal, alternating, enzymatic, and consecutive pathways, depending on whether the end-on or side-on N 2 adsorption pattern is more energetically favorable (Fig. 3). In end-on N 2 adsorption, the N atom further from the active center is first hydrogenated to form the first NH 3 molecule in the distal pathway, while the two N atoms are hydrogenated simultaneously in the alternating pathway. Similarly, in side-on N 2 adsorption, one of the N atoms is first hydrogenated to form the first NH 3 molecule in the consecutive pathway, and two N atoms are simultaneously hydrogenated in the enzymatic pathway. In the determination of the potential-limiting step (PLS) and the maximum Gibbs free energy change (ΔG), it is noteworthy that unlike in the NRR and thermal nitrogen fixation under basic conditions, the last step of NH 3 desorption in the NRR under acidic/slightly basic conditions is usually not considered in the calculations since protonation from *NH 3 to NH 4 + is usually facile [55,56].

Computational Schemes for the NRR
Because of the difficulty in the experimental determination of the catalytic active sites and underlying mechanisms for ASCs, theoretical calculations, especially DFT calculations, play an essential role in the study of ASCs [57]. The prevailing computational schemes for the NRR are based on Nørskov's computational hydrogen electrode (CHE) model, in which the Gibbs free energy of the proton-electron pair is calculated by G(H + + e − ) = 0.5G(H 2 ) − eU, where U is the applied potential [58,59]. The applied potential required to diminish the energy barrier of all steps in the NRR is defined as the limiting potential (U L ), and overpotential η is calculated by η = U equil − U L , where U equil ≈ − 0.16 V is the equilibrium potential for the NRR [53]. The U L values (at the DFT-generalized gradient approximation (GGA) level), energetically favorable pathway, and PLS with the maximum ΔG of recently reported NRR ASCs are summarized in Table 1 based on categories of the substrates (note that * denotes an atomic active site in this review unless otherwise specified).

High-Throughput DFT Calculations
Due to the various TM elements, coordination environments, and substrates, a comprehensive search for NRR ASCs can lead to a large number of combinations, appealing for highthroughput DFT calculations. A typical example is TM atoms supported on N-doped graphene (Table 1), where   [89] many types of coordination are thermodynamically possible and can be controlled experimentally (C 3 , C 2 N 1 , C 1 N 2 , N 3 , C 4 , C 3 N 1 , C 2 N 2 , C 1 N 3 , and N 4 , Fig. 4a) [60][61][62][63][64][65]. Du, Wang, and coworkers proposed a two-step high-throughput screening strategy for designing NRR ASCs with a significantly lower computational cost [62]. With the nine above-mentioned types of coordination, 30 TM element selections, and two adsorption patterns (end-on and side-on), 540 ASCs constituted the database for screening. In the first step, the authors used the N 2 adsorption energy (ΔE N 2 ) and free energy change required to break the N≡N triple bond (the first hydrogenation step, Δ G N 2 −N 2 H ) as descriptors to rule out ASCs with low activity. In the second step, the descriptors were the free energy for NH 3 desorption and the hydrogenation from -NH 2 to -NH 3 (ΔG NH 2 −NH 3 ). The detailed threshold values are listed in Fig. 4b, and these steps theoretically require a large energy input and are usually the PLS of the NRR (Table 1). A total of 97 ASCs were selected in the first step ( Fig. 4c) because they strongly interact with N 2 and can hydrogenate N 2 to N 2 H with high performance (red region), while 10 highly active ASCs were selected in the second step (Fig. 4d). Among the ASCs, W 1 C 3 exhibited the best NRR performance with a U L value of − 0.25 V through the enzymatic mechanism (Fig. 4e). This benchmarking work provided an example of high-throughput DFT calculations that accelerated the discovery of NRR ASCs.

Machine Learning
Even though the above-mentioned high-throughput DFT method can reduce the time cost for screening NRR ASCs, further optimization of the screening process is indispensable when the dataset contains hundreds or thousands of systems. Machine learning (ML), an interdisciplinary approach, has been widely used in screening energy materials, including lithium-ion conducting materials [120,121], perovskites [122], and CO 2 RR/HER electrocatalysts [123]. Specifically, for electrocatalyst design, ML approaches can build the relationship between catalytic performance and intrinsic structural/electronic/bonding properties by analyzing the feature importance. Kim and coworkers used the deep neural network (DNN) ML approach for the high-throughput screening of NRR ASCs on boron-doped grapheme [71]. They first narrowed down the number of key features for describing a Different types of coordination environments for TMs supported on N-doped graphene, b schematic flowchart of the two-step screening process for NRR ASCs, c first-step and d second-step screening results for TMs supported on N-doped graphene, and e NRR free energy diagram of the enzymatic mechanism for W 1 C 3 at different applied potentials. Reprinted with permission from Ref. [62]. Copyright © 2018, Wiley-VCH NRR performance to seven (including electronegativity, the atomic number, and the atomic radius) by DNN and used them to select eligible candidates for the NRR by the criteria described in Ref. [62] (Fig. 5a). Then, after training the data with a light gradient boosting machine (LGBM) model, the ML results were similar to the DFT-calculated adsorption energy values (Fig. 5b). Furthermore, by analyzing the correlations between the features (Fig. 5c), the authors found that the most important features are the TM coordination number and number of hydrogen atoms in the DNN and LGBM models, respectively. Finally, among ASCs screened out by DFT and ML, CrB 3 C 1 showed the smallest overpotential of 0.13 V (Fig. 5d). The authors also extended their methodology to the NRR on TM borides, defective 2D materials, and ASCs on π-conjugated polymers for fast screening and descriptor discovery [124]. These pioneering works pave the way for combining DFT with ML to reduce the computational cost of screening NRR ASCs, but more effort in analyzing different ML models and further refinement of the screening procedure are necessary.

Activity Descriptors of NRR ASCs
Activity descriptors play a significant role in understanding the reaction mechanisms and constructing the relationships between geometric structure, electronic properties, and catalytic activity. Therefore, developing activity descriptors is an urgent task for catalyst design [125]. To this end, descriptors of hydrogen and oxygen electrocatalysts have been intensively investigated since Hammer and Nørskov came up with the d-band theory in the 1990s [126], and more sophisticated descriptors such as surface distortion, d-band shape, and generalized/orbitalwise coordination number have been developed for the HER, oxygen evolution reaction (OER), and ORR [125,127]. The well-defined active sites of the ASCs provide an ideal platform for developing accurate structure-activity relationships and descriptors. For Reprinted with permission from Ref. [71]. Copyright 2020, the Royal Society of Chemistry instance, a universal descriptor φ based on valence electrons in the d orbital and electronegativity was proposed to describe HER/OER/ORR activity of ASCs supported on N-doped graphene [128]. However, unlike hydrogen/oxygen electrocatalytic reactions, the NRR is a more complicated process with a much larger number of possible intermediates and pathways, making the discovery of activity descriptors a more challenging task.

Adsorption Energy Descriptors
As proven by many reports [73,83,99], similar to those of other catalytic reactions, such as the OER, ORR, and CO 2 RR, the adsorption energies of *N x H y NRR intermediates on ASCs may also follow linear scaling relations. In this case, the NRR activity descriptor based on the adsorption energy of only one intermediate (instead of a complicated multiple-parameter problem) is possible, so the rational design of NRR ASCs can be greatly simplified. For example, for ASCs supported on Ti 3−x C 2 O y and Ti 2−x CO y MXenes, the limiting potential and adsorption energy of *NNH satisfy a linear relationship [105]. The descriptors of different substrates may also vary in their ΔE *N (ASCs on PtS 2 [101] and graphdiyne (GDY) [76]) and ΔG *NH (ASCs on MoS 2 [99] and with MB 4 /MC 4 /MN 4 coordination [73]). A systematic study by Qiao et al. built a system for describing the NRR activity of ASCs based on the scaling relations between the adsorption energies of *NNH, *NNH 2 , *NH, *NH 2 , and *N [83]. By plotting the NRR limiting potential as a function of ΔE *N (Fig. 6a, b), the authors showed that ΔE *N can act as a facile descriptor for predicting the NRR limiting potential, since the DFT-calculated data points (scatter plots) and predicted limiting potential (line graphs) agreed well. When the ΔE *N value is located in different regions, the PLS and limiting potential for the NRR change accordingly. Furthermore, based on the scaling relations, the authors provided a more intuitive contour plot of the NRR limiting potential as a function of the Gibbs free energies of *NNH and *NH 2 (Fig. 6c). Ru@g-C 3 N 4 , which was located very close to the center of the red region due to the scaling relations, had the lowest limiting potential of − 0.33 V (Table 1). However, the adsorption energy descriptors heavily depend on the type of TM elements; for example, for the early TM elements (Fig. 6a) and late TM elements Society. e The energy barrier of the PLS as a function of the integrals of the unoccupied d states (UDSs) for TMs on PtS 2 . Reprinted with permission from Ref. [101]. Copyright 2020, American Chemical Society. f Relationship between integrated descriptor EL(l, θ, Z) and ΔG of *N 2 + H + → *NNH. Reprinted with permission from Ref. [113]. Copyright 2020, American Chemical Society on g-C 3 N 4 ( Fig. 6b) have different U L − ΔE *N relations. In addition, the ligands also influence the linear relations, as indicated by the different dotted lines for TM@N 3 , TM@ N 4 , and TM@g-C 3 N 4 in Fig. 6c. These factors call for more universal descriptors for NRR ASCs.

Electronic Structure Descriptors
Another important descriptor correlating the intrinsic properties of ASCs with their catalytic performance is electronic structure. In many systems, the well-known d-band center and the highest peak position of the density of states (DOSs) below the Fermi level have a poor linear relationship with NRR performance [83]. Thus, there has been a search for more sophisticated electronic structure descriptors. For example, in Qiao's work, following the limiting potential-ΔE *N relationship, ΔE *N further exhibits a linear relationship to the integrated crystal orbital Hamiltonian population (ICOHP) (Fig. 6d) [83]. Increased ICOHP suggests increased filling of the antibonding orbital population and decreased *N adsorption, and thus, a linear relation is formed. Wu et al. used the percentage of empty d orbitals as an electronic structure descriptor, and a higher proportion of empty d orbitals led to a lower NRR overpotential [94]. In another work, Cai et al. [101] correlated the ΔG of the NRR PLS with the integral of the unoccupied d states (UDSs) and found that the Ru on PtS 2 , which had a high UDS integral value, exhibited the best NRR performance (Fig. 6e). In summary, the electronic structure of ASCs, especially the overlapping bonding/antibonding metal-nitrogen orbitals, plays a decisive role in the adsorption strength of the NRR intermediates and in the overall activity.

Other Descriptors
In addition to those mentioned above, other descriptors combining electronic structure and atomic structure parameters were also explored. According to Xu and Chen et al., an integrated NRR descriptor EL(l, , ASCs on metal diborides was identified by high-throughput DFT calculations and mathematical regression technology ( Fig. 6f) [113]. Here, l, θ, Z abs , and Z B denote the metal-nitrogen bond length, the metal-nitrogen-boron bond angle, the atomic number of adsorbed metal, and the atomic number of boron, respectively. EL(l, θ, Z) exhibited a linear relationship with the ΔG of the PLS, proving its effectiveness in predicting NRR performance. Li and coworkers used the dipole moment of the N≡N triple bond as a descriptor and found that N≡N dipole moments have a similar trend to that of the N 2 adsorption energy on TMs anchored on Pc [114]. Nevertheless, all the above-mentioned activity descriptors are only applicable to one group of ASCs, and more universal activity descriptors for NRR ASCs are still lacking.

Selectivity Descriptors for NRR ASCs
The selectivity of NRR ASCs is greatly affected by the HER side reaction, and the active sites for the NRR can be poisoned by adsorbed hydrogen (*H), leading to poor NRR selectivity. The prevailing theoretical selectivity descriptor is the adsorption Gibbs free energy difference between *H and *N 2 or *NNH, depending on whether N 2 adsorption or NNH formation is the first step of the NRR [73]. For example, in Fig. 7a, the ASCs located in the blue regions exhibit negative ΔGN 2 -ΔG H , indicating that the adsorption of N 2 is more favorable. Hence, they have higher NRR selectivity [62]. In contrast, Xiao et al. [73] considered the concerted mechanism where N 2 + H + + e − → *NNH is the first step of NRR, and the corresponding selectivity descriptor is ΔG[N 2 (g) + H + + e − → *NNH] − ΔG(H + + e − → *H) (Fig. 7b). In a simplified model where the HER is the only competing reaction of the NRR and the mass and electron transfer are not rate-determining factors, the Faradaic efficiency (FE) for the NRR can be estimated by the Boltzmann distribution [63,83]: where δG is the difference between the HER and NRR PLS and k B and T represent the Boltzmann constant and temperature, respectively.

Perspective and Challenges for Theoretical Design of NRR ASCs
Despite all the progress, there are several challenges in the theoretical design approaches to NRR ASCs. First, the reliability of the DFT-calculated limiting potential should be carefully examined. Even for the same system, different conclusions are sometimes drawn in different studies. For example, for ASCs supported on g-C 3 N 4 , at the same level of theory (GGA-PBE with van der Waals correction), different authors reported that Ti (U L = − 0.51 V) [81], W (U L = − 0.35 V) [82], Ru (U L = − 0.33 V) [83], and Pt (U L = − 0.24 V) [84] on g-C 3 N 4 are the best systems for the NRR (Table 1), leading to questions about the standardization of such Gibbs free energy calculations. The origins of this discrepancy may include the use of slightly different initial geometric structures and input parameter settings for the DFT calculations. Second, the neglect of several complex factors in the theoretical calculations may result in a large gap between calculations and experiments; such factors include the following: (1) Kinetic analysis. Most works only focus on thermodynamic analysis, while kinetic analysis has long been overlooked because of the difficulty in the computation of electrochemical activation barriers. In a pioneering work by Azofra et al. [26], Janik's DFT methodology was used in NRR calculations to estimate the activation barrier for species *A: G act (U) = G 0 act + βF(U − U 0 ), where G 0 act is the activation barrier for *A hydrogenation, U and U 0 are the applied electrode potential and energy of *A + *H reactants to *A + H + + e − , respectively, β is set as 0.5, and F is the Faraday constant [129]. Another solution is Nørskov's charge-extrapolation method, in which the total charge of a system remains constant and the potential can change during an electrochemical step [130,131]. These methods can also be used to estimate the kinetic energy barriers of ASCs in the NRR, as such barriers play an important role in NRR activation [73,81,105]. For instance, *N 2 → *N 2 H, which has the highest activation barrier of 1.12 eV, is the rate-determining step for Mo@Ti 3−x C 2 O y ASCs (Fig. 8a), and this value is much larger than the corresponding thermodynamic barrier (0.44 eV) [105].
(2) Simulation of the NRR catalytic performance under realistic conditions. A turnover frequency (TOF) map can be calculated at different temperatures and pressures by microkinetic modeling, providing a prediction of the NRR reactive performance of ASCs under realistic conditions. For example, within the pressure range of 1-100 bar (1 bar = 100 000 Pa) and temperature range of 300-1 000 K, the TOF of Mo-BHT ASCs  (Fig. 8c, d) [70]. As such, the calculation results of the CCM should be reexamined with the CPM for the NRR to consider the charge effects and simulate more realistic situations. (4) Solvation effects. Solvation corrections are also needed to narrow the gap between theoretical calculations and experimental conditions, and applying the implicit solvation model to correct DFT calculation results can supplement the current calculations [133]. Solvation effects are usually neglected in theoretical modeling of the NRR because their changes are usually smaller than 0.1 V [14], but explorations are needed to obtain more accurate results. Recently, Carter's group employed a mixed implicit/explicit solvation model to study the NRR performance of a single TM atom-doped g-GaN monolayer, where explicit H 2 O molecules were added in the first solvation shell, and the implicit solvation model was also applied for the long-range solvation effects [134]. Their results suggest that the solvation effect in the NRR should not be neglected for more rigorous theoretical investigations.
These are only some of the urgent challenges for the theoretical design of NRR ASCs. Fortunately, with the development of computing power and new approaches, the gap between theory and experiments is rapidly becoming narrower.

Practical Applications of TM ASCs in the NRR
In this section, we review the current progress in the experimental realization of TM ASCs in the NRR. Based on the type of TM elements (active sites), we divide the cases into three categories: noble metal (Au, Ag, Pt, and Ru) ASCs; nonprecious metal (Fe, Co, Ni, Cu, and Mo) ASCs; and rareearth metal (Y and Sc) ASCs. The metal content, electrolyte, and key NRR performance indicators (NH 3 yield rate and FE) for each ASC are summarized in Table 2 for comparison. Since DFT calculations are effective and common in the design of ASCs and reveal the origin and mechanism of their NRR performance, the theoretical limiting potential and PLS values (at the DFT-GGA level) for the best samples in experiments are also listed in Table 2. Unless specified otherwise, all the potential values in this review are referenced to the reversible hydrogen electrode (RHE). It is noteworthy that most of the synthesized ASCs are supported on nitrogen-doped (porous) carbon materials (NC or NPC) through metal-nitrogen (M-N x ) coordination bonds due to the outstanding electrical conductivity, abundant active sites with high porosity, and good mechanical strength and stability [135][136][137], and the corresponding synthetic strategies, including wet-chemistry methods, spatial confinement, and coordination site construction strategies (especially stabilizing ASCs by MOFs and their derivatives), are used in fabricating ASCs on NC [50,138]. Very recently, other support materials for ASCs have been developed for the NRR, including Pt on WO 3 [139], Fe on nitrogen-free lignocellulose-derived carbon [140], Fe on MoS 2 [141], and Mo on GDY [142], and the family of NRR ASCs has been enriched by these processes.

Noble Metal ASCs
Catalysts based on noble metals such as Au, Ag, Pt, and Ru usually exhibit superior catalytic performance. For instance, Pt is an ideal catalyst for the HER and ORR, while the noble metal oxides RuO 2 and IrO 2 exhibit high OER activity [168].  However, their high cost and scarcity greatly hamper the large-scale application of noble-metal-based catalysts.
Atomically dispersed ASCs provide a solution to increase the AUE and selectivity.
Pioneering studies on the electrochemical NRR confirmed that the activity of Au-based catalysts [28][29][30][31] is possibly due to the high N 2 activation and HER suppression ability of Au. It is reasonable to conjecture that Au ASCs exhibit higher selectivity and AUE while retaining high NRR activity. Based on this conjecture, Wu and coworkers downsized Au nanoparticles (NPs) to Au ASCs supported on C 3 N 4 (Au 1 -C 3 N 4 ) by adsorbing C 3 N 4 with HAuCl 4 and reducing them with H 2 [143]. Au 1 -C 3 N 4 exhibited a high FE value of 11.1% and an NH 4 + yield rate 22.5 times as high as that for Au NPs due to the high AUE of ASCs. Oschatz et al. NRR tested Au ASCs and Au NPs on NPC as well as metal-free NPC for the NRR, and they concluded that Au ASCs showed stable NH 3 production and much-improved FE [144].
The rational design of ASCs based on theoretical calculations can greatly reduce the cost of trial-and-error experiments. Based on Luo et al.'s calculations, the end-on N 2 adsorption configuration and oblique *NNH admolecules favor high NRR performance. Thus, Ag-N 4 ASCs on NC (SA-Ag/NC) were designed, and their instability inhibited the side-on *NNH configuration [145]. The fabricated Ag ASCs showed a high NH 3 yield rate of 270.9 µg h −1 mg −1 cat and FE of 21.9% at − 0.65 and − 0.6 V, respectively, as well as long-term (60 h) stability. This work highlighted the importance of theoretical calculations in the targeted design of novel ASCs for NRR.
Conventionally, Pt-based catalysts are considered unsuitable for the NRR due to their high HER activity, which suppresses the NRR [31,169]. However, a different story was depicted by Shen et al., who downsized Pt NPs to isolated atoms supported on a WO 3 substrate (Pt SAs/WO 3 ), which exhibited excellent NRR activity [139]. Pt SAs/WO 3 were prepared by in situ photodeposition combined with a hydrothermal method, and the atomic dispersion could be observed by aberration-corrected high-angle annular darkfield scanning transmission electron microscopy (HAADF-STEM) (Fig. 9a). Diffuse reflectance infrared Fourier transform spectroscopy using CO (CO-DRIFTS) exhibited a CO adsorption band at 2 114 cm −1 , which can be attributed to the CO top adsorption configuration, further proving the formation of Pt ASCs (Fig. 9b). At − 0.2 V, Pt SAs/WO 3 showed an optimized production rate of 342.4 µg h −1 mg −1 Pt and an FE of 31.1%, which are more than 10 times higher than those of the NRR-inert Pt NP catalysts (Pt NPs/WO 3 ). DFT calculations suggested that N 2 adsorption on the Pt-3O site may be less hindered by *H adsorption, leading to suppressed HER. This work provides another example in which the rational design of ASCs can lead to the entirely different scenario of an NRR with high efficiency and selectivity.
Unlike Ag and Pt, Ru metal is conventionally identified as a second-generation catalyst for ammonia synthesis after Fe [14,[170][171][172][173]. However, the NH 3 production rate and FE for Ru NRR catalysts are still not satisfactory [170]. In 2018, Zeng's group reported the synthesis of Ru ASCs supported on N-C (Ru SAs/N-C) by high-temperature pyrolysis of a Ru-containing zeolite imidazolate framework (ZIF-8) (Fig. 9c) and investigated their applications in the NRR [146]. Extended X-ray absorption fine structure (EXAFS) spectra showed that compared with Ru NPs/N-C, the only peak for Ru SAs/N-C can be attributed to Ru-N bonds, and no peaks corresponding to Ru metal and oxides were  [167] observed, indicating that Ru SAs/N-C are atomically distributed (Fig. 9d). The designed Ru ASCs could achieve an FE of 29.6% (Fig. 9e) and a record-high NH 3 production rate of 120.9 µg h −1 mg −1 cat. (Fig. 9f). In a recent publication, Ma et al. used a theoretical simulation-directed strategy to design Ru ASCs supported on Cu oxides; this strategy involved weak single atom-substrate interactions that led to improved N 2 adsorption and reduced alkaline HER (compared to Ru-N ASCs), which was further proven by experiments [148]. They also proposed two descriptors-water dissociation energy and d-band center-for NRR and alkaline HER activities to accelerate ASC screening. In another work by Sun et al., the addition of ZrO 2 was shown to dramatically increase the NRR FE of Ru ASCs from 9% to 21% (Table 2) [147]. DFT calculations revealed that while Ru supported on NC exhibited high theoretical NRR performance, Ru@ ZrO 2 can provide high NRR/HER selectivity due to the less hindered N 2 adsorption by H (Fig. 9g, h), in line with experimental results. Therefore, Ru@ZrO 2 /NC ASCs have the bifunctional capability to simultaneously enhance the NH 3 production rate and FE. The low Ru loading (< 0.2%) and outstanding NRR activity (Table 2) endow Ru ASCs with great potential in NRR applications. However, the relatively low yield of Ru single sites and difficulty in the synthesis [147] appeal for more research into this area.

Non-Precious Metal ASCs
Substituting noble metal elements with nonprecious metal elements can reduce the cost of ASCs. Among the nonprecious metal elements, Fe and Mo attract attention because they act as active centers in natural FeMo nitrogenases (FeMo-co) [7,9,174], and Fe/Mo ASCs may also exhibit high NRR activity. For example, Yan and coworkers prepared Fe ASCs by modulating polypyrrole-iron complexes [149]. The synthesized Fe SA -N-C can achieve striking FE of 56.55% at 0 V (Fig. 10a), which is the highest among all reported ASCs for the NRR to date (Table 2), and the competing HER can be suppressed (Fig. 10b). Molecular dynamics (MD) simulations indicated that the Fe SA -N-C structure can attract N 2 molecules with a considerably small energy barrier of 2.38 kJ mol −1 (Fig. 10c), promoting subsequent NRR thermodynamics. In another work, Liu and coworkers reported Fe ASCs with high NRR performance in a neutral environment (0.1 M phosphate buffer solution) [150]. The EXAFS and DFT results ascribed the performance to the Fe-N 4 configuration, which can effectively activate N 2 .
In addition to ASCs supported on NC, Zeng's group used Fe single atoms to decorate HER-active MoS 2 edge sites, and the Fe-MoS 2 ASCs with lower PLS (0.37 eV) for NRR (0.50 eV for MoS 2 ) and higher HER barrier resulted in much-improved NRR reactivity [141]. Zhao et al. reported the facile synthesis of Fe ASCs with a new nitrogen-free coordination environment (Fe on lignocellulose-derived carbon, Fe SAs/LCC) [140]. X-ray absorption near edge structure (XANES) spectroscopy implied the atomically dispersed nature of Fe by the absence of Fe-Fe bonds, and it indicated the presence of Fe-O bonds (Fig. 10d). Additionally, the EXAFS fitting results proved that the Fe single atoms formed a novel Fe-(O-C 2 ) 4 bond (Fig. 10e), while DFT calculations suggested that such a structure can promote N 2 activation through a backdonation mechanism. Consequently, electrochemical tests with ASCs immobilized on a glassy carbon electrode showed near-record NRR performance (yield rate of 307.7 µg h −1 mg −1 cat. and FE of 51.0%). Most recently, Fe ASCs on N,O-doped porous carbon were designed and fabricated, and their outstanding NRR activity originated from the optimized charge transfer between Fe and  [140]. Copyright 2020, Wiley-VCH. f 3D spatial contour plot for the HOMO and the LUMO near the Fermi level for Mo 0 /GDY. Reprinted with permission from Ref. [142]. Copyright 2019, American Chemical Society. g NH 3 yield rate and FE of Mo SAs/Mo 2 C/NCNTs, Mo 2 C/NCNTs, and Mo SAs/NCNTs at different potentials, and h NH 3 yield rate with increasing cycle numbers at − 0.25 V potential. Reprinted with permission from Ref. [161]. Copyright 2020, Wiley-VCH adjacent O atoms [154]. This work highlighted the influence of the local coordination environment and the ligands on the NRR performance of ASCs.
In 2019, cost-effective and highly efficient Mo ASCs (SA-Mo/NPCs) were first utilized in NRRs [160]. Owing to the highly exposed active sites with the designed hierarchical porous carbon framework, SA-Mo/NPC exhibited a high NH 3 yield rate of 34.0 µg h −1 mg −1 cat. and FE of 14.6% as well as high stability. Li's group designed zerovalent Mo ASCs on graphdiyne (Mo 0 /GDY), which can be synthesized through a facile and scalable method, and the Mo ASCs have novel bifunctional NRR/HER applications that are not reported elsewhere [142]. Compared with traditional ASCs, zerovalent ASCs with excellent stability and performance can accelerate fundamental investigations of the electrocatalytic mechanisms [44]. Calculations suggested that the Mo sites close to the corner of the alkyne ring serve as electron-rich centers for catalysis ( Fig. 10f) with prominent electronic structure and NRR energetics. The XANES spectra of the Mo 0 /GDY and Mo references are almost identical, confirming the zerovalent nature of the ASCs. M Na 2 SO 4 , Mo 0 /GDY exhibited outstanding performance under ambient conditions, while in the non-N 2 -saturated electrolytes, high HER activity was also achieved. This work was the first to apply GDY-based ASCs to a bifunctional NRR/HER in different environments.
Assembling Mo ASCs with Mo 2 C nanoparticles on nitrogen-doped carbon nanotubes (Mo SAs-Mo 2 C/NCNTs) can improve the NRR activity and selectivity more than either the nanoparticles or nanotubes alone due to the synergistic mechanism, as confirmed by Wang et al.'s theoretical calculations and experiments [161]. DFT calculations suggested that the Mo 2 C(101) surface is more catalytically active for the NRR, while Mo ASCs show higher HER reactivity. Following the newly proposed surface hydrogenation mechanism [175], hydrogen coverage of Mo ASCs can reinforce N 2 activation on the surrounding Mo 2 C, providing a cooperative effect between Mo ASCs and Mo 2 C. Electrochemical NRR tests verified the theoretical calculations that Mo SAs-Mo 2 C/ NCNTs exhibited NRR performance several times as high as that of either Mo SAs/NCNTs or Mo 2 C/NCNTs alone (Fig. 10g) and a small amount of decay during the reusability test (Fig. 10h). As illustrated by this work, the synergistic effect between ASCs and other types of active sites may enhance the NRR performance. In addition to Fe and Mo, nonprecious metal Co [156][157][158], Ni [159], Cu [162,176], and Mn [163] ASCs have also been examined as catalysts for the NRR.

Rare-Earth Metal ASCs
Catalysts based on rare-earth metal elements are conventionally considered ineffective because the multi-shell nature of electrons from rare-earth elements may lead to strong binding with reactant molecules that are difficult to release [164]. Nevertheless, after nitrogen and carbon coordination, electronic structure modulation can endow rare-earth metal ASCs with greatly enhanced catalytic performance compared with rare-earth metal NPs, as proven by Shui's group [164]. Y 1 /NC and Sc 1 /NC ASCs were synthesized by using Y/Sc-doped ZIF-8, followed by 1 000 °C pyrolysis. Both DFT calculations (Fig. 11a) and Fourier transform EXAFS (FT-EXAFS) spectra (Fig. 11b) confirmed the six-coordinate structures (Y 1 -NC, preferably with N 3 C 3 and Sc 1 -NC with N 6 coordination), which are different from the prevailing TM-N 4 coordination structures of other TM-based ASCs because of the large atomic radii of rare-earth elements. The synthesized Y and Sc ASCs exhibit high ammonia production rates of 21.8 and 19.2 µg h −1 cm −2 (Fig. 11c, d) and outstanding FE of 12.1% and 11.2% at − 0.1 V, respectively, even under strict experimental control, indicating that the coordination environment can greatly enhance the inert nature of rare earth metal-based catalysts. Finally, DFT calculations further proved that Sc and Y ASCs have lower N 2 activation barriers than the corresponding oxides (Fig. 11e). This interesting work extended the current family of ASCs to the rare-earth elements (Fig. 11f).

Perspectives and Challenges
In our closing remarks, we summarize the current reports on NRR ASCs categorized by the element type (Fig. 11f). Nearly half of the reports are focused on nonprecious Fe ASCs, and among them, Fe SA -N-C exhibits the highest FE of 56.55% [149] and Fe SAs/LCC or Fe-(O-C 2 ) 4 shows the highest NH 3 yield rate of 307.7 µg h −1 mg −1 cat [140] . Considering the lower cost than noble-metal-based catalysts and stability, Fe ASCs are among the most promising nextgeneration electrocatalysts for the NRR. In addition, Mo, Co, and Ru ASCs have attracted the attention of researchers. The coordination environment also has a significant influence on the catalytic activity, and the rational design of ASCs based on the coordination environment calls for the combination of theoretical calculations and characterization techniques (e.g., XANES and EXAFS), which is particularly important for ASCs (with electrocatalytic activity very sensitive to the local coordination environment) compared with other electrocatalysts.
A more urgent task for the development of the electrocatalytic NRR (not only with ASCs as the electrocatalysts but also with other electrocatalysts) is the identification and suppression of false positives caused by other nitrogen Page 17 of 32 6 resources in the experiments, such as nitrogen-containing compounds in the gas stream, environment, or even the electrocatalysts themselves. Standard experimental procedures to eliminate false positives have been suggested [48,49,177]. Nevertheless, according to Choi et al.'s [178] evaluation, only two of 127 papers on the aqueous NRR up to April 2020 could satisfy two of the three criteria they set (sufficiently high NH 3 yield rate; sufficient and reliable 15 N experiments; rigorous control and quantification of NO x ), and none met all three criteria due to the neglect of key experimental controls (especially NO x contamination), leading scientists to doubt the reproducibility and reliability of the reported NRR electrocatalysts and the true catalytic origin. Therefore, standardization of experimental protocols is very important for the healthy development of this newly emerging field, and similar to other fields, establishing internationally recognized certification laboratories for electrocatalytic NRR may also be a solution.
With respect to the theoretical design prior to experiments, the comparison between the adsorption of N 2 and that of O, NH 3 , and NO x is seldom reported, which is also a long-overlooked yet necessary approach to evaluating the N 2 adsorption ability with the influence of those contaminants before experimental realization. In addition, the thermodynamic stability of the catalysts can also be predicted by theoretical calculations. All these challenges must be sufficiently addressed before the electrocatalytic NRR can reach the level of industrial ammonia synthesis.

Rational Design of Double-Atom Catalysts
Despite all the advantages of the atomically dispersed nature of single-atom catalysts, they also suffer from structural simplicity; that is, single sites may not provide both an ideal yield rate and FE [56]. For complicated multielectron-transfer reactions such as the NRR, tuning the adsorption strength of intermediates on active sites is an urgent issue for the rational design of highly efficient catalysts with high selectivity. Specifically, for the NRR, scaling relations between the adsorption energy of *N x H y radicals hinder the optimization of theoretical NRR overpotential [83], so the circumvention of scaling relations is one of the most significant tasks.
Double-atom catalysts (DACs), or atom-pair catalysts, such as metal dimers supported on substrates have emerged as a newly extended group of ASCs with practical applications in the HER [179,180], OER [181,182], ORR [182][183][184][185][186], and CO 2 RR [187][188][189]. Theoretically, metal dimer sites, unlike single atomic sites, may break the intrinsic scaling relations and optimize the adsorption behavior of intermediates, thus surpassing the theoretical limit of the catalysts, as revealed by the pioneering theoretical works in OER/ORR [190] and CO 2 RR [191]. However, for the NRR, both experimental and theoretical benchmarking investigations about the rational design of DACs are scarce, possibly due to the difficulty in controlling their synthesis and the complexity of the NRR mechanism.
Li et al. [165] synthesized FeMo dimers supported on N-doped graphene (FeMo@NG) by an in situ sacrificial template anchoring method. The NH 3 yield rate of 14.95 µg h −1 mg −1 at − 0.4 V and FE of 41.7% at − 0.2 V were recorded in an electrochemical test (Table 2), and they were much superior to those of the corresponding single-atom Fe and Mo catalysts (Fe@NG and Mo@NG). DFT calculations indicated that FeMoN 6 structures act as the active sites for the NRR and have a lower limiting potential (− 0.906 V) than Fe@NG and Mo@NG. A very recent publication highlighted the importance of incorporating Cu into Pd/NC ASCs to form PdCu/NC DACs in electrocatalytic NRR [166]. DFT calculations indicated that the incorporation of Cu can shift the partial density of states of Pd toward the Fermi level and promote d-2π* coupling between Pd and N 2 , leading to much-improved N 2 adsorption and protonation with suppression of the HER. The experimental results validated the conclusion that PdCu/NC [(24.8 ± 0.8)% FE and (69.2 ± 2.5) µg h −1 mg cat.
−1 ] outperformed Pd/NC in the NRR. These works laid a solid foundation for utilizing synergy of DACs to enhance the activity and selectivity for the NRR. However, more experimental efforts into the controllable synthesis of DACs are still in great demand to gain more fundamental insights into the mechanisms for the enhanced NRR performance, while theoretical design can compensate for the deficiency in experimental results (Table 3) [19].
In a comparative study, Jiang et al. [193] concluded that TM 2 -C 2 N DACs were more suitable for the NRR than TM-C 2 N single-atom catalysts, and Mn 2 -C 2 N exhibited the lowest limiting potential of − 0.23 V. Huang and coworkers screened homonuclear and heteronuclear DACs supported on 2D phthalocyanine (M 2 -Pc and MM'-Pc, Fig. 12a) by high-throughput DFT calculations using *N 2 H adsorption energy Δ E N 2 H * as a simple descriptor for NRR activity [56]. They first screened 30 M 2 -Pc structures with thermodynamic and electrochemical stability for their N 2 adsorption properties and found that eight of them can achieve N 2 chemisorption and activation. Among them, Ti 2 -Pc, V 2 -Pc, and Re 2 -Pc exhibited outstanding NRR U L values of − 0.75, − 0.39, and − 0.82 V, respectively (Fig. 12b), superior to the Ru(0001) benchmark (U L = − 0.98 V) with the highest theoretical NRR activity among metal surfaces [59]. Δ E N 2 H * was selected as a descriptor to obtain a volcanolike relationship between NRR U L and Δ E N 2 H * (Fig. 12c), and obviously, V 2 -Pc, which had the best U L among all the DACs, was located near the top of the volcano. The authors further extended the relationship toward heteronuclear DACs MM'-Pc (Fig. 12d), and in a selectivity investigation performed by using ΔG(H*) − ΔG(N 2 H*) as a descriptor (Fig. 12e), five DACs, Ti 2 -Pc, V 2 -Pc, TiV-Pc, VCr-Pc, and VTa-Pc, were found to simultaneously achieve both high NRR activity and selectivity. Deng et al. [203] studied the NRR performance of homonuclear DACs TM 2 -N 6 @G (Fig. 12f). They constructed a contour plot of limiting potential as a function of Gibbs free energy changes of the key elementary steps to screen for DACs with outstanding U L (Fig. 12g). The author further used the d-band center of 0-2 eV as a descriptor and found a volcano-like relationship between U L and the d-band center and further related them to ICOHP. Wang et al. [195] proposed a pull-pull effect in which dual-metal sites can maximize N 2 activation by pulling the two lone pair electrons of N 2 at both ends (Fig. 12h,  i). The constructed donor-acceptor couples can circumvent the scaling relations by synergistic effects that promote high NRR activity (Fig. 12j). The FeMo/g-C 3 N 4 DAC has a low limiting potential of − 0.23 V and high selectivity, while TiMo, NiMo, and MoW/g-C 3 N 4 are promising NRR photocatalysts. These theoretical works pave the way for designing DACs for high-performance NRRs.
In addition to DACs, several theoretical investigations proved that triple-atom catalysts such as Fe 3 /Al 2 O 3 (010) [209] and Fe 3 -GDY/Gra [210] are NRR-active. The synthesis of Ru 3 triple-atom catalysts was reported with a confined pyrolysis method within MOFs [211], and the application of double-or triple-atom catalysts for the NRR in the laboratory is promising.
In summary, replacing single-atom catalysts with DACs might be a potential solution to improve the theoretical NRR activity and selectivity. Nevertheless, the reported doubleatom electrocatalysts for the NRR are relatively rare compared with those for the ORR, water splitting, and CO 2 RR [19], and their NH 3 yield rate and FE do not show a pronounced improvement compared with those for single-atom catalysts. Currently, the discovery of DACs for the NRR is still driven by theoretical design, appealing for more efforts in this field.

Rational Design of Boron-Based Metal-Free ASCs
As stated above, metal-based catalysts, including TM ASCs, remain the most widely applied catalysts for the NRR. Despite recent progress, metal-based catalysts still suffer from relatively low efficiency because of the competition between the formation of metal-N and metal-H bonds.
Since the d-orbital electrons in TM centers are usually more inclined to form metal-H bonds [212,213], the FE of the NRR is greatly compromised. Metal-based catalysts may also raise stability, economic, and environmental concerns. Designing a new group of metal-free catalysts is therefore of great significance for the development of electrocatalytic and photocatalytic NRRs [214][215][216][217]. Pioneering works suggested that NPC [218] and B 4 C nanosheets [215] are superior NRR catalysts, but until now, metal-free NRR catalysts have generally been far less explored than metal-based catalysts.

Theoretical Design
For a typical TM-based NRR catalyst, TM elements with both empty and occupied d orbitals exhibit high N 2 activation ability, and the origin of effective N 2 activation can be ascribed to the "acceptance-donation" mechanism (also known as the "σ donation-π * back-donation" mechanism): TM atoms accept lone-pair electrons from N 2 and donate the d electrons to the antibonding orbital so that the inert N≡N bonds can be weakened (Fig. 13a) 6 Page 20 of 32 [214]. In a pioneering work by Légaré et al. [219], the d-block element boron was found to be NRR-active, and since N 2 is a weak Lewis base, boron atoms with empty orbitals can act as Lewis acid active sites for the NRR [219,220]. For sp 2 -hybridized boron atoms, N 2 can be activated through a B 2 N 2 motif with an end-on adsorption Reprinted with permission from Ref. [56]. Copyright 2020, American Chemical Society. f Structure of TM 2 -N 6 @G DAC, and g color-filled contour plot of the limiting potential of TM 2 -N 6 @G DACs as a function of ΔG *NNH and ΔG[*NH 2 + (H + + e − ) → *NH 3 ] through the distal pathway. Reprinted with permission from Ref. [203]. Copyright 2020, American Chemical Society. Schematics of N 2 activation on h singlemetal sites and i dual-metal sites through the pull-pull effect and j the binding energy of N 2 H versus N on g-C 3 N 4 -supported dual-metal sites and scaling relations for the TM surface. Reprinted with permission from Ref. [195]. Copyright 2020, American Chemical Society configuration [219]. Similarly, sp 3 -hybridized boron atoms can also activate N 2 effectively, as they can have one occupied and one empty sp 3 orbital for bonding via the "acceptance-donation" mechanism ( Fig. 13b), and the possibility of adopting a side-on adsorption pattern may induce different behavior in the NRR [214]. Following this concept, Wang et al. designed metal-free boron SASCs, i.e., B/g-C 3 N 4 through DFT calculations, where N 2 can be activated in both end-on and side-on configurations through the "acceptance-donation" mechanism, as illustrated by the negative adsorption energy as well as charge accumulation and depletion around the boron atom and N 2 (Fig. 13c, d) [214]. Detailed energetics analysis confirmed that through the side-on adsorption configuration and enzymatic NRR pathway, a very low onset potential of 0.20 V can be achieved (Fig. 13e).
Such design principles for boron-based metal-free ASCs can also be applied to NRR electrocatalysts. Sun's group designed 21 conceptual single-boron catalysts, including adsorbed (A), substituted (S), and lattice (D) boron on graphene, boron nitride, boron sulfide, black phosphorus, g-C 3 N 4 , and MoS 2 (Fig. 13f) [221]. By using the N adsorption energy on Ru E(N*) Ru as a reference [59], a volcanoshaped plot of the maximum Gibbs free energy change for the NRR elementary steps ΔG max versus E(N*)-E(N*) Ru was constructed (Fig. 13g), and boron adsorbed on graphene (G-A), boron nitride (BN-A), and substituted on h-MoS 2 (h-S1) were identified as the most promising NRR ASCs with ΔG max values of 0.31, 0.45, and 0.46 eV, respectively. For selectivity issues, the authors compared ΔG max HER with ΔG max NRR (Fig. 13h) and concluded that G-A and h-S1 have high NRR selectivity. By correlating ΔG max HER and Fig. 13 Design of boron ASCs for the NRR. Schematic of N 2 activation on a TM and b boron centers, difference charge density for c side-on and d end-on N 2 adsorption on B/g-C 3 N 4 , and e Gibbs free energy diagram of the NRR on B/g-C 3 N 4 through the enzymatic mechanism. Reprinted with permission from Ref. [214].  (Fig. 13i), they found that a small or negative charge on boron favors the NRR and a positive charge favors the HER, so the charge on boron can act as a descriptor for the design of boron ASCs for the NRR.
Another question for boron-based ASCs is whether the optimized boron species for the NRR are sp 2 -hybridized [222,223] or sp 3 -hybridized [214]. Yin et al. [224] provided theoretical insight by designing NRR ASCs with boron on C 2 N. Boron adsorbed on C 2 N is sp 3 -hybridized (denoted as B/C 2 N-a), while substituted boron is sp 2 -hybridized (denoted as B/C 2 N-s); thus, C 2 N is an ideal platform for analyzing the hybridization. B/C 2 N-s exhibited better activity for the NRR (all the steps along the alternating pathway are exothermic, Table 4) and higher HER suppression ability so that more sp 2 -hybridized boron sites can enhance the NRR ability. This work provides guidance for NRR catalyst design and is in line with experimental results [222,223]. More theoretical progress in the rational design of boronbased ASCs is summarized in Table 4.

Experimental Evidence
Experimentally, boron-doped graphene (BG) was first applied as sp 2 -hybridized NRR boron ASCs by Zheng et al. [222]. BG was fabricated by H 3 BO 3 /graphene oxide thermal reduction in H 2 /Ar mixed gas, and according to B 1s X-ray photoelectron spectroscopy (XPS) results, four structures, including BC 4 (lattice defect of graphene), BC 3 (B replacing a carbon atom in graphene framework), BC 2 O, and BCO 2 (B replacing edge/defect site carbon), could be observed after boron doping (Fig. 14a). The temperatureprogrammed desorption (TPD) curve indicated that boron doping at different levels (BG-1, BG-2) can significantly enhance N 2 chemisorption (Fig. 14b), and further electrochemical tests suggested that at a doping level of 6.2% (BG-1), boron ASCs exhibited NH 3 yield rates of 9.8 µg h −1 cm −2 and 10.8% FE (in 0.05 M H 2 SO 4 ), which were several times higher than those for undoped graphene (Fig. 14c, d). DFT calculations proved that the NRR is prohibited on graphene due to large energy barrier, and compared with those of BC 2 O (1.04 eV) and BCO 2 (1.30 eV), the BC 3 structure has the PLS (*NH → *NH 2 ) with the lowest energy barrier (0.43 eV), contributing to the high NRR activity of B ASCs (Fig. 14e). Boron-doped nanostructured diamond also performed well in the electrocatalytic NRR with an NH 3 yield rate of 9.8 µg h −1 cm −2 and an FE of 21.2%, and DFT calculations illustrated that boron atoms are active sites that can facilitate charge accumulation and reduce the free energy barrier of the PLS [234].

Perspective and Challenges
Apart from B-doped ASCs, S- [235], O- [236], and F-doped [237] carbon materials can also improve the NRR activity of pristine carbon because the electronegativity differences between S, O, F and C can create Lewis acid sites for activating N 2 and repelling H.
All these examples open up a new avenue for the rational design of metal-free ASCs for the electrocatalytic NRR. There are also several challenges for the design of metal-free ASCs, including: (1) Experimental synthesis and characterization of nonmetal atomic sites: Compared with metal-based ASCs, the control and characterization of nonmetal ASCs are difficult, so the atomically dispersed nature cannot be firmly verified. Nevertheless, with the aid of theoretical calculations, the NRR-active or NRR-inactive local structure of metal-free ASCs can be identified so that it can be determined whether single atomic sites act as active sites [222]. (2) Exploration of the interaction between boron, substrate, and N 2 molecules: While some studies have investigated the interaction between boron and carbon or N-doped carbon [214,222,[224][225][226][227][228], the interplay between other substrates and boron has not yet been fully illustrated. Additionally, whether substrates and ligands can affect the NRR activity of metal-free ASCs remains unclear and requires more attention. In addition, the effects of strain-, defect-, or doping-tailored electronic structures on the performance of metal-free NRR ASCs have not yet been revealed. (3) Design of metal-free ASCs other than those that are boron-based, such as F-doped carbon materials [237]: Recently, nonmetal single-phosphorus-atom catalysts were fabricated for the HER [238], and more evidence is needed for their applications in the NRR.

Conclusion and Outlook
In this review, we comprehensively summarized the recent progress in the theoretical and experimental design of ASCs for the NRR. The computational schemes, activity and selectivity descriptors, and experimental design were subsequently introduced, and corresponding progress was highlighted. Furthermore, the use of metal dimers instead of single-atom sites can synergistically improve the NRR activity, and boron-based ASCs can be a typical example of metal-free catalysts with similar or even superior NRR activity. Despite the vast progress that has occurred in recent years (Fig. 15), several challenges that may hinder the effective design of NRR ASCs with high activity and selectivity still remain; they include but are not limited to: (1) Standardization of the experimental and computational setup. As we discussed in Sect. 3.4, the nitrogen-containing compounds in catalysts, electrodes, and environments may cause false positives in the detection of ammonia and induce accuracy and reliability issues in pathway and corresponding energy changes for BC 3 , BC 2 O, BCO 2 , and C. Reprinted with permission from Ref. [222]. Copyright 2018, Elsevier the design of ASCs for the NRR [48,49,177]. Therefore, standardization of the control experiments is necessary for the elimination of contamination, and benchmarking protocols proposed by Anderson et al. [48] can be applied to ASCs as a solution; in particular, quantitative isotope measurements of 15 N 2 gas were used to increase the reliability and reduce the cost. In computational experiments, standardization is also important to ensure that the simulated activity and selectivity can reflect the real situations in experiments to a greater extent. (2) Developing and applying advanced theoretical approaches to narrow the gap between theory and experiments. These approaches include analysis of the kinetic energy barrier in the NRR [26], introduction of the charge effect [70] and the solvation effect, and microkinetic simulation of the TOF map at different temperatures and pressures [117]. In addition, more rigorous functional and van der Waals (vdW) corrections for the calculations of adsorption energy should be developed and applied. For example, the Bayesian error estimation functional with vdW (BEEF-vdW) can lead to a more accurate description of the energetics of catalytic reactions [239,240]. (3) Advanced instrumentation of experimental characterization and design of NRR ASCs. In situ and operando characterization of ASCs plays a significant role in understanding the underlying origins of NRR activity. For example, in situ XANES and EXAFS can be applied to detect the role of local coordination environments and electronic structures of active sites in the NRR process, and in situ DRIFTS or surface-enhanced infrared absorption spectroscopy (SEIRAS) [31,173] can be used to measure the intermediate species during NRR, both contributing to a more comprehensive mechanistic understanding. Further utilization of other advanced dynamic characterization approaches can be a great opportunity for the rational design of NRR ASCs. (4) Mechanistic investigation. Although most current theoretical investigations adopt a concerted PCET mechanism and CHE model in the calculations, there are controversies regarding whether concerted PCET or sequential PCET can better describe the proton and electron transfer process. For the CO 2 RR, Koper concluded that pH plays an important role in the competition between concerted and sequential PCET mechanisms [241]. For the NRR, more mechanistic investigations are needed. In addition, for the (111) facets of TM nitrides, a lattice-nitrogen-related Marsvan Krevelen mechanism is also possible for NRR [25]; thus, similar investigations of nitrogen-containing ASCs are needed. 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 licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence 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 licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.