Decoding active sites in high-entropy catalysts via attention-enhanced model

Abstract

Identifying active sites is decisive for optimizing catalysts, but this remains challenging, especially in high-entropy materials with multiple random sites. Here, we developed an attention-enhanced, multiobjective predictive model to precisely identify active sites and their corresponding overpotentials, a key parameter for catalytic activity. Using this model to predict the overpotential of oxygen evolution reaction (OER) process and doping formation energies in high-entropy CoOOH materials, we screened 17,500 catalysts and identified 8 with optimal catalytic activity. Subsequent automated synthesis and validation found a high-performance catalyst, TiFeNiZn-CoOOH, which exhibited an exceptional OER overpotential of 263 millivolts at 100 milliamperes per square centimeter. Feature importance and statistical analysis of more than 5 million structures confirmed that Zn consistently shows the highest active site occupation probability, and the [CoNiZn] coordination yields the lowest overpotential. Electronic structure analysis revealed that Zn activates gap states, critically lowering the OER energy barrier. This work paves a broad avenue for screening high-performance catalysts with identified catalyst structures.

AI sifts through millions of materials to find optimal catalysts for clean water splitting.

INTRODUCTION

The development of high-entropy (HE) catalysts with enhanced activity and stability remains a critical area of research in electrochemical applications, including water electrolysis, fuel cells, and CO₂ electroreduction (1–5). Optimizing catalytic performance in these domains is essential, and various strategies, including multielement doping, defect engineering, and edge structure modifications (6–9), have been explored to create active sites that enhance catalytic efficiency. Despite the success of these methods, much of the research still relies on trial-and-error approaches (10), lacking systematic frameworks for accurately predicting and identifying catalytic active sites within the vast HE design space (11). The primary challenge lies in understanding how atomic-level structural modifications influence catalytic activity (12) and how to design stable and efficient HE catalysts. Recent advances indicate that high-throughput computational methods (13) and in situ characterization experiments (14–16) can identify active sites. However, the effective application of these methods often requires extensive computational resources and sophisticated experimental setups.

To address these challenges, more effective models and methodologies are needed to predict active configurations responsible for catalytic behavior accurately. Artificial intelligence (AI) methods present a promising solution for the precise prediction of catalytic active sites and the design of high-performance catalysts (17–19). Deep learning models, such as EquiformerV2 (20), GemNet-OC (21), coGN (22), and PaiNN (23), show great potential in multiple material domains due to their attention mechanisms, which efficiently capture complex relationships between catalyst structure and catalytic activity. However, these models primarily rely on correlations between structure and energy within non-HE systems (24, 25), and their predictive accuracy diminishes in multielement doped systems. By integrating a small quantity of high-precision domain data for knowledge transfer (26, 27), the model’s prediction capability in complex systems can be substantially enhanced (28). This “pretraining–fine-tuning” strategy provides a practical solution to overcome the design bottleneck of HE catalysts.

Although AI models offer powerful predictive capabilities, they face substantial challenges when it comes to elucidating the structure-activity relationships that govern catalyst behavior (29, 30). Feature importance analysis using attention values can offer insights into the catalytic mechanisms (31), facilitating a deeper understanding of active sites (26). To identify high-activity configurations within vast compositional spaces is a challenging task due to the complexity of the design space and the interactions between multiple elements. By leveraging the constructed HE catalyst space for statistical analysis (32) and combining it with feature importance analysis, we can establish a clearer relationship between active sites and structural components, ultimately advancing the design of efficient and stable HE catalysts.

In this study, we developed an integrated catalyst discovery strategy that integrates advanced machine learning models with automated characterization techniques. Our approach combines multiobjective transfer learning, active site identification, and mechanistic exploration to efficiently unravel the underlying factors governing catalytic performance. Leveraging a dataset of 4822 HE CoOOH catalysts, we constructed an EquiformerV2-based model that accurately predicts catalytic properties, achieving an average absolute error of 3.6 meV atom⁻¹ for doping formation energy and 4.5 mV atom⁻¹ for theoretical overpotential. By screening 17,500 potential catalysts, we identified eight optimal candidates, and subsequent automated synthesis and characterization experiments revealed that three of these, TiFeNiZn-CoOOH, CrFeCuZn-CoOOH, and MnFeNiZn-CoOOH, exhibited exceptional oxygen evolution reaction (OER) activity. Furthermore, to understand the fundamental drivers of this performance, our feature importance and statistical analyses were expanded to a combined prediction space of more than 5 million structures, spanning binary, ternary, and quaternary systems. This comprehensive analysis revealed critical, generalizable design principles: Zn consistently dominates the active site coordination, and the [CoNiZn] local environment is consistently identified as the optimal, lowest-overpotential configuration across all levels of elemental complexity. These results highlight the precision of our active site identification approach and its interpretability for elucidating generalizable mechanistic relationships between local coordination environments and catalytic performance.

RESULTS

Because of the vast selection space of transition metals (TMs) and the inherent uncertainty of HE components, the number of potential oxygen evolution catalysts for HE CoOOH is theoretically infinite. The overpotential and doping formation energy serve as key parameters for evaluating the activity and stability of HE catalysts (33, 34), providing a solid theoretical foundation for the development of high-performance catalysts. Overall, the design and screening of high-performance catalytic systems is a multitask process (28, 35, 36). Using pretrained models with fine-tuning on a limited set of high-quality data enables a highly efficient and effective approach for the rapid computational design and screening of high-performance HE CoOOH catalysts.

Linking model predictions with automated experiments offers a highly efficient approach for accelerating catalyst discovery and optimization (37, 38). The AI models predict key properties, such as overpotential and doping formation energy, enabling the identification of promising catalyst candidates. These predictions are subsequently validated in an automated experimental laboratory capable of rapidly synthesizing and evaluating a large number of materials with high consistency and minimal human error (39), substantially expediting the screening and performance validation stages. This integration establishes an effective feedback loop between computational models and experimental results, facilitating continuous refinement and more efficient optimization of catalyst performance, ultimately reducing the time required for material discovery and development. By combining the strengths of both approaches, the development of high-performance catalysts is markedly accelerated.

The proposed workflow for HE CoOOH catalyst design, as depicted in Fig. 1, consists of four parts: multiobjective transfer learning, materials and active site search, support mechanism discovery, and automated laboratory (Auto-Lab) validation. We constructed an HE CoOOH dataset for transfer learning on EquiformerV2, based on the developed Post-Att Adapter, the overpotential (η_OER) and doping formation energy (E_doping) are trained using the EquiformerV2 model (Fig. 1A). With the fine-tuning model, we accurately identify theoretical high-activity materials and active sites within the material space (Fig. 1B). By integrating the interpretability method with the Post-Att Adapter, the attention values of doped TM active units are analyzed, providing valuable insights to help uncover underlying mechanisms (Fig. 1C). Last, the material screening process is validated using Auto-Lab for material synthesis and property characterization (Fig. 1D).

Fig. 1. Schematic diagram of HE CoOOH catalysts workflow.

(A) Overall model architecture of multiobjective transfer learning. The model takes HE-LDHs dataset as input for doping energy and overpotential tasks. The doping energy model is trained with slab structures, while the overpotential model uses slab, ^*OH, ^*O, and ^*OOH structures. (B) Materials and active site search. The fine-tuned model identifies materials that meet the criteria for overpotential and doping energy, selecting potential candidates. The lowest overpotential active site is located through a systematic search in the active site space. Light gray points represent the catalyst system, brown points denote screened candidates, and cyan points indicate potential active sites. (C) Support mechanism discovery. The Post-Att adapter of fine-tuning model combined with interpretability method is used to find the active unit and facilitate the discovery of mechanism. Small gray spheres represent oxygen atoms, large spheres represent TMs, and different colors represent different TMs. (D) Automated laboratory validation. Automated synthesis and characterization platform; the white-dotted box refers to the instrument used.

Multiobjective transfer learning model

We constructed a diverse computational dataset (4822 HE CoOOH), focusing on binary, ternary, and quaternary combinations of three-dimensional (3D) TMs (Ti, V, Cr, Mn, Fe, Ni, Cu, and Zn); dataset details are provided in note S1. On the basis of HE CoOOH dataset, we introduced a fine-tuning approach (Post-Att Adapter) applied to the EquiformerV2 model (Fig. 1A) within a multiobjective transfer learning framework. The model is trained for two target properties, η_OER and E_doping. The OER overpotential is determined by evaluating the thermodynamic stabilities of the reaction intermediates, following the model proposed by Norskov and co-workers (40), which uses a four-electron reaction mechanism to simulate the OER potential energy profile. In this approach, the free energies associated with the adsorption of key intermediates (e.g., ^*OH, ^*O, and ^*OOH) are systematically computed to identify the potential-determining step by

$${\mathrm{\eta }}_{\text{OER}}=\text{MAX} (\mathrm{\Delta }{G}_1,\mathrm{\Delta }{G}_2,\mathrm{\Delta }{G}_3,\mathrm{\Delta }{G}_4)-1.23 \mathrm{V}$$ (1)

Concurrently, the thermodynamic feasibility of doping is assessed through the calculation of the doping formation energy (41, 42), defined by

$$E_{\text{doping}}=E_{\text{doped}}-E_{\text{host}}-\mathrm{\Delta }{E}_{\text{dopant}}$$ (2)

where, E_doped denotes the total energy of the doped system calculated by density functional theory (DFT), E_host represents the energy of the pristine host material calculated by DFT, and ΔE_dopant is the reference energy of the dopant species. This methodology provides a robust framework for evaluating both catalytic performance and material stability, thereby facilitating the rational design of high-performance catalysts (2, 34). Detailed descriptions of the calculation η_OER and E_doping can be found in note S2. A Post-Att adapter is inserted after the attention layer, and only its weights are updated during training, while the initial weights of EquiformerV2 remain frozen. The output tensor from the attention layer is processed through the adapter, which includes an expanded and restored channel dimension to optimize the activation function for capturing high-dimensional features. This approach substantially reduces the number of trainable parameters compared to full-scale fine-tuning.

Two key performance metrics for HE CoOOH are presented in Fig. 2 (A and B), η_OER and E_doping. In the η_OER prediction task, we fine-tune the model using datasets ranging from 1000 to 4822 data points. With 1000 data points, the model has a mean absolute error (MAE) of 4.48 (mV atom⁻¹). As the dataset size increased to 2000, 3000, 4000, and 4822 data points, the corresponding MAE were 4.62, 4.71, 4.77, and 4.50 (mV atom⁻¹), respectively. This minor nonmonotonic trend is a statistical fluctuation inherent to the fivefold cross-validation sampling process, where minor variations in the composition of the random data splits can lead to small changes in the averaged MAE. These results show that the performance variation across dataset sizes is within 6%, indicating near-optimal performance, even with a small dataset. This stability is due to the efficient use of prior knowledge from the Post-Att Adapter fine-tuning approach. In contrast, the impact of dataset size on E_doping prediction was more pronounced. With only 100 data points, the model yielded an MAE of 4.32 (meV atom⁻¹), which improved to 3.59 (meV atom⁻¹) as the dataset size increased to 586 data points, showing consistent performance enhancement.

Fig. 2. MAE results under different dataset size fine-tuning and potential materials search.

(A) η_OER. (B) E_doping, where total represents the use of 4822 and 586, respectively. The baseline model is the other machine learning model, which is used for each of these two tasks. (C) The t-distributed stochastic neighbor embedding plot embedding layer in the EquiformerV2 model of potential materials space and prediction materials space. (D) Two criteria for η_OER and E_doping were set to screen the catalysts. The density of the scatter points is represented on the scale on the left, with potential catalysts marked by brownish-yellow dots.

The fine-tuning EquiformerV2 model outperforms three baseline models (EquiformerV2, GCN, and XGBoost), demonstrating superior performance even under small dataset conditions. This highlights the effectiveness of fine-tuning pretrained models in data-scarce scenarios. Further details on the training processes and a performance comparison against other fine-tuning methods (such as LoRA) and adapter locations are provided in note S3 and fig. S1.

Accurate prediction of catalyst active site

In HE CoOOH catalysts, the vast number of potential active sites, combined with the variation in TM doping, substantially increases the difficulty of screening for optimal active sites. To address this challenge, we used a transfer learning model to systematically identify the theoretically optimal active sites on the (001) plane of HE CoOOH (detailed in note S4). Given the extensive compositional design space, a complete prediction across all structures is computationally infeasible, and thus, we selected representative points that cover the entire compositional space for prediction. We used the Latin Hypercube Sampling method to select 17,500 data points (prediction space) in a total space of 3,057,600 data points. Next, we use t-distributed stochastic neighbor embedding to visualize the embedding vector of the prediction space and total space in 2D space (26). Figure 2C shows that prediction space is clustered in the middle of total space and separated from other data points, indicating that the prediction space is representative.

Following this, the structures of different adsorbents (^*O, ^*OH, and ^*OOH) were constructed on the slab, and a total of 1,543,543 structures were obtained (in quaternary system). With the aim of better selecting the catalysts, we set two criteria for obtaining potential catalysts: (i) η_OER below 0.35 V and (ii) E_doping less than 0 eV (Fig. 2D). The threshold for η_OER, derived from investigations that do not incorporate nanostructural modifications or defect engineering, suggests that lower overpotentials are associated with more favorable adsorption energies for key intermediates (^*OH, ^*O, and ^*OOH) in the four-electron OER mechanism, thereby reflecting superior intrinsic activity (43, 44). Prior studies indicate that catalysts with theoretical overpotentials below approximately 0.35 V tend to exhibit enhanced catalytic activity (27, 45, 46). Similarly, the selection of a criterion E_doping < 0 eV is underpinned by thermodynamic principles; a negative doping formation energy implies that dopant incorporation into the host lattice is exothermically favorable (47), ensuring greater structural stability (48). Together, these criteria provide a rigorous theoretical framework for screening catalysts that combine high catalytic activity with structural stability.

On the basis of these standards, we subsequently prioritized the top eight candidate catalysts for further investigation, specifically selecting those with the lowest predicted OER overpotentials that also satisfied the negative doping formation energy criterion. Among these, TiFeNiZn-CoOOH emerged as the most promising candidate, exhibiting a theoretical OER overpotential of 287 mV. Detailed compositions and corresponding overpotential values for these systems are provided in table S1. To validate the model’s ability to predict active sites, we constructed and analyzed 16 potential sites on the TiFeNiZn-CoOOH (001) surface, as illustrated in Fig. 3 (A and B). For each site, the relaxed energies of intermediates (^*OH, ^*O, and ^*OOH) were computed to determine the corresponding overpotentials. These theoretical calculations identified site O[CoNiZn] (labeled as #11) as the active site, which is consistent with the model’s predictions. In addition, fig. S2 provides the free energy diagrams for the OER of TiFeNiZn-CoOOH and the other candidate catalysts, demonstrating that the active site predictions align with theoretical calculations. Specifically, the model suggests that the active sites in TiFeNiZn-CoOOH involve Zn, Ni, and Co within their TM environment, with the coordination configurations of the other candidate catalysts and the comparison between DFT calculations and model predictions detailed in note S5. These findings confirm the model’s accuracy in predicting active sites and offer valuable insights into the catalytic mechanisms of complex systems.

Fig. 3. Active site search and validation.

(A) Diagram of the 16 potential active sites of the slab of TiFeNiZn-CoOOH material. (B) Comparison of site overpotential calculated by DFT and predicted by the model, where labeled ^#11 is the active site. (C) Theoretical overpotential η_OER as function of free energies ΔG_O − ΔG_OH and ΔG_OH for various potential active site. (D) Optimized geometries of the ^*O, ^*OH, and ^*OOH intermediates on the active site (number 11 site) of TiFeNiZn-CoOOH.

We investigate the energetics of all intermediates (^*OH, ^*O, and ^*OOH adsorption) and evaluate the theoretical overpotentials for η_OER of TiFeNiZn-CoOOH (001). DFT calculations show that coordination configuration [CoNiZn] site exhibits the lower η_OER (Fig. 3, B and C). The large deviation at site #8 is attributed to structural instability (as detailed in fig. S3A), the unfavorable local environment at this site causes the OOH intermediate to fail via OH detachment during DFT relaxation, while the model’s correct identification of the active site (#11) as the minimum confirms its robustness for catalyst screening. At the OH-adsorbed sites with coordination configurations [CoCoCo] and [CoNiZn], the latter configuration ([CoNiZn]) exhibits a higher number of active electrons in the O (2p) orbitals (fig. S3, B and C). This configuration enhances the adsorption of OH, lowering the energy barrier of the rate-determining step (ΔG_OH from 2.7 to 1.6 eV), thereby improving the overall catalytic performance. The improvement in η_OER at diverse sites at coordination configuration [CoNiZn] site is attributed to optimal ΔG_OH, which weakens ^*OH binding at active site affects the OER reaction moderately (Fig. 3, C and D). More details of the results can be found in fig. S4. Using this approach, we can precisely trace the active sites, and their locations align with those identified through theoretical calculations.

Automated laboratory validation

Thus far, we have demonstrated the model ability to predict theoretical overpotentials and identify active sites within the HE system. However, to fully validate its predictive capabilities, experimental verification is crucial. Therefore, we used fully automated experimental techniques through the Auto-Lab system, which enables precise and high-throughput material synthesis and electrochemical characterization. This approach highlights the potential of integrating fine-tuning models with state-of-the-art experimental technologies for accelerating materials discovery and evaluation. On the basis of the model’s predictions, we selected the top eight candidate catalysts that exhibited the most promising theoretical catalytic performance. These materials were synthesized and systematically characterized to assess their practical applicability. As illustrated in Fig. 1D, the Auto-Lab workflow is composed of multiple modules, including automated sample preparation, deposition, and electrochemical performance testing, ensuring consistency and reproducibility. The detailed step-by-step experimental procedures, including synthesis conditions, testing parameters, and data acquisition methods, are provided in the movie S1 for reference.

To rigorously assess the electrocatalytic performance of the selected materials, linear sweep voltammetry (LSV) measurements were performed in 1.0 M KOH electrolyte at a scan rate of 5 mV s⁻¹. The LSV curves for the top eight candidate materials provided in Fig. 4 (A and B) show that TiFeNiZn-CoOOH only requires a small overpotential of 263 mV to reach the benchmark current density of 100 mA cm⁻², notably outperforming other candidates including CrFeCuZn-CoOOH (295 mV), MnFeNiZn-CoOOH (298 mV), VMnFeNi-CoOOH (323 mV), TiVFeNi-CoOOH (313 mV), VMnFeZn-CoOOH (328 mV), TiVCrFe-CoOOH (340 mV), TiCrMnFe-CoOOH (311 mV), and CoOOH (356 mV). Notably, the overpotential is down by 25.8%, decreasing from 418 mV for CoOOH to 310 mV for TiFeNiZn-CoOOH at 500 mA cm⁻². At a large current density of 1000 mA cm⁻², it can even decrease by 25.7% from 455 to 338 mV. To establish a direct comparison with theoretical predictions, comparative analysis across multiple current densities (100, 500, and 1000 mA cm⁻²) consistently identifies TiFeNiZn-CoOOH as the most active electrocatalyst among the investigated catalysts. The results demonstrate a qualitative agreement between the model predictions and experimental measurements, confirming the model’s capability in screening high-performance catalysts.

Fig. 4. Auto-Lab experimental verification and prediction of electrochemical properties of materials.

(A) LSV curves of the predicted eight material systems. (B) Comparison of experimental and theoretical overpotential properties of the predicted eight material systems at 100, 500, and 1000 mA cm⁻² in 1 M KOH electrolyte. (C) Tafel slope for TiFeNiZn-CoOOH, CrFeCuZn-CoOOH, and MnFeNiZn-CoOOH. (D) Galvanostatic stability of the OER at the geometric current density of 100 mA cm⁻² for TiFeNiZn-CoOOH. (E) LSV curves of TiFeNiZn-CoOOH, TiFeZn-CoOOH, TifeNi-CoOOH, and TiFe-CoOOH. (F) STEM images of TiFeNiZn-CoOOH with elemental mapping for Co, Ti, Fe, Ni, and Zn. (G) High-resolution TEM images of TiFeNiZn-CoOOH. The inset, a magnified view of the selected area (indicated by the red box), clearly shows ordered lattice fringes. The measured interplanar spacing is 0.28 nm, which is in good agreement with the d-spacing of the (101) crystallographic plane of the CoOOH phase. h, hours.

Lower Tafel slopes indicate faster electron transfer during the OER process, which aligns with the enhanced catalytic activity of the materials. The Tafel slopes for TiFeNiZn-CoOOH, CrFeCuZn-CoOOH, and MnFeNiZn-CoOOH were determined to be 39.2, 51.6, and 55.3 mV dec⁻¹ in Fig. 4C, respectively. The lower Tafel slope of TiFeNiZn-CoOOH demonstrates its superior reaction kinetics compared to its counterparts (42, 48). In addition to activity and kinetics, the long-term stability of the catalysts is a crucial factor for practical applications. Galvanostatic chronopotentiometric stability testing of TiFeNiZn-CoOOH under a constant current density of 100 mA cm⁻² (Fig. 4D) demonstrates sustained OER activity over 120 hours with a potential retention rate of 97.5%, indicating exceptional operational durability. To further investigate the mechanistic origin of this stability and to determine the contribution of specific elements, we performed additional comparative experiments. We conducted both comparative long-term galvanostatic stability tests (fig. S5) and 5000-cycle accelerated durability tests (ADTs) on the primary quaternary TiFeNiZn-CoOOH catalyst against its key ternary subsystems: TiFeNi-CoOOH, TiNiZn-CoOOH, TiFeZn-CoOOH, and FeNiZn-CoOOH. The FeNiZn-CoOOH sample (lacking Ti) exhibited the poorest stability, showing the largest fluctuations in the long-term test and a notable overpotential increase of 10 mV at 100 mA cm⁻² after the 5000-cycle ADT. In contrast, all Ti-containing ternary samples (TiFeNi-, TiNiZn-, and TiFeZn-CoOOH) showed stability, with overpotential increases of only 4 to 5 mV (fig. S6). This provides experimental evidence that Ti is the key elemental contributor to the enhanced structural durability of the catalyst. Furthermore, prolonged galvanostatic stability testing of CrFeCuZn-CoOOH under identical operational conditions (100 mA cm⁻²) revealed a remarkable overpotential retention rate of 96.3% after 120 hours of continuous OER operation, as visualized in fig. S7. This result highlights the material’s exceptional electrochemical durability, comparable to the stability metrics observed for TiFeNiZn-CoOOH.

In the active site search, we specifically focused on the TiFeNiZn composition and its active sites within the [CoNiZn] coordination configuration. To experimentally verify this hypothesis, we additionally synthesized three materials: TiFe-CoOOH, TiFeZn-CoOOH, and TiFeNi-CoOOH. The corresponding LSV results (Fig. 4E) revealed that TiFeNiZn-CoOOH outperformed the other three materials, confirming that the optimal catalytic performance is achieved only when the [CoNiZn] coordination configuration is present. This finding aligns with our earlier identification of the [CoNiZn] structure as the active site, reinforcing its crucial role in enhancing catalytic activity. In addition, the inclusion of Zn in the material notably improved its catalytic performance, further validating the critical role of Zn in enhancing activity.

The morphology, elemental distribution, and crystalline nature of the TiFeNiZn-CoOOH were further elucidated using scanning transmission electron microscopy (STEM) and high-resolution TEM, as depicted in Fig. 4 (F and G). The STEM image reveals that the TiFeNiZn-CoOOH consists of aggregated hexagonal nanosheets. The corresponding energy-dispersive x-ray mapping indicates a uniform distribution of all constituent elements (Ti, Fe, Ni, Zn, and Co), each exhibiting high-intensity signals throughout the nanostructure. Furthermore, elemental analysis from different regions of the sample (detailed in figs. S8 and S9) confirmed similar atomic ratios of (Ti:Fe:Ni:Zn): Co, at approximately (1.16:1.04:1.08:1.55):10 and (1.54:1.12:1.30:1.41):10, respectively. The result is in excellent agreement with the nominal stoichiometric design of the material. Figure 4G displays well-defined lattice fringes with an interplanar spacing of 0.27 nm, which corresponds to the (101) crystallographic plane of the CoOOH phase, confirming the high crystallinity of the composite.

The uniform elemental distribution, coupled with the well-preserved lattice structure, suggests that the synthesis method effectively prevents phase segregation. This homogeneous incorporation of multiple TMs is crucial for enhancing structural stability and is anticipated to optimize the local electronic structure. By tuning the adsorption energies of key OER intermediates, this tailored composition is expected to considerably boost electrocatalytic activity. Similar morphological and compositional uniformity, which is critical for high catalytic performance, was also observed in other top-performing catalysts from our screening, such as CrFeCuZn-CoOOH (see STEM images in fig. S10).

Model-driven insight into catalysis

The complexity of active sites in HE systems is largely governed by the notable influence of local TM coordination configurations, which play a crucial role in determining catalytic properties. Since attention scores highlight key features in predicting target properties, the fine-tuning model’s adapter layers assigned high scores to the most notable features (26), providing deeper insights into the link between HE CoOOH structure and catalytic behavior. To further explore this correlation, we conducted a feature importance analysis using the fine-tuning model predicting η_OER, focusing on TiFeNiZn-CoOOH, a representative high-activity OER catalyst in the HE CoOOH family. As shown in Fig. 5A, the slab structure of TiFeNiZn-CoOOH includes doped TMO₆ octahedra, which represent different local active units. The colors of the TMO₆ octahedra correspond to the attention scores derived from the first layer of the Post-Att Adapter. Notably, the ZnO₆ octahedra active unit exhibits the highest attention value when predicting η_OER, reinforcing the strong dependence of active sites on local structural features, as further detailed in note S6 and figs. S11 and S12. Previous analyses identified the coordination configuration of the active site in TiFeNiZn-CoOOH as [CoNiZn], demonstrating the distinctive role of Zn in this system. This finding further emphasizes Zn’s critical function in modulating catalytic reactions, highlighting its contribution to enhancing the material’s overall catalytic performance.

Fig. 5. Interpretability approaches to research insights gained in predictive materials.

(A) Normalized attention values of the doped TM structural units in TiFeNiZn-CoOOH. The color represents the normalized attention unit scores. (B) Probability of the presence of doped TMs around the active sites. (C) Probability of the local TM environment combinations in Zn-containing systems and the corresponding overpotential box plot for these combinations. The blue boxplots correspond to the left y axis (η_OER), and the light green bars correspond to the right y axis (frequency ratio).

Beyond the case study of TiFeNiZn-CoOOH, we conducted an in-depth analysis of active site TM coordination configuration. Statistical methods were used to calculate the probability of each doping atom appearing in active sites across quaternary predicted datasets (details in note S7). As shown in Fig. 5B, Zn exhibits the highest occurrence probability at 72%, in the quaternary systems. To test the generalizability of this finding, we extended this analysis to binary- and ternary-doped systems. This trend was consistently observed, with Zn demonstrating the highest mean occurrence probability at active sites (for sites with η_OER < 0.6 V) in both binary (93.3%, fig. S13A) and ternary (93.0%, fig. S13B) systems. This highlights its important and general role in regulating active sites across different levels of elemental complexity.

Beyond analyzing TM coordination configurations around active sites, we investigated systems with overpotentials below 0.6 V, which correspond to a lower η_OER for CoOOH. In these systems, the types of coordination configurations were identified, and their corresponding overpotentials were statistically evaluated. The most prominent configurations for the quaternary systems are shown in Fig. 5C. Among them, the [CoCoZn] configuration exhibited the highest occurrence probability of 63.1%, with a median η_OER of 524 mV. In contrast, the [CoNiZn] configuration, with a lower occurrence probability of 7.2%, showed the lowest median η_OER of 499 mV. This key finding was validated in the simpler systems as well. For binary systems (fig. S14A), the [CoCoZn] configuration was also the most probable (81.6%) with a median η_OER of 520 mV, while the [CoNiZn] configuration (2.7% probability) again showed the lowest median η_OER (446 mV). Similarly, for ternary systems (fig. S15A), [CoCoZn] dominated (68.0% probability, 527 mV median) and [CoNiZn] (7.3% probability) yielded the lowest median overpotential (479 mV). The [CoNiZn] configuration shows the lowest overpotential among the others, suggesting that this coordination configuration corresponds to the most efficient catalytic performance in terms of η_OER. The robustness of this trend was analyzed by examining the distribution probability without the η_OER restriction, as shown in fig. S16 (for quaternary systems). In this case, the [CoCoZn] configuration was still found to have the highest occurrence probability (61.6%) with a median η_OER of 553 mV, while the [CoNiZn] configuration had a 5.9% occurrence probability and the lowest median η_OER of 516 mV. This robustness holds true for the other systems, where [CoCoZn] remained the most frequent configuration and [CoNiZn] retained the lowest median overpotential in both the binary (fig. S14B) and ternary (fig. S15B) full datasets. Together, this comprehensive analysis across binary, ternary, and quaternary systems provides strong, generalizable evidence that while [CoCoZn] is the most common Zn-based site, the [CoNiZn] coordination environment consistently represents the most efficient active site configuration in terms of η_OER.

The [CoCoZn] configuration, as a representative TM coordination, provides key evidence for understanding the critical role of Zn in HE CoOOH. As shown in fig. S17, we further analyzed the projected density of states (pDOS) for CoO₂ and Zn_0.33Co_0.67O₂. In the O (2p) orbital region, Zn incorporation into the CoO₂ matrix induces a large shift toward the Fermi level. Initially, the O (2p) orbitals were positioned at lower energy states. However, upon Zn doping, they shift closer to the Fermi level. This redistribution of electron density creates additional unoccupied orbitals near the Fermi level, enhancing the availability of electronic states for the OER. Such changes in the electronic structure, particularly the positioning of O (2p) states relative to the Fermi level, are in agreement with previous studies (10), which highlight the role of dopants in modifying the electronic environment to promote catalytic activity. By simplifying the multimetallic doping system into more manageable models represented by typical structures, the unique role of specific TMs like Zn in material performance can be more clearly understood. This approach not only helps explore the mechanisms behind metal contributions but also provides a fresh perspective for optimizing the catalytic performance of HE CoOOH.

DISCUSSION

In this work, we introduced a dataset-driven catalyst discovery strategy, integrating advanced machine learning models and automated characterization techniques to efficiently identify active sites and explore catalytic mechanisms. A dataset of 4822 HE CoOOH catalysts was used to develop an EquiformerV2-based model that accurately predicts E_doping and η_OER, achieving an MAE of 3.6 meV atom⁻¹ for doping energy and 4.5 mV atom⁻¹ for overpotential. A systematic, large-scale statistical analysis across a combined prediction space of more than 5 million structures (spanning binary, ternary, and quaternary systems) yielded two generalizable conclusions: Zn consistently exhibits the highest active site occupation probability (mean values: 93% in binary/ternary, 72% in quaternary), and the [CoNiZn] coordination configuration consistently yields the lowest median overpotential. These insights were further validated by automating the synthesis and characterization of several high-performance catalysts, including TiFeNiZn-CoOOH, CrFeCuZn-CoOOH, and MnFeNiZn-CoOOH, which demonstrated excellent OER performance (263, 295, and 298 mV at 100 mA cm⁻²). The identification of the [CoNiZn] coordination environment as a key active site configuration was experimentally confirmed, with Zn doping playing a pivotal role in enhancing catalytic activity. This dataset-driven approach demonstrates its potential in quickly uncovering active sites and exploring catalytic mechanisms, providing a strong framework for accelerating the design and development of high-performance catalysts in complex HE systems.

MATERIALS AND METHODS

Materials chemicals

MnCl₂·4H₂O, FeCl₂·4H₂O, NiCl₂·6H₂O, ZnCl₂·6H₂O, CuCl₂·2H₂O, VCl₃, CrCl₂, CoCl₂·6H₂O, TiCl₄ aqueous, urea, NH₄F, and KOH were purchased from Aladdin Chemical Reagent Company (Shanghai, China). Ethanol (>99.5%) was purchased from Sigma-Aldrich. All aqueous solutions were prepared using ultrapure water (resistivity = 18.2 MΩ cm⁻¹, Milli-Q water) throughout the whole experiments.

Automated laboratory synthesis process

All samples were synthesized using the hydrothermal method on nickel foam. The automated experiment of TiFeNiZn-CoOOH is presented as an example below (Fig. 1D). Before starting a new experimental activity, consumables (such as crucibles, ball milling jars, and test tubes) and precursor vials containing 50 to 100 g of powder are manually loaded into the reactor and crucibles storage area (Fig. 1D-1) and the powder proportioning instrument (Fig. 1D-2), along with the required TiCl₄ solution and deionized water into the liquid transfer device (Fig. 1D-3). Using the AGV car (Fig. 1D-0), a ball milling jar is transferred into the high-throughput powder system, and 1 mmol CoCl₂·6H₂O, 0.1 mmol FeCl₂·4H₂O, 0.1 mmol NiCl₂·6H₂O, 0.1 mmol ZnCl₂·6H₂O, 2 mmol urea, and 15 mmol NH₄F precursor powders are dispensed into a ball milling jar. Thereafter, the AGV car transfers the jar to the liquid transfer device, which adds 35 ml of deionized water and 0.1 mmol of TiCl₄ solution. The solid-liquid mixture is homogenized using the mixing instrument (Fig. 1D-4). The ball milling bar is opened using the automatic cap opening machine (Fig. 1D-5), and the resulting uniform slurry is transferred into a 50-ml polytetrafluoroethylene reactor using the liquid handling and dispensing system. The reactor is then sealed using the automatic cap opening machine. The sealed reactors are placed into the drying oven (Fig. 1D-6), where each batch is heated to 120°C at a slow heating rate of 5°C min⁻¹ and held for 5 hours. After the holding period, the reactors are allowed to cool naturally to 50°C. At this point, the robotic arm on AGV car removes the samples from the oven, and they are left to cool to room temperature for an additional 10 min. After the nickel foam has dried, it is placed on a cutting device and trimmed into sizes of 1 cm by 3 cm and 1 cm by 1 cm. A robotic arm then transfers the 1 cm–by–3 cm and 1 cm–by–1 cm samples to the electrochemical workstation (Fig. 1D-7) for electrochemical testing and analysis.

Materials characterizations

The scanning electron microscope (FEI Quanta 250 FEG) and TEM (JEOL JEM-2200F) were used to characterize the morphologies and structures of the as-synthesized samples. The x-ray diffraction (XRD) was performed by a LabX XRD 6000 with Cu K_α radiation (λ = 1.5416 Å).

Electrocatalytic activity for OER

Electrochemical measurements were carried out using a Gamry electrochemical workstation (Gamry Reference 1000, Gamry Instruments, USA). The OER performance of electrocatalysts was tested by the as-prepared TMs-CoOOH/NF samples (1 cm by 1 cm), Hg/HgO (1 M KOH) electrode and graphite rod as working electrode, reference electrode, and counter electrode, respectively. For OER electrochemical tests, all tests were performed in 1 M KOH solution. LSV curves were performed with a scan rate of 5 mV s⁻¹ at the range from 0.2 to 1.0 V. All the potentials were calibrated to reversible hydrogen electrode (RHE) by the equation

$$E_{\text{RHE}}=E_{\text{Hg}/\text{HgO}}+0.098+0.0591\times \text{pH}\left(\mathrm{V}\right)$$ (3)

The electrochemical active surface area was estimated from the electrochemical double-layer capacitance (C_dl). The C_dl was determined with typical cyclic voltammetry (CV) tests at various scan rates (20, 30, 40, 50, and 60 mV s⁻¹) in the range of non-faradic region. The C_dl was estimated by plotting the half of the difference between anodic current density and cathodic current density (ja-jc) at 1.17 V (versus RHE) against the scan rate.

Computational methods

In this study, DFT calculations were performed using the Vienna Ab initio Simulation Package. The exchange-correlation effects were described using the generalized gradient approximation (GGA) with the Perdew-Burke-Ernzerhof functional. To address the strong correlation effects in the d and f orbitals of TMs, the GGA + U method was used, incorporating an on-site Coulomb interaction parameter, Hubbard U, with specific U values provided in table S2. Projector-augmented wave pseudopotentials were used to represent the core electrons, and the plane-wave cutoff energy was set to 500 eV. The k-point sampling was performed using a 3 by 3 by 1 Monkhorst-Pack grid for the geometry optimization and electronic structure calculations of all intermediates on the CoOOH (001) surface. The energy convergence criterion was set at 10⁻⁴ eV. Grimme’s DFT-D2 empirical dispersion correction was applied to account for long-range van der Waals interactions. A vacuum slab of 22.5 Å was introduced in the z direction to ensure surface isolation and prevent interactions between neighboring surfaces. In addition, zero-point energy and vibrational entropy corrections were included, consistent with previous theoretical studies. The calculations for η_OER and E_doping were conducted on the basis of the aforementioned parameters, with detailed methodologies provided in the notes S2 and S3.

Supplementary Materials

The PDF file includes:

Supplementary Notes S1 to S7

Figs. S1 to S17

Tables S1 to S9

Legend for movie S1

References

Other Supplementary Material for this manuscript includes the following:

Movie S1