The electrochemical reduction of carbon dioxide (CO2RR) offers a promising pathway for converting CO2 into valuable chemicals and fuels, thereby mitigating climate change and addressing energy challenges. However, the development of efficient and selective electrocatalysts for CO2RR remains a significant hurdle. Traditional catalyst discovery methods, relying heavily on experimental trial-and-error approaches, are time-consuming and resource-intensive. High-throughput experimentation (HTE) has shown promise in accelerating the process, but it still faces limitations due to the vast chemical space of potential catalysts. The inverse design approach, using generative models to design materials with target properties, and data-driven high-throughput virtual screening (HTVS), offer efficient alternatives for exploring this immense chemical space. HTVS initially creates a materials pool and then predicts properties. While different from inverse design, which generates materials with specific properties, HTVS achieves the same objective of identifying desirable materials within an unknown chemical space without time-consuming steps, provided the materials pool is sufficiently large. This study focuses on developing and utilizing an HTVS strategy to rapidly identify active and selective catalysts for CO2RR, aiming to circumvent the limitations of traditional approaches.

Previous research has explored various computational methods for predicting the activity and selectivity of CO2RR catalysts. Density Functional Theory (DFT) calculations have been extensively used, but they are computationally expensive and thus limit the exploration of the chemical space. Machine learning (ML) models, trained on DFT data, have emerged as a powerful tool for accelerating the prediction process. Various ML models, including those based on crystal graphs such as LS-CGCNN, have demonstrated high accuracy in predicting binding energies. However, these often require precise geometric information and extensive surface structure modeling, hindering their application in HTVS for large-scale chemical space exploration. The concept of CO2RR selectivity maps, based on the binding energies of various intermediates, allows for the prediction of product selectivity. These maps usually use scaling relations to estimate the binding energies, reducing computational cost, but at the cost of reduced accuracy. This study utilizes the previous work of Tang et al. for the construction of a CO2RR selectivity map, but improves upon it by directly utilizing predicted binding energies rather than scaling relations to minimize uncertainties, creating a more refined 3D selectivity map.

This research developed a high-throughput virtual screening (HTVS) workflow that combines machine learning (ML) models based on the structure-free active motif-based representation (DSTAR) and a CO2RR selectivity map to predict the activity and selectivity of CO2RR catalysts. The DSTAR method represents catalyst surfaces using elemental descriptors of nearest neighbors, which avoids time-consuming steps like slab structure generation and binding site identification. This allows for the exploration of a vast chemical space. The authors utilized DFT and DSTAR to generate 2,463,030 active motifs for 465 binary combinations of 30 elements, significantly expanding the chemical space compared to existing datasets (GASpy dataset with 1,089 bulk structures). Trained ML models were used to predict the binding energies (ΔE_CO, ΔE_OH, ΔE_H) for all generated active motifs, showing reasonable accuracy, with mean absolute errors (MAEs) of 0.118 eV, 0.227 eV, and 0.107 eV, respectively. The CO2RR selectivity map, an improvement over previous work by Tang et al., utilizes these three predicted binding energies to predict the selectivity of four major CO2RR products: formate, CO, C1+ (products beyond CO/CO*), and H2. This 3D map uses six thermodynamic boundary conditions (BC1-BC6) derived from seven reaction steps, which consider various reaction pathways and potential surface poisoning. The potential-dependent activity and selectivity were evaluated at different applied potentials. A new metric, “productivity,” was introduced to quantify both activity and selectivity. Productivity integrates the probability of the active motif being within the range of predicted binding energy errors, offering a more comprehensive evaluation. The top 20 candidates for each product were identified based on productivity. The methodology included experimental validation of predicted results for Cu-Pd and Cu-Ga alloys, employing standard electrochemical techniques.

The HTVS workflow successfully identified numerous potential CO2RR catalysts. The 3D selectivity map accurately predicted the activity and selectivity of various known catalysts, including those selective for formate (Pb), CO (Au, Ag), H2 (Rh, Ir, Pt), and further reduced products (Cu). The study revealed several novel catalysts with high productivity for specific products, especially within binary alloy systems, many not previously reported in literature. Analysis of Cu-Al alloys showed that C1+ productivity increases with decreasing Al content and coordination number, providing insights into structure-activity relationships. The Cu-Pd and Cu-Ga alloys experimentally validated high selectivity for C1+ and formate, respectively, matching predictions, demonstrating the effectiveness of the HTVS workflow. A productivity heatmap showcased the potential-dependent activity and selectivity of various materials, highlighting how selectivity can shift with applied potential. The experimental validation and theoretical predictions show a high level of agreement, although some discrepancies were noted and attributed to extrinsic factors such as kinetics and local field effects not captured by the thermodynamic model. While there is some level of uncertainty due to the error in prediction and thermodynamic limitations, the overall accuracy and successful experimental verification points to the value of this methodology.

The results demonstrate that the developed HTVS workflow represents a significant advancement in catalyst discovery for CO2RR. The combination of DSTAR representation, enabling high-throughput predictions, and the refined 3D selectivity map, considering potential-dependent behavior, provides a powerful tool for accelerating the search for efficient and selective catalysts. The introduction of the “productivity” metric offers a more comprehensive evaluation of catalyst performance, integrating activity and selectivity with uncertainty considerations. While some discrepancies exist between predicted and experimental selectivities for certain materials, these are primarily attributed to the limitations of the thermodynamic model and the absence of kinetic and local field effects. The experimental validation of Cu-Pd and Cu-Ga alloys strongly supports the accuracy and utility of the HTVS approach. The analysis of Cu-Al alloys highlights the potential for tailoring catalyst composition and structure to optimize selectivity. The findings suggest that this methodology can be applied to other electrocatalytic reactions, offering broader applications in materials discovery.

This study successfully demonstrates a high-throughput virtual screening (HTVS) workflow for accelerating the discovery of active and selective CO2RR catalysts. The combination of DSTAR-based ML models and a refined 3D selectivity map enables the efficient exploration of a vast chemical space. The introduction of a new productivity metric provides a more comprehensive evaluation of catalyst performance. Experimental validation supports the effectiveness of the workflow. Future work should focus on incorporating kinetic effects and local field effects into the model to improve prediction accuracy and further expand the scope of the HTVS strategy to other electrocatalytic reactions and material systems.

The study's limitations primarily stem from the inherent assumptions of the thermodynamic model used in the 3D selectivity map. Kinetic effects and local field effects are not explicitly considered, leading to potential discrepancies between predicted and experimental selectivities for certain materials. The accuracy of the ML predictions also contributes to uncertainty, although the reported MAEs are reasonably low. The focus on binary alloys, while offering a substantial expansion of the chemical space, limits the exploration of more complex compositions. Further refinement of the ML models and incorporation of kinetic and other relevant parameters could enhance the predictive power and applicability of the HTVS workflow.