The full name of this dataset is “Computational Dataset for Catalytic Conversion of C1 Molecules Based on Iron-Based Catalysts,” with version number V1.0. The dataset centers on the 001, 111 and 510 crystal planes of Fe₅C₂ as the core research subjects, covering three fundamental reactions of C1 molecules on the catalyst surface: adsorption, dissociation, and formation (Table 1). It relies on density functional theory to obtain structural and energy parameters. The directory structure is hierarchically organized, with clear data architecture, and file naming reflects the reaction system, pathway, and type, facilitating user identification and retrieval.

From a theoretical perspective, this dataset systematically stores high-precision computational data (such as adsorption energies, dissociation barriers, and product distributions) for the complete reaction pathways of C1 molecules on multiple crystal facets (001, 111, 510) of Fe₅C₂, providing abundant case studies for investigating the synergistic adsorption mechanisms of C/H/O. The cross-facet and multi-reaction comparative analyses within the dataset can further reveal the intrinsic relationship between crystal facet effects (e.g., step density, atomic coordination number) and reaction selectivity, advancing C1 catalytic conversion theory from single-facet studies toward multidimensional mechanistic coupling.

Regarding catalyst design, the energy and structural parameters in the dataset can be directly used to establish reaction kinetic models and predict product selectivity under different reaction conditions. Research shows that the presence of surface OH groups can significantly alter the product distribution of iron-based catalysts and enhance olefin selectivity^[5]. This dataset provides a theoretical basis for the rational design of surface-modified Fe₅C₂ catalysts by detailing the influence of water molecules on the surface C/H ratio. For example, the results on how water molecules affect surface carbon coverage allows for the precise regulation of the active environment on the catalyst surface, thereby reducing the formation of undesired by-products like methane and increasing the production efficiency of higher-value products with longer carbon chains (C₂₊, such as ethylene and propylene).

At the methodological level, the dataset provides a high-quality computational benchmark for surface catalytic reactions, which can be used to validate the accuracy of computational methods or train machine learning force fields. The data includes nearly 700 optimized adsorption structures and reaction transition states, covering various chemical environments of elements such as C, O and H on the Fe₅C₂ surface, serving as reference data for surface catalytic reaction simulations. Additionally, the rich structure-energy correlations in the dataset offer training samples for developing rapid catalytic activity descriptors and machine learning models.

In terms of application prospects, the dataset covers multi-scenario reaction data across the entire cycle of C1 molecule conversion (adsorption → dissociation → formation), such as syngas-to-CH₄ and CO₂ catalytic conversion, directly serving the optimization of industrial catalytic processes. For example, by analyzing the HCOO formation pathway on the Fe₅C₂(510) crystal surface in a “2H₂O+2H environment”, the impact of water conditions on CO₂ hydrogenation conversion can be investigated. Additionally, the correlation between environmental parameters (e.g., H₂O partial pressure) and product distribution in the dataset can guide the dynamic adjustment of operating conditions in industrial reactor designs (e.g., fluidized bed, fixed bed), facilitating breakthroughs in efficient carbon resource conversion and low-carbon emission technologies.

All calculations in this dataset were performed based on density functional theory using the Vienna Ab initio Simulation Package (VASP) software^[6−7]. The projector augmented wave (PAW) pseudopotential^[8] was employed to describe the atomic nuclei and core electrons, while the Perdew-Burke-Ernzerhof (PBE) functional^[9] was used to treat electron exchange-correlation interactions. The plane-wave cutoff energy was set to 400 eV, and the k-point mesh was generated using the Gamma-centered method to ensure a sampling density in reciprocal space better than 0.04 Å⁻¹. During structural optimization, the convergence criteria were set to 0.03 eV/Å for atomic forces and 10⁻⁴ eV for energy. Transition state searches were conducted using the CINEB method^[10], with each transition state confirmed to possess only one imaginary frequency through vibrational frequency analysis.

All surface models were represented using periodic slab models. For instance, the Fe₅C₂(510) surface (Figure 1) features a slab model with a thickness of 10 atomic layers, where the bottom 5 layers are fixed, and a vacuum layer of 15 Å is set to eliminate interactions between periodic images. The surface calculations employed a supercell of 10.55 Å × 13.42 Å, corresponding to a stoichiometric ratio of 58 iron atoms and 24 carbon atoms. In the reaction pathway calculations, the bottom two atomic layers were fixed at their bulk positions, while the top two layers and adsorbed species were fully relaxed.

The data generation process consists of four main steps: (1) optimization of the clean surface; (2) adsorption structure search; (3) transition state calculation; and (4) vibrational frequency analysis. First, the Fe₅C₂(510) clean surface is fully relaxed to obtain a stable surface structure. Subsequently, by systematically varying adsorption sites (such as top, bridge, and hollow sites) and molecular orientations, the most stable adsorption configurations of reaction intermediates are searched. For each reaction step, the CINEB method is employed to connect the initial and final states, identifying the reaction transition state. Finally, vibrational frequency analysis is performed for each stable point (initial state, transition state, final state) to obtain thermodynamic corrections and vibrational properties.

The calculations in this dataset were performed using optimized INCAR parameter settings to ensure a balance between accuracy and efficiency. The plane-wave cutoff energy was set to 400 eV, slightly higher than the ENMAX values of most elements in the POTCAR file, ensuring the completeness of the plane-wave basis set. The electronic self-consistent convergence criterion EDIFF was set to 1×10⁻⁴ eV to guarantee the precision of energy calculations^[4,11]. Ionic relaxation employed the conjugate gradient algorithm (IBRION=2), with the force convergence criterion set to 0.03 eV/Å, ensuring the rationality of the optimized structure^[4,11]. The k-point grid was generated using the Monkhorst-Pack method and set to 2×2×1. This configuration was optimized for the specific characteristics of surface calculations: the k-point density in the x and y directions (surface plane) was sufficient to describe the periodic electronic structure of the surface, while in the z direction (perpendicular to the surface), only one k-point was necessary due to the presence of a 15 Å vacuum layer. The broadening method utilized the Methfessel-Paxton technique (ISMEAR=2) with a parameter SIGMA=0.20 eV, suitable for metallic systems such as Fe₅C₂. Spin-polarized calculations (ISPIN=2, MAGMOM = 3 for Fe) were enabled to accurately describe the magnetic properties of the iron-based system. Additionally, D3 dispersion correction (IVDW=12) was included to properly account for long-range van der Waals interactions, which is particularly important for calculating molecular adsorption energies.

This dataset primarily investigates the reaction processes of C1 molecular catalytic conversion on the Fe₅C₂ polycrystalline surfaces (001, 111, 510), with the corresponding structural counts for each surface being 96, 93 and 472, respectively. Among them, the Fe₅C₂(001) surface encompasses three types of reactions: adsorption (e.g., adsorption of species like COH and H₂ at ms/ma sites), dissociation (e.g., decomposition of HCO under H environment), and formation (e.g., generation of products like COH under H environment). The reaction logic on the Fe₅C₂(111) surface is similar. Due to its complex stepped structure, the Fe₅C₂(510) surface exhibits more diverse reactions: the adsorption phase involves mixed species such as CO₂ and H₂; dissociation covers CO/CO₂ conversion under multiple environments like 2H₂O+2H and H₂O+1H; and the formation phase can produce low-carbon hydrocarbons like CH_x and CH₄, as well as intermediates such as HCO and H₂COOH, comprehensively demonstrating the regulatory effects of H₂O coexistence and multi-H environments on reaction pathways. Each subdirectory contains important structural files from VASP calculations, where POSCAR represents the initial structural input, CONTCAR the optimized structure, and OUTCAR includes detailed energy and electronic structure information.

Typical data samples visually illustrate the data encoding rules and reaction information: “Fe001_ads_1H2ma” denotes the adsorption of one H₂ molecule at the ma site on Fe₅C₂(001); “Fe001-0.00-Fe56C20” represents the pure surface structure of Fe₅C₂(001), presented with 0.00 termination, containing 56 Fe atoms and 20 C atoms, etc. For chemical reactions involving transition states, we use a special symbol (~) to denote them. The field before the ~ symbol represents the reactant, and the field after the ~ symbol represents the reaction product. For example, Fe510_H-O-H-OH-H2O~H-OH-OH-H2O_00 indicates a chemical reaction occurring on the Fe₅C₂(510) surface, where the reactant H-O-H-OH-H2O forms the product H-OH-OH-H2O, with 00 representing the initial state structure and 11 representing the final state structure; Fe510_H-O-H-OH-H2O~H-OH-OH-H2O_01 represents the first structure on the transition state path, and so on. These samples, spanning the entire process from adsorption to dissociation and formation, validate the dataset’s systematic recording of multidimensional information encompassing “surface-environment-species-reaction”, as detailed in Table 2.

To ensure the accuracy and reliability of the computational data, this study rigorously validated the calculation methods. First, the convergence of computational parameters was verified by testing the plane-wave cutoff energy and k-point density. The test results showed that a cutoff energy of 400 eV and a k-point density of 0.04 Å⁻¹ could achieve total energy convergence within 1 meV/atom. Additionally, the accuracy of the computational method was confirmed by calculating the lattice constants of Fe₅C₂ bulk phase (a = 11.545 Å, b = 4.496 Å, c = 4.982 Å), which closely matched the experimental values (a = 11.588 Å, b = 4.579 Å, c = 5.069 Å). For transition state calculations, each suspected transition state underwent vibrational frequency analysis to ensure the presence of only one imaginary frequency (ranging from −200 to −50 cm⁻¹). For particularly complex reaction pathways, the dimer method was employed for cross-validation to guarantee the reliability of the transition state results.

All adsorption structures and reaction intermediates were optimized using stringent convergence criteria: energy changes less than 10⁻⁴ eV and atomic forces less than 0.03 eV/Å. To ensure the attainment of global minima rather than local minima, multiple initial adsorption configurations were tested and compared for each adsorption system. For instance, for the HCOO* species, various modes such as bidentate adsorption (O atoms interacting with two Fe atoms) and monodentate adsorption (one O atom interacting with Fe) were tested, with the most energetically stable configuration ultimately selected for inclusion in the dataset.

For aqueous systems, the initial arrangement of water molecules considered multiple orientations and positions to ensure the most stable adsorption structure was found. Additionally, frequency calculations were performed to exclude the presence of imaginary frequencies, confirming that all optimized structures reside at minima rather than saddle points on the potential energy surface.

In the domains of machine learning and data mining, the proliferation of high-dimensional data—characterized by a vast number of features or attributes—presents significant challenges that extend beyond mere storage and computational inefficiencies. The phenomenon known as the “curse of dimensionality” exacerbates these issues, as the sparsity and geometric properties of high-dimensional spaces complicate the intuitive understanding of data distribution and intrinsic structure, thereby hindering effective analysis and modeling. Therefore, as shown in Figure 2, we employed principal component analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE) methods to perform cluster analysis on all our generated structural and energy data. In principal component analysis (PCA), Component 1 and Component 2 represent orthogonal dimensions derived from reducing high-dimensional atomic environment descriptors (such as SOAP vectors), where Component 1 captures the maximum variance in the data and corresponds to the most significant structural patterns, while Component 2 accounts for the second-largest portion of residual variance. PCA highlights differences between systems through linear dimensionality reduction, while t-SNE emphasizes local correlations among configurations (such as the clustering of reaction intermediates in t-SNE space) via nonlinear mapping. The PCA similarity plot (Figure 2) visually demonstrates structural similarity through the spatial distribution of points—the closer the points, the more similar the atomic configurations, with different color codes clearly identifying clustering features among systems. The core advantage of the t-SNE visualization lies in preserving local similarity, where the color bar maps formation energy from blue (approximately −800 eV, low-energy stable configurations) to yellow (approximately −200 eV, high-energy metastable/transition states), directly correlating with structural stability: continuous blue regions correspond to smooth energy evolution during catalyst surface relaxation, while discrete yellow clusters suggest energy barriers or phase transition nodes along the reaction pathway.

This dataset adopts a hierarchical directory structure for organized access. Each computational subdirectory includes standard VASP files: POSCAR (initial structure), CONTCAR (optimized structure), and OUTCAR (primary output data). Users can select relevant files for analysis based on research objectives. The typical analysis workflow for this dataset includes: (1) Structural parameter extraction: reading bond lengths, bond angles, and other structural information from CONTCAR files; (2) Energy extraction: retrieving energy data from OUTCAR to calculate reaction energies and energy barriers; (3) Electronic structure analysis: analyzing electronic information in OUTCAR to understand the reaction mechanism; (4) Kinetic parameter calculation: combining vibrational frequency data to compute thermodynamic corrections and kinetic constants.

It is recommended to use the following tools for data reading and analysis: (1) Official VASP post-processing tools (such as vaspkit, p4vasp, etc.) can be used for extracting basic structural parameters and energies; (2) Python materials science libraries (such as pymatgen, ase, etc.) provide powerful VASP output parsing capabilities, supporting customized data analysis workflows; (3) Visualization software (such as VESTA, OVITO, etc.) is suitable for structural display and electron density analysis. For beginners, it is advisable to start with the CONTCAR files of typical reaction pathways, visualize the structures of key intermediates and transition states, and establish an intuitive understanding of structure-performance relationships. Advanced users can delve into detailed electronic structure information in the OUTCAR files, such as density of states, charge population, and work function variations.

It should be emphasized that this dataset has certain limitations: Firstly, it employs an idealized periodic slab model, primarily exposing ideal terminations without explicitly including the effects of various defects such as point defects and grain boundaries, which differs from actual industrial Fe-based catalysts (with particle sizes of 10−100 nm and surface defects interacting with supports). Secondly, the calculations are based on 0 K conditions and do not account for industrial reaction parameters such as temperature and pressure. In practical applications, the influence of temperature on dynamic surface structural evolution should be considered. Thirdly, this dataset focuses on the Fe₅C₂ phase and does not include data on other iron carbide phases. In real catalytic processes, phases such as Fe₂C, Fe₃C and Fe₇C₃ may also be involved in addition to Fe₅C₂. It is recommended that users treat this dataset as foundational reference data and conduct extended analyses tailored to specific research questions. For instance, zero-temperature energy barriers can be converted into apparent activation energies through microkinetic modeling, or structure-activity relationships can be established using machine learning methods to predict the catalytic performance of similar systems.