Multi-Task Mixture Density Graph Neural Networks for Predicting Catalyst Performance

This is a dataset containing 3225 various Cu-based single-atom alloys (SAAs), with CO adsorbed on surfaces of the configurations. First-principles calculated adsorption energy for each CO-SAA complex is also included. The dataset is constructed with 43 different element species are selected as dopants, including 26 from the sub-group (d-block) and 17 from the main group IIIA-VIA (p-block). The selected doping species cover most of the potentially applicable Cu-based SAAs, excluding group 18 elements, toxic and radiative elements, as well as alkali metal, alkaline-earth metal and halogens due to their strong reactivity. By varying the surface index, the doping position, and the adsorption site, e.g., top, bridge and hollow site, of the CO molecule, 75 distinct structures are constructed for each doping species, composed of 6, 8, 8, 27, and 26 structures on Cu(100), Cu(110), Cu(111), Cu(210), and Cu(411) surfaces, respectively, with representative structures. In related works (1, 2) published with the dataset, C and O are usually excluded to avoid introducing complicated interactions among multiple C and O atoms. In that case, the number of structures included is 3075. We provide "CuSAACO-vasp_data.zip", "CuSAACO-CONTCAR.zip", "CuSAACO-csv.zip" and "combined.zip" files. The former three files contain VASP output files (OUTCARs and CONTCARs) and two .csv files saving the structure index ("No."), doping element species ("element") of each CO-SAA complex as well as their adsorption energy ("E") derived from first-principles calculations. With the provided files, Cu-based SAA datasets for various tasks (I2I, R2I, R2R, MT I2I, see this article) can be generated using "struct2dataset.py" script in this github project. As an example of generated datasets, the "combined.zip" provides datasets for the MT I2I task generated with 10 different seeds, which are the ones applied in this article. In addition, coordinates of all atoms of corresponding configurations included in the dataset are also saved in the two .csv files for unrelaxed and relaxed structures, respectively. If an dataset of Open Catalyst Project form is not compatible with your project, it is also possible to process only with the .csv data using custom code to generate datasets.