These are the initial structures tested for model selection that were filtered out

For the structures selected for further analysis and the structures discussed in the manuscript, please refer to the main attachment on https://zenodo.org/records/15384894

The POSCAR files are the starting geometries, and the CONTCAR files are the most recent geometries from a relaxation run (but not necessarily converged; in many cases structures were discarded before reaching convergence if the geometry is off or doesn't make sense); for each structure, the OUTCAR files provided are from the latest run, meaning that there are multiple relaxation runs for each initial structure (i.e. multiple restarts), and only the latest OUTCAR was provided to save space 

