Discrepancies between "legacy" and "current" data. What should I believe?

Hi,

TL/DR
The optimized crystal structures from the legacy materials project website don’t always agree with the current website. For instance, compare:
mp-554573: SiO2 (Orthorhombic, Ama2, 40) vs. mp-554573: SiO2 (orthorhombic, Ama2, 40)
One cannot simply compare the cell parameters as the crystallographic settings are different, but it’s obvious that the optimized structural parameters are different when one compares the densities. That said, the reported calculated stiffness tensors are the same. What’s up with that? What data should I rely on? I see this happening for quite a few calculations that I’m interested in.

Longer explanation…
I’m currently a civil engineer in the SF Bay Area. I was the first user of VASP in the US back over 25 years ago. Around 20 years ago, I used VASP to parameterize a set of transferable interatomic potentials for mixed metal oxide crystals and glasses. Although never published (except as “personal communication”), these potentials have been used in hundreds of studies. Twenty years later, I’m trying to finish this up by parametrizing new interatomic potentials based upon my own DFT linear response calculations using ABINIT/PBE/ONCV and data from the Materials Project. I’m hoping to exploit as much data as possible from the Materials Project but am having lower and lower confidence about what I see. Any help/suggestions would be appreciated.

Thanks,
David

Hi David,

I can’t offer an official response any more, since it’s been a long time since I worked on the MP stack directly, but the calculation provenance for each materials page is provided for both the legacy and the current site.

Based on this - it looks like the lattice parameters for the materials details page are taken from the r2scan structure optimization, presumably because r2scan is a more accurate functional.

For what it’s worth, even as a person with much higher than average knowledge of the MP stack, I didn’t realize this would be the case and it took me a few minutes to find and confirm the cause of the discrepancy (all with public info). But the r2scan data being on MP has been the case for a while now, and it’s consistent with the current documentation, and I think it’s a reasonable decision to privilege the r2scan data for the gains in accuracy, even if it makes the provenance a little more complicated.

2 Likes

@David_Teter, some of the information you are looking for can be found in our documentation.

@Joseph_Montoya is totally right here. The structure for a material is defined with a “blessed” calculation (task). This can change when new data is added to the database. For example, now that we have many metaGGA calculations in the core database, the canonical structure is taken from an r2SCAN calculation if available.

In general, each different property can be sourced from different available calculations. This information can be found under the “Calculations” section in the details pages (https://materialsproject.org/materials/mp-554573/tasks), or by the task_id or origins information from each respective API endpoint.

– Jason

1 Like

Hi @Joseph_Montoya - Ah. That makes sense. The last time I spent any time on this was over 5 years ago prior to the widespread use of any new meta-GGA functionals. I will roll back to the legacy data so that it is consistent across my parameterization data sets. Thanks!

1 Like

@munrojm - Thanks Jason. Out of curiousity, does the Materials Project have any calculated elastic properties from both PBE and r2SCAN functionals? Thanks!

@David_Teter, no problem. Unfortunately, all of the current elastic properties have been calculated with only PBE/PBE+U. We do have a large update in the pipeline for the elastic data, but it consists of a clean up and expansion of the dataset with the same workflow.

– Jason

2 Likes

Understood. Thank you.