What is the expected uncertainty of the input DFT data? For example, if you ran the DFT simulations for the same structure using multiple starting points, what is a typical spread like (stochastic uncertainty)? Specifically, we’ve been looking at the S2EF 200k data.
Great question. Unfortunately we haven’t run a thorough enough set of experiments to provide a quantitative value here, hard to really know without it, but it’s usually small.
We had experimented early on by perturbing initial structures slightly to see if the relaxed states changed but didn’t notice anything alarming here. We had another experiment where we compared forces of images with different initialized WAVECARs, e.g. comparing the forces of images along a trajectory (WAVECAR initialized at the first image) to those of singlepoint calculations of those same images (WAVECAR initialized for each image) and observed a force MAE of ~0.00145eV/A on free atoms. Note these experiments were on a fairly small scale (few hundred images) and I would hesitate to claim this for the entire dataset, but just to give you an idea. Also, the DFT functional used in OC20 (RPBE) doesn’t provide intrinsic uncertainty estimates like say BEEF-vdW does so that information isn’t available.
I hope this helps, sorry I don’t have more details for you.
1 Like