MACE-mp float, size and optimizer tests

[kavanase]

Hi!
Thanks for developing and maintaining such useful tools.
Just thought it might be useful to mention, in this preprint I tested a range of MACE-mp settings (model size, float64/32, ASE optimizer choice) etc for speed and accuracy on a large dataset of defect supercell relaxations, which may be of interest – e.g. float64 is recommended for geometry optimisations but these tests found that in most cases it gave essentially the same result as float32 (with a mean (total) absolute difference of 1.4 meV, standard deviation of absolute differences of 1.1 meV, for supercells of 80-800 atoms; much less than 0.1 meV/atom).

[ilyes319]

Very cool thanks for sharing