STOP: Input matrix is linearly dependent

[glwhart]

This error happens for the POSCAR below. The structure was generated by MLIP (during a relaxation?). Something is amiss because the input lattice vectors are not linearly dependent---mathematica says that the triple product (volume) of these (i.e., the determinant) is ~140 cubic angstroms.

1
          1.51776       1.51776      -1.51776
          7.58878       1.51776       7.58878
          1.51776      -4.55327      -1.51776
 2 2 6
cart
                0             0             0
          3.03551            -0             0
          4.55327      -1.51776       1.51776
          1.51776      -1.51776      -1.51776
          6.07102      -3.03551       3.03551
          7.58878       1.51776       4.55327
          3.03551      -3.03551             0
          4.55327       1.51776       1.51776
          6.07102            -0       3.03551
          7.58878      -1.51776       4.55327

[glwhart]

The symptomatic problem is that Wiley's Niggli cell finder is failing to find a match with one of 44 niggli cells and so the output cell is all zeros, having never been set to anything. The underlying problem is that the limited number of digits in the MLIP output format is breaking symmetry unintentionally.

[glwhart]

Considerable work has been done on the Niggli cell finder, so this likely has gone away. Also MTP outputs more digits now (when compiled with "lossless precision" or something like that).

But MTP still writes out new structure files that are slightly distorted and this still tends to cause finite precision errors here and there. In the long term, we need to use a symmetry preserving optimizer (as in ASE) and stop using MTP to relax structures.