Routine ZPOTRF ZTRTRI failed

[basant_ali]

I am trying to run an OPTICS job with Algo=Exact and NBANDS=560.
The error disappears with Algo= (any other thing rather than Exact). However, Algo=Exact is mandatory to run the following GW calculation.
However, I receive an error
scaLAPACK: Routine ZPOTRF ZTRTRI failed! 52684 1 1
LAPACK: Routine ZTRTRI failed!

I think this is an error related to the compilation of scaLAPACK and LAPACK, I used intel-mkl/2020.4.304 to compile the attached make file for VASP5.4.4.

Please, advise!

Thank you,
Basant

[manuel_engel1]

Hi Basant,

I see that this is related to your other post. They might share a common issue. I don't believe that this is related to your installation since the library call functions correctly but the library returns with an error, indicating that there is a problem with the orthonormalization of the orbitals. Could you please upload the input and output files of the optics run? I would like to see if there is an underlying issue with the system you care calculating.

[basant_ali]

Thank you for your reply, Kindly find the INPUT and OUTPUT attached.

[manuel_engel1]

There are a few things you could try:

• Try calculating without symmetrizing (ISYM=0 or ISYM=-1).
• Try generating a WAVECAR and CHGCAR in a preparatory calculation using ALGO=Normal (or similar). Then restart from those preconverged files using ALGO=Exact.
• Disabling real-space projectors (LREAL=False).
• Symmetrize your POSCAR. Judging by the lattice vectors, it seems like the symmetry is slightly broken. Maybe this is on purpose, but otherwise it could greatly help converge the system.

I have a strong suspicion that the final output of this calculation is responsible for the NaN values in your subsequent GW calculation.

[basant_ali]

Thank you for your reply.

I have tried all the suggested steps, however, the error is still there.
The run freezes for 2 hours after starting to write the OSZICAR then stops with a (Routine ZPOTRF ZTRTRI failed) error without writing any electronic step.

Thank you,
Basant

[manuel_engel1]

Some final things you could try:

• Switch to PREC=Accurate to increase your FFT grid dimensions.
• Swap the POTCAR of some of the atoms for a different one. For example, you could use the GW POTCAR for the Br atom.
• Use a smaller cell. I have noticed while looking at the POSCAR that your structure could be approximately described by a smaller unit cell (there is a 2x2 replication of the unit cell with some small atomic distortions). This would help with convergence.

If everything fails, you may not have any other choice than to switch to a different ALGO. In this case, since you aim to perform GW calculations afterwards, you have to be extra careful to include enough unoccupied states and to converge them properly. Unoccupied states usually converge much slower than the occupied ones. Therefore, it will be necessary to increase the number of electronic minimization steps manually.

[basant_ali]

Thank you for your advice.

I have already tried PREC= Accurate and I can not decrease the unit cell since I have four different distorted tetrahedral angles.

Using a different Algo and conversion of the system let us go back to the error of the GW calculation of NaN dielectric values.

Thank you for your help and support.

[manuel_engel1]

I understand. Maybe another note: ALGO=Exact is not a strict necessity for subsequent GW calculations. What's important is that enough unoccupied states are well converged. ALGO=Exact is usually an easy way to accomplish this but with enough care, you can achieve the same result with other diagonalization methods. This is a rather large system for what you are trying to accomplish so proper convergence might be very difficult.