My Community

Hi,

I accidentally found that for my LDA+U calculations with slow convergence, it is actually easier to reach convergence by breaking up the calculation into two steps,
1) Starting from scratch, limiting NELM to say 12 with default NELMDL= -5. 12 electronic steps are not enough to fully converge.

2) Read the CHGCAR and WAVECAR from step 1 with normal NELM (e.g. 20) and no other change.

This actually works much better than a one-step approach, also starting from scratch of course; the latter often fails to reach convergence at all even after many (>50) number of steps. All tests were done with ALGO=normal (works better than fast).

Sorry if this is a repost. I did not find anything related in the manual or in the forum. My question is, why would breaking up the calculation help? Something to do with the iterative nature of diagonalization algorithm? And is there any setting to change so that a single step can obtain similar results? I tried the suggestion about the mixing parameters, and found no obvious improvement.

Best regards,
Fei

P.S. in step 2, since a WAVECAR was used, there was no delay in mixing by default.

Hi,

The first few ionic steps in a LDA+U calculation can be hard to converge. It depends on the number and nature of the atoms, on d or f orbitals etc. There are also a lot of metastable states in LDA+U calculations so it can be tough for the algorithm to find the ground state at the beginning of the calculation.

If you start from scratch, wavefunctions are randomly determined by vasp, so are the occupation matrices of the correlated orbitals. If these random values are far from the ground state, then the LDA+U calculation does not converge very well.

If you use the chgcar or the wavecar files, you use previous wavefunctions instead of random wavefunctions, so you are supposed to be closer to the ground state than if you start from scratch. That's why it converges faster.

If you're using LDA+U, be sure that your system is not trapped in a metastable state.

Boris