howto:troubleshooting
Differences
This shows you the differences between two versions of the page.
Both sides previous revisionPrevious revisionNext revision | Previous revisionNext revisionBoth sides next revision | ||
howto:troubleshooting [2020/08/31 06:45] – [Incorrect initial geometry] Xavier Gonze | howto:troubleshooting [2021/01/25 13:38] – Xavier Gonze | ||
---|---|---|---|
Line 6: | Line 6: | ||
Many other tips might be found on the [[https:// | Many other tips might be found on the [[https:// | ||
+ | |||
+ | |||
+ | ==== Configure problems ==== | ||
+ | |||
+ | config.status: | ||
+ | |||
+ | This problem might come from the unpacking of the tar.gz. It was reported that using the Krusader file manager does not unpack correctly the tar.gz, so some files are missing. A manual unpack (tar -xvf ...) should solve the problem. | ||
+ | |||
+ | |||
+ | configure script won't accept hostnames containing a hyphen | ||
+ | |||
+ | Also apparently an error due to incorrect unpacking of the tar.gz . | ||
Line 52: | Line 64: | ||
==== SCF cycle does not converge ==== | ==== SCF cycle does not converge ==== | ||
- | (to be filled) | + | The default ABINIT algorithm for SCF (see input variable iscf) is an excellent compromise between speed and robustness. |
+ | It is referred | ||
+ | N. D. Woods and M. C. Payne and P. J. Hasnip, J. Phys.: Condens. Matter 31, 453001 (2019), and appear on the Pareto frontier. | ||
+ | Some parameters in the Kerker preconditioning have also an effect on the balance between speed and robustness, and the ABINIT default values | ||
+ | are good for metallic systems. As thouroughly discussed in this reference, there is no fool-proof fast algorithm. | ||
+ | |||
+ | If your SCF cycle do not converge (see nres2 or vres2 column in the output file), the reasons can be: | ||
+ | (1) Insufficient underlying accuracy in the solution of the Schrödinger at fixed potential | ||
+ | (2) Transient non-linear behaviour of the SCF, due either to (2a) sudden change of occupation numbers (usually only for metals), or (2b) long-wavelength fluctuations of potential, bigger than the gap. | ||
+ | |||
+ | Start first to address (2), by some tuning which can come without significantly slowing ABINIT. Test different values of diemac. The default is very large, and suitable fo metals. Try running with a small value (perhaps 5) if your system is a bulk semiconductor or a molecule. If it is a doped solid, some intermediate value (perhaps 50) might be appropriate. | ||
+ | |||
+ | If this does not work, try to address (1), set tolrde to 0.001, increase the value of nline (e.g. 6 or 8 instead of the default 4), as well as of nnsclo (e.g. 2 instead of the default). Your residm should be significantly lower than without such modifications, | ||
+ | |||
+ | If this still does not work, but your residm did not look bad after all before addressing (1), then revert back to the default values of tolrde, nline an nnsclo, as this indeed slows down ABINIT. | ||
+ | |||
+ | Then, try to use iscf 2. This is a very slow but inconditionally convergent algorithm provided diemix is small enough. At some stage of convergence, | ||
==== Geometry optimization does not converge ==== | ==== Geometry optimization does not converge ==== |
howto/troubleshooting.txt · Last modified: 2021/07/12 15:28 by Jean-Michel Beuken