ATK 11.2.3 fixes several minor problems in 11.2.x.
The update can be obtained from the usual download page.
Changes in ATK 11.2.3:
- There was one more small detail regarding the device DOS fix (for the NaN problem) that was not included in the 11.2.2 update. Now the problem should be finally solved.
- Some Hamiltonian matrix elements could become wrong near the boundaries; seems like a compiler bug, but it's fixed now and it cures convergence for systems where all charge would be lost near the edges.
- Better conversion of bulk systems to two probe systems. If you have problems with convergence, this could fix it! Please contact us for more information if you have convergence issues.
- The total energy entropy term was missing a factor 2 for non-spinpolarized systems.
- New feature: the "repeat" method on a bulk system now repeats the density matrix and Hamiltonian.
- VNL scripter: You can now select both periodic and Dirichlet boundary conditions for the FastFourier2DSolver.
- Calculation of forces now consumes less memory.
In addition, we have included a bonus: a performance improvement.
There is one term in the Hamiltonian, H0, which doesn't change between iterations. So, it's computed once and for all, before you get into the self-consistent loop. However, the term is present both for electrodes, the equivalent bulk and the device calculation itself. In ATK 11.2.3 we have improved the algorithm for this calculation, and reduced the time essentially from minutes to seconds.
Notably, this term also needs to be computed when you read a configuration from a NetCDF file, to perform analysis, for instance. So this will also be much faster now.
The picture below shows a performance graph of a device calculation. The encircled part is the H0 term for the equivalent bulk, and we can see the dramatic reduction in time (the 11.2.2 graph actually doesn't fit entirely on the screen). The speed-up is almost 100x in this example, which contains 360 atoms in the central region; the improvement actually gets better the larger the system is, as we have mapped an N2 operation to an N operation (N=number of atoms).
Changes in ATK 11.2.2:
- Some stress relaxations with fcc unit cells could fail unless you used a very high k-point sampling and mesh cut-off, esp. for GGA. The new algorithm applying the unit cell symmetries also to the stress tensor, this is less of an issue now.
- Device DOS could become NaN (not a number) if the unit cell had no symmetries. This is now fixed.
- VNL/Builder: further improvements in the algorithm for converting a bulk system to a device configuration: if no periodicity is found, the user can specify it manually
- Grouping is again available from the menu in the Builder
Changes in ATK 11.2.1:
- Checkpoint file in MPI: if the user specified his own name for the checkpoint file, the slave nodes also tried to write to the file when running in parallel, causing the program to fail.
- LDOS: fixed an issue with parallelization
- VNL/Builder: improvements in the algorithm for converting a bulk system to a device configuration
- Touch-ups to manuals
For other features in 11.2, see the 11.2 release letter.
NOTE: Please uninstall any existing installation of ATK 11.2.x before installing this update, in order to avoid confusion with paths, desktop icons, etc.