3.54 Deprecated keywords

Outsources the distribution of integration grid batches to the external qhull and METIS libraries. Only relevant if these libraries are compiled into the code. However, the associated grid_partitioning_methods are less useful than the default maxmin algorithm, and the internal work distribution method of FHI-aims usually performs rather well. Therefore, this option is deprecated and kept only for experimental purposes, for now.

In a parallel run of FHI-aims, each processor holds a certain part of the real-space integration grid, which in turn are each touched by all atom-centered multipole components (splined) of the real-space Hartree potential. So, to construct the electrostatic (Hartree) potential on each grid point, an array of splined atom-centered multipole components $\delta {\tilde{v}}_{\text{at},lm}(r)$ must be available on every MPI sub-process (see Sec. 3.7 for details regarding the electrostatic potential). The memory use to store these components grows rapidly and with a large prefactor with system size. Thus, keeping a local copy of all the splined multipole components of the Hartree potential on each CPU is not advisable.

The actual handling of these components is instead controlled by keyword communication_type. The following choices for option are possible:
calc_hartree : The default, and usually very efficient. Hartree potential components for each atom are integrated on the fly on each CPU when needed (usually less CPU time than communication time).
shmem : If compiled with shared memory support (see Section I.2), each compute node of a parallel run keeps the components of the Hartree potential in a separate shared memory segment, only internode communication is needed. Performance test show hardly any benefet over calc_hartree. Note that the (legacy) keyword distributed_spline_storage should be false for shmem, at least.

This keyword is superseded by the compensate_multipole_errors keyword which solves the problem of small residual charge integration errors in slab calculations much better than distribute_leftover_charge .

Keyword distribute_leftover_charge could introduce noticeable total energy inaccuracies especially for “light” integration grid settings.

See Ref. [36] regarding the correct shape of the energy functional that treats the nuclear and electronic parts of the electrostatic energy together on a per-atom basis.

If requested, keyword force_smooth_cutoff ensures that the radial function $u(r)$ and its first and second derivatives remain below tolerance at the outermost point of the logarithmic grid where any of them is non-zero at all. The code stops if the onset of the radial function is too abrupt.

It would be a good idea to switch this option on if reducing the width parameter of keyword cut_pot to a very low value (say, below 1 Å).

This keyword is retained for experimental purposes only, for the moment. Since the related grid_partitioning_method group was a proof-of-concept to show that the default maxmin performs better, this keyword is now deprecated. See Ref. [128] for more details if interested.

Several large work space arrays across the integration grid are used in the construction of the Hartree potential. Their size can grow quite large, especially when forces are computed for large structures (then, three arrays per atom are required for all atoms).

FHI-aims can circumvent this by computing the final output (integrated energies and forces) in “chunks” of the whole integration grid, limiting the work space used for each chunk. This modification is especially important on low-memory-per-processor architectures such as the BlueGene.

This keyword is no longer supported, do not use it. It will be superseded by an improved handling of scalar relativity during the s.c.f. cycle in the future.

On few-CPU systems and for mid-sized systems (several hundred atoms), the stored electron density components from past iterations are a large part of the memory used. If this becomes a bottleneck, the stored Pulay arrays can in principle be swapped to disk, instead, to be read only during Pulay mixing.

If anyone has a strong need for this currently unsupported feature, please contact us.

Meaningful only in the spin-polarized case (spin collinear in control.in). On a technical level, this is a special case of the more general, locally spin-constrained DFT formalism available within FHI-aims (see Sec. 3.14). Note that the underlying constraint_electrons keyword can be used to enforce a non-integer fixed spin moment, in addition to allowing to fix electron or spin numbers in given subsystems

Also, be sure to check what the Kohn-Sham eigenvalues mean if you need them, do not use them blindly. multiplicity shifts the eigenvalues!

For practical purposes, the integration grid should always be repartitioned after a relaxation step; the associated overhead is low, and the shape of the batches will remain optimal in the face of individual points that “move” with different atoms.

This flag currently only applies to the initialization iteration, in case that self-adapting angular grids are used (not recommended; see keyword angular_acc if interested). Even then, switching off the subdivision of radial shells can only decrease the performance.

Deprecated flag because basis_dep_cutoff should supersede this functionality in a more organized way. Having a separate, tighter cutoff for core radial functions sounds like a good idea, but core radial functions are already rather localized anyway. Our experience is that the separate core setting either does not matter for CPU time, or already introduces noticeable total energy changes when it matters.