DiagonalizationSolver

class DiagonalizationSolver(bands_above_fermi_level=None, processes_per_kpoint=None, optimize_for_speed_over_memory=None)

Calculate the density matrix by direct diagonalization.

Parameters:

  • bands_above_fermi_level (int | Automatic | All) – The number of bands above the Fermi level. Must be either a non-negative integer, All or Automatic. When set to All the total number of bands will be equal to the total number of basis functions. When set to Automatic an adaptive method will be used. For most systems, Automatic provides a major speed-up as compared to All. If the number of states above the Fermi level changes significantly from one diagonalization step to another, the adapative method may lead to slowing down of the calculation.
    Default: Automatic

  • processes_per_kpoint (int) – The number of processes to use per kpoint. Must be a positive integer.
    Default: The number will be determined automatically from the total number of k-points and processes such as to keep the number as small as possible. One may set this number manually in order to reduce the memory requirements for each process.

  • optimize_for_speed_over_memory (bool) – Whether to optimize the diagonalization for speed over memory. If True, the solver will determine the most optimal algorithm in terms of speed. Note that optimizing for speed can increase the memory usage by up to 30 %. Note also that this flag only has an effect if processes_per_kpoint=1, or the automatic value for processes_per_kpoint=1 turns out to be 1. It is intended to be used for medium-sized to large systems with many k-points.
    Default: False

bandsAboveFermiLevel()
Returns:

The initial guess for the number of bands above the Fermi level.

Return type:

int | All

optimizeForSpeedOverMemory()
Returns:

Whether the diagonalization should be optimized for speed over memory.

Return type:

bool

processesPerKpoint()
Returns:

Number of processes per kpoint.

Return type:

int

uniqueString()

Return a unique string representing the state of the object.

Usage Example

Define a density matrix method that uses the DiagonalizationSolver:

algorithm_parameters = AlgorithmParameters(
    density_matrix_method = DiagonalizationSolver())

Notes

Diagonalization algorithm

The library used for diagonalization of the LCAO Hamiltonian depends on the value of the parameter processes_per_kpoint:

  • LAPACK is used when a single processor is used for each k-point, i.e. processes_per_kpoint=1.

  • ELPA [1] is used when more than one processor is used for each k-point, e.g. processes_per_kpoint=2.

Number of bands included in the diagonalization

When the parameter bands_above_fermi_level is set to Automatic (default), the diagonalization solver uses an adaptive method. The method determines the number of occupied bands to include when diagonalizing the Hamiltonian after each diagonalization step. Then, it increases the number of occupied states by a large enough margin such that the next diagonalization step is guaranteed to include all occupied bands.

Attention

If the number of bands above the Fermi level is too low such that not all occupied states are included, the diagonalization is repeated with an increased number of bands. This can slow down the calculation, and therefore setting bands_above_fermi_level manually is not recommended.

The behavior of bands_above_fermi_level with regard to the spin type UNPOLARIZED, POLARIZED, and NONCOLLINEAR is described in Notes of the Bandstructure analysis object.