TBLITE

References: Katbashev2025

Section used to specify options for an xTB computation using tblite. This section is required when XTB/GFN_TYPE is set to TBLITE. Periodicity follows SUBSYS/CELL/PERIODIC; the deprecated XTB/DO_EWALD keyword is read consistently with CP2K-internal xTB for diagnostics/debugging. Initial densities and restarts are controlled by the regular CP2K SCF settings, for example DFT/SCF/SCF_GUESS, not by a separate tblite guess in this section. Spin-polarized CP2K/tblite calculations are selected with the regular DFT/UKS keyword or its LSD/SPIN_POLARIZED aliases; DFT/MULTIPLICITY sets the number of unpaired electrons. UKS uses tblite’s spin-polarized Hamiltonian; AUTO selects the native tblite SCC mixer for XTB/GFN_TYPE TBLITE, while SCC_MIXER CP2K mixes both total-charge and magnetization SCC variables through CP2K’s TB charge-mixing infrastructure. Native tblite SCC convergence follows ACCURACY; SCC_MIXER CP2K follows the regular CP2K SCF convergence settings such as DFT/SCF/EPS_SCF. With QS/LS_SCF, CP2K disables the separate SCC-variable mixer and lets LS_SCF optimize the density matrix directly. The same direct SCC-variable update is available for Gamma-point OT calculations, where smearing and ADDED_MOS are not supported. [Edit on GitHub]

Subsections

REFERENCE_CLI

Keyword descriptions

SECTION_PARAMETERS: logical = F 

Lone keyword: T

Marks the tblite options section as present. The actual switch to the tblite backend is XTB/GFN_TYPE TBLITE; this section then collects the numerical model, SCC mixing, and optional native-reference settings used by that backend. [Edit on GitHub]

ACCURACY: real = 1.00000000E+000 

Usage: ACCURACY 1.0

Mentions: ⭐Extended Tight Binding

Numerical accuracy factor for the CP2K/tblite backend. This scales tblite’s internal accuracy-dependent thresholds. With the native tblite SCC mixer, this is the user-visible convergence control and corresponds to native tblite’s –acc setting. With SCC_MIXER CP2K, SCF convergence instead follows DFT/SCF/EPS_SCF; ACCURACY still controls the tblite numerical setup, such as the distance cutoff used to build the xTB basis and neighbor data. The default matches native tblite. Smaller values make the calculation tighter and usually more expensive; larger values loosen the numerical setup and can affect energies, forces, and stress. Native REFERENCE_CLI diagnostics inherit this value so the production and external reference calculations use the same accuracy setting. [Edit on GitHub]

METHOD: enum = GFN2 

Usage: METHOD (GFN1|GFN2|IPEA1)

Valid values:

GFN1 Uses the GFN1-XTB method by Grimme.
GFN2 Uses the GFN2-XTB method by Grimme.
IPEA1 Uses the IPEA1 method by Grimme.

Mentions: ⭐Extended Tight Binding

Selects the xTB Hamiltonian model used by tblite. Without PARAM this chooses one of tblite’s built-in parametrizations, including the basis, reference occupations, repulsion, dispersion, electrostatic, and coordination-number dependent terms. When PARAM is given, METHOD remains the declared model family that the custom parametrization is expected to implement and that CP2K uses for method-dependent setup and diagnostics. [Edit on GitHub]

PARAM: string

Usage: PARAM filename

Mentions: ⭐Extended Tight Binding

Custom tblite parametrization file. The file is read by tblite and used to construct the xTB calculator instead of the built-in parameter table for METHOD. It may therefore change model constants such as Hamiltonian, repulsion, dispersion, electrostatic, and coordination-number response parameters. It does not change the CP2K charge, spin, cell, k-point, or SCF setup. The parametrization must be compatible with the selected METHOD; CP2K does not reinterpret or validate the physical content of the file. [Edit on GitHub]

REFERENCE_CLI: logical = F 

Lone keyword: T

Usage: REFERENCE_CLI

Enable an optional diagnostic comparison against an external native tblite command line calculation. CP2K writes the current geometry, charge, spin, method, optional PARAM file, electronic temperature, and requested reference settings to temporary files, runs the external program, and compares the resulting energy, gradient, and optionally virial with the CP2K/tblite values. This check is meant for debugging backend consistency, not for production sampling. The native tblite CLI exposes only an electronic temperature for the occupation broadening; if CP2K uses another smearing method, the reference run prints a warning and uses Fermi-Dirac smearing with the CP2K electronic temperature. If true, the XTB/TBLITE/REFERENCE_CLI section must be present. The check is also enabled whenever that section is present. It is skipped for CP2K multi-k-point calculations, because the native tblite CLI does not reproduce CP2K KPOINTS sampling. [Edit on GitHub]