This guide deal with the general setup and local installation of the
There are several ways to obtain the
A precompiled version of the program can be obtained from the
latest release page
To get the newest features an automatically generated
is available for the latest commit on the default branch.
At the release page you find the a compressed tarball (
tar xf xtb*.tar.xz)
usually containing the following content:
. ├── bin │ └── xtb ├── include │ └── xtb │ └── xtb.h ├── lib │ ├── libxtb.so -> libxtb.so.6 │ ├── libxtb.so.6 -> libxtb.so.6.3.1 │ ├── libxtb.so.6.3.1 │ └── pkgconfig │ └── xtb.pc └── share ├── man │ ├── man1 │ │ └── xtb.1 │ └── man7 │ └── xcontrol.7 ├── modules │ └── modulefiles │ └── xtb │ └── 6.3.1 └── xtb ├── param_gfn2-xtb.txt ├── param_gfn1-xtb.txt ├── param_gfn0-xtb.txt ├── param_ipea-xtb.txt └── .param_gfnff.xtb
The binary is usually compiled with the Intel Fortran compiler and statically
linked against Intel’s Math Kernel Library (Intel MKL).
Newer versions of
xtb (6.2 and newer) additionally include a shared library,
the header specification of the C-API.
First check the version by
> xtb --version ----------------------------------------------------------- | ===================== | | x T B | | ===================== | | S. Grimme | | Mulliken Center for Theoretical Chemistry | | University of Bonn | ----------------------------------------------------------- * xtb version 6.2.1 (bf8695d) compiled by 'ehlert@majestix' on 2019-10-25 normal termination of xtb
This should print some fancy banner, the version number, say 6.2.1, the
git commit hash, the user and host name of the programmer compiled the
program and at date.
Make sure the version is matching the version from the release page and
the commit hash is the actual commit hash of the release, also the date
of the compilation should match.
We compile the program on our clusters in Bonn, so the host name is usually
hedy20 for the old-kernel version).
You get also the user name of the programmer drafting the release,
testing the binary and packing the tarball.
This line is very important when reporting a bug to us, because we can validate the binary you were using and easier reproduce it.
xtb from the conda-forge channel can be achieved by adding conda-forge to your channels with:
conda config --add channels conda-forge
Once the conda-forge channel has been enabled,
xtb can be installed with:
conda install xtb
It is possible to list all of the versions of
xtb available on your platform with:
conda search xtb --channel conda-forge
The conda package manager can become quite slow when adding large channels like conda-forge, for a more performant alternative you can try to use mamba instead, which can be conveniently installed from the conda-forge channel with
conda install mamba -c conda-forge
Alternatively, the mambaforge installed can be used to bootstrap a conda installation with mamba preinstalled.
This section will give you the basic information you need to
know about the
xtb program. Some of the steps are elemental
for your calculation to succeed, so please consider to follow
my instructions carefully.
Some part of the
xtb program can be quite wasteful with stack memory,
to avoid stack overflows when calculating large molecules, you should
unlimit the system stack, e.g. with
> ulimit -s unlimited
Note that the memory management of
xtb is constantly improved to avoid
using large amounts of stack memory, but to be on the save side
include this option for production runs.
xtb program uses shared memory OpenMP parallelisation, to calculate larger systems
an appropriate OMP stacksize must be provided, chose a reasonable large number by
> export OMP_STACKSIZE=4G
Note that the memory requirement will increase with the system size and the number of requested threads. Currently, for 3000 atom system each thread requires around 16GB of memory due to the usage of dense matrix algebra for the xTB Hamiltonian. You have to account for this by allowing an accordingly larger OpenMP stacksize.
To distribute the number of threads reasonable in the OpenMP section it is recommended to use
> export OMP_NUM_THREADS=<ncores>,1
You might want to deactivate nested OMP constructs by
> export OMP_MAX_ACTIVE_LEVELS=1
Most OpenMP regions allow to customize the scheduling by setting
for many threads the
dynamic schedule has proven to give a good load-balance
between all threads.
The default linear algebra backend of xtb is the Math Kernel Library,
which usually comes as OpenMP threaded version and respects the settings
OMP_NUM_THREADS. You can however still adjust the MKL parallelisation
separately by setting
> export MKL_NUM_THREADS=<ncores>
A number of environment variables is used by
xtb to perform calculations.
Please set the
XTBPATH variable to include all locations were
you store information relevant for your
xtb calculation, like configuration
files and parameter files.
The present working directory is implicitly included for most files that
are searched in the
Available since version 6.3.2
A tcl environment module is provided and can be used with usual module systems.
For installations from the tarball the
prefix variable in the module file
has to be adjusted accordingly
--- ./share/modules/modulefiles/xtb/6.3.1 +++ ./share/modules/modulefiles/xtb/6.3.1 @@ -1,5 +1,5 @@ #%Module -set prefix / +set prefix /absolute/path/to/xtb module-whatis "Semiempirical Extended Tight-Binding Program Package"
share/modules/modulesfiles directory is included in your
you should be able to load
module load xtb
If you plan to use the
xtb shared library in your build system you have
to do a similar adjustment to the
Example scripts to be sourced in your shells rc file are included in the distributed tarball:
and should setup all environment variables correctly in most cases.
Beside this manual you can check the in-program help by
> xtb --help
Unfortunately, this might be outdated,
therefore, you should refer to the man-pages distributed with the
Please check for the man-pages of
There is also an online documentation, but you already now that one, of course.
If you think some information is missing in your calculation you can
switch to the verbose mode by using
--verbose in the command line
arguments. This will increase the print level almost everywhere in the
xtb program, also the input parser will print a lot of information
that might be interesting for your current calculation.
Overall this can be an awful lot of information, so it is not recommended as a default option.
Orca 4.2 implements support for xTB calculations using an IO based interface
xtb binary and parsing its output.
The binaries of Orca will call an executable called
should be placed in the directory containing the Orca binaries.
We recommend to create a symbolic link to your local
xtb binary by
> ln -s $(which xtb) otool_xtb
xtb version 6.2.3 produces an energy printout which cannot
be processes by the reader in Orca, to fix this issue, use
the provided script to wrap the
xtb binary instead
of creating a symbolic link.
You can invoke xTB calculations in Orca by using one of the simple keywords
! XTB1 # for GFN1-xTB ! XTB2 # for GFN2-xTB
in your Orca input file, for more details refer to the Orca manual.
Orca will communicate with
xtb mainly by using commandline arguments,
requesting singlepoint calculations and parsing the total energy and
gradient from the program output.
Of course you should setup the
xtb related environment variables,
xtb can find its parameter files and configuration files.
.xtbrc is still read if it is contained in
XTBPATH and can
be used to change the behaviour of xTB calculations in Orca, e.g. for
setting the electronic temperature.