Skip to content

README.md

Latest commit

 

History

History
123 lines (84 loc) · 4.32 KB

README.md

File metadata and controls

123 lines (84 loc) · 4.32 KB

MAD 2.0: multi-target foundation data computed with FHIaims

Wiki authors: Joe Abbott, Cesare Malosso

Tip

What is this guide for? The idea is to provide a common set of FHIaims settings for all COSMO-nauts who want to run DFT on their dataset to use. By doing so, it will allow us to coherently perform fine-tuning, transfer learning, and application-focussed test cases of the (not yet existing) PET-MAD-2 mutli-target foundation model.

User Guide

DFT settings

We have constructed a single FHIaims settings file (a control.in) that contains the standard DFT, computational, and target-specific settings for generating data. Calculations are performed in a two-step process. First, SCF is converged and the following targets generated.

Those generated as default:

- energy
- forces
- stress
- Density Matrix (k-space, also acts an SCF restart file)
- Hamiltonian (real-space, printed with HDF5)
- Density of States

Then an SCF restart is optionally performed to generate the extra targets:

- polarization (in the case of non-metals only)
- electron density basis decomposition

Warning

We are currently in the process of testing the restart functionality for generating the polarization and density decompositions. As these are submitted as subsequent and separate calculations, generation of the above targets can be performed as usual first, with extension to these optional targets if necessary. Check back soon for details.

The outputting of these targets and all other required DFT settings are provided in the example control.in with header:

# ==============================================================================
# One control.in to rule them all, and in the darkness bind them.
# 
# Created by COSMOnauts, for COSMOnauts.
#
# ==============================================================================
...

Species Defaults

The tight defaults are taken from FHIaims version 250610, from directory FHIaims/species_defaults/defaults_2020/tight.

Then, the defaults of the Lanthanides and Actinides are modified in the following way, removing an extra confinement basis function:

for file in *_default; do
    if [ -f "$file" ]; then
        # Only edit files that contain a line starting with "# confined"
        if grep -q "confined" "$file"; then
            sed -i 's/confined/# confined/' "$file"
            echo "Commented 'confined' in $file"
        fi
    fi
done

To give the species settings found in species_defaults/tight_modified.

Example calculation - Jed (SCITAS)

An example demonstrating its usage is available in the example/ directory. The provided submission script is configured for the Jed system but can be adapted to other HPC environments as needed.

To run this example, copy the contents of example/ to somewhere in your home/ folder. Run sbatch run-aims_jed.sh to compute all the targets.

Scripts

Python script for reading an xyz trajectory file and setting up an FHIaims calculation for each frame is available at scripts/prepare_aims_input.py.

Extra Info

How FHIaims was built

Important

To be updated once this merge request on the FHIaims GitLab has been merged to the main branch.

FHIaims has been compiled from source in the directory build/.

Steps to reproduce the build on Jed:

  1. Download the source code from GitLab. This requires an account with access permission to FHIaims. Switch to the commit pinned to the 250626 version, which specifically enables support for ScaLAPACK and HDF5.
git clone https://aims-git.rz-berlin.mpg.de/aims/FHIaims.git
# enter username and password
cd FHIaims
git checkout bbb7258b1c1ed9f3283031dda3bea1bc74bd3a26
  1. Create a build directory and copy the build files (specific to Jed) into it.
mkdir build
cd build
cp  ../../jed_settings/* .
  1. Run cmake and make
./run-cmake.sh

Check that FHIaims version: 250626 is printed. Once finished successfully, run the build:

sbatch run-make.sh

The executable is then under the file name aims.250626.scalapack.hdf5.mpi.cosmo.mad.2. For convenience, we also store this on SCITAS at path: /work/cosmo/COSMO-MAD-2/executable/aims.250626.scalapack.hdf5.mpi.cosmo.mad.2.