Methods
Accurate XC functionals for surface and materials chemistry
The XC functionals we develop are based on a combination of data science and physical constraints. Training data consists of experimental and quantum chemistry reference data. We thus devise functionals that are accurate for a number of relevant materials properties.
The metaGGA MCML [Brown, Maimaiti, Trepte, Bligaard, & Voss, J. Comput. Chem. 42, 2004 (2021)] is a multipurpose functional for materials science yielding good surface chemical as well as bulk structural predictions. Fᴏʀᴛʀᴀɴ source code implementing MCML can be downloaded from our github page. Our functionals are also included with libxc, which is interfaced to many DFT codes. Below is an example how to use MCML via libxc with the VASP code. A recent version of libxc is needed, which can be checked out from gitlab: git clone https://gitlab.com/libxc/libxc.git cd libxc autoreconf -i ./configure --prefix=libxcinstallationprefix --disable-fhc CC=gcc FC=gfortran make && make install
Replace libxcinstallationprefix with an installation path for libxc (and, e.g., gcc with icc and gfortran with ifort, respectively, if these compilers are going to be used to compile VASP instead of the GNU compilers). Following the instructions on how to compile VASP with libxc support, add the following (towards then end of) VASP's makefile.include before compilation: CPP_OPTIONS += -DUSELIBXC LIBXC_ROOT = libxcinstallationprefix LLIBS += -L$(LIBXC_ROOT)/lib -lxcf03 -lxc INCS += -I$(LIBXC_ROOT)/include
MCML can then be employed in VASP simulations with the following INCAR settings (in addition to other useful settings for metaGGA simulations, such as LASPH=.TRUE. or switching to conjugate gradients with ALGO=A in cases of poor convergence) # parameters for MCML DOI:10.1002/jcc.26732 METAGGA = LIBXC LIBXC1 = MGGA_X_MCML LIBXC2 = GGA_C_REGTPSS
The metaGGA VCML-rVV10 [Trepte & Voss, J. Comput. Chem. 43, 1104 (2022)] consists of simultaneously trained semi-local and non-local interactions, yielding an accurate and balanced description of chemi- and physisorption on metal surfaces in addition to a good description of bulk structural properties. VCML-rVV10 can be employed in VASP simulations with the following INCAR settings: # parameters for VCML-rVV10 DOI:10.1002/jcc.26872 METAGGA = LIBXC LIBXC1 = MGGA_X_VCML LIBXC2 = GGA_C_REGTPSS LUSE_VDW = .TRUE. BPARAM = 15.35
Density functional tight-binding parametrizations for very efficient, yet accurate electronic structure predictions for bulk materials and surfaces
Machine-learning interatomic potentials (MLIPs) and lattice kinetic Monte Carlo (kMC) methods allow for simulation of atomic structural evolution with great efficiency. In contrast to computationally more expensive density functional methods, information on the evolution of the electronic structure is not provided by these approaches.
Density functional tight-binding (DFTB) predicts electronic structures at a fractional cost of DFT. DFTB total energy calculations require parametrizations of interatomic repulsive terms to be fitted against DFT benchmark data. The electron structure-only terms of a DFTB Hamiltonian can be parametrized chemical element by chemical element, and can thus be efficiently parametrized. We develop such tight-binding parametrizations for bulk materials and surfaces, where the atomic structures are evolved with MLIPs or kMC, and DFTB is used to compute electronic structures on structural snapshots. In our first publication on this topic [Balzaretti & Voss, J. Chem. Theory Comput. 20, 7272 (2024)], we have optimized such parametrizations for transition metal alloys and surfaces. The corresponding Slater-Koster tables can be downloaded from this repository: https://github.com/Voss-Lab/SK_repository and employed in a DFTB code that can read such tables in the SKF file format, such as DFTB+.