1. JARVIS-FF

JARVIS-FF was created to evaluate classical force-fields for various applications. Many FFs are designed for a particular application in mind, hence, its necessary to assess their performance uniformly in a systematic way.

1.1 Crystal structures

As shown in the slides, atomic structures in terms of their coordinates and translational vectors are key inputs for any atomistic calculation. We obtain crystal structures from publicly available DFT databases such as materials project, open quantum material database, aflowlib and literature. We utilize the API framework available with all these databases to obtain the crystal structures. Unless specified we used conventional cells of materials during our FF calculations.

1.2 Force-fields/interatomic potentials

Force-fields were downloaded from Interatomic Potential Repository (IPR) and LAMMPS source-code distribution.

1.3 Elastic constant calculation:

We use LAMMPS as energy and force calculator. We use minimize along with fix box/relax command to minimize the cell. The high-throughput setting of LAMMPS jobs was done using the jarvis code. Using such a high-throughput framework we calculated the 21 distinct elastic constants using the script provided in the LAMMPS distribution version 15May2015. It is to be noted that the above script is one of the many other methods to calculate elastic constants and comparison and integration of other scripts with the one used here is subject to future work. Such a script allows to computation of elastic constant and energetics for any available empirical potential type and any crystal structure. In our runs we used 10^-06 as strain, 10^-10 eV/Å for force convergence criteria during the minimization to optimize the structure and 1000 maximum iteration for structure optimization. These are generalized computational set-up parameters, and the energetics and elastic constant data may or may not depend on them. We tested strain parameters for a range of values (10^-04, 10^-06 and 10^-08) but obviously evaluating such set of parameters for all the calculations was too extensive a work and was not carried out here.

1.4 Vacancy formation:

The vacancy structures were created by deleting symmetrically distinct sites on the relaxed structure from the above minimization step. We used Pymatgen code for making defect structures. We relax only internal positions during vacancy formation energy calculations. The reference elements for vanacancy formation energies were taken from DFT databases (materials project) most stable configuration (energy above hull zero). Only electrically neutral defects were considered for this project.

1.5 Surfaces:

Similar to vacancies, symmetrically distinct surfaces were created with Pymatgen on the relaxed structure with maximum miller index upto 3. All of the surfaces have normal along z-direction.

1.6 Phonon:

Phonon calculations were done with Phonopy package. Symmetrically distinct perturbed structures were created with Phonopy and then forces were created with JARVIS-FF. We used ASE package to transfer structural data in Phonopy input format. The phonon density of states and bandstructures were created with Phonopy.

Details of the method can be found in our paper Evaluation and comparison of classical interatomic potentials through a user-friendly interactive web-interface

2. JARVIS-DFT

JARVIS-DFT was created to find new two-dimensional layered materials and create a consistemt database of functional materials. Complete database can be downloaded in JSON format now!

2.1 Convergence:

Calculations are done using VASP software. Convergence on KPOINTS and ENCUT is done with respect to total energy of the system within 0.001 eV tolerance. Please note convergence on KPOINTS and ENCUT is generally done for target properties, but here we assume energy-convergence with 0.001 eV should be sufficient for other properties also. The points on the curves are obtained with single-point calculation (nuber of ionic steps,NSW=1). However, for very accurate calculations, NSW>1 might be needed.After convergence, we optimize crystal structure with strict force and energy convergence criteria.

2.2 X-ray diffraction pattern:

X-ray diffraction (XRD) pattern and the Radial distribution function (RDF) plots were obtained from the crystal structure using pymatgen. XRD peaks should be comparable to experiments mainly for bulk 3D materials. Relative intensities may differ.

2.3 Electronic structure:

Electronic structure is characterized using electronic density of states and bandstructure. DFT is generally predicted to underestimate bandgap of materials. Accurate band-gaps are obtained with higher level methods (with high computational requirement) such as HSE, GW, which are under progress. Total DOS, Orbital DOS and Element dos buttons are provided for density of states options at each webpage. Energy is rescaled to make Fermi-energy zero. In the bandstructure plot, spin up is is shown with blue lines while spin down are shown with red lines. Non-degenerate spin-up and spin-down states (if applicable) would imply a net orbital magnetic moment in the system.

2.4 Electrostatic potential:

The plane averaged electrostatic potential (ionic+Hartree) along x, y and z-directions at webpages. The red line shows the Fermi-energy while the green line shows the maximum value of the electrostatic potential. For slab structures (with vacuum along z-direction), the difference in these two values can be used to calculate work-function of the material.

2.5 Optoelectronic properties with local/semilocal functional:

Incident photon energy dependent dielectric function is calculated. Only interband optical transitions are taken into account.Please note the underestimatation of band-gap problem with DFT will reflect in the spectra as well. For very accurate optical properties GW/BSE calculation would be needed, which is yet to be done because of their very high computational cost. Optical properties for layered materials needs to be rescaled with the actual thickness to simulation z-box ratio.

2.6 Optoelectronic properties with MBJ functional:

Single point DFT calculation was carried out with meta-gga MBJ functional. This should give reasonable bandgap, and optical properties assuming the calculation was properly converged. Incident photon energy dependence of optical is shown at webpages. Only interband optical transitions are taken into account.

2.7 Elastic constants and non-converged phonons:

Elastic tensor was calculated for the conventional cell of the system with finite-difference method. For layered materials, the elastic constants are rescaled with respect to vacuum padding (see the input files) and the units for elastic coefficients are in N/m. Phonons obtained from this calcuation are also shown. WARNING: Please note this may not be the exact phonon modes of the system as we did not test the cell-size dependence of phonons yet. At least 1.2 nm x1.2 nm x1.2 nm or more is needed for obtaining reliable phonon spectrum. For systems having primitive-cell phonon representation tables, I denotes infrared activity and R denotes Raman active modes (where applicabale).

2.8 Thermoelectric properties with constant time relaxation:

Thermoelectric properties are calculated using BoltzTrap code. Electron and hole mass tensors are given at 300 K. Plots show the Seebeck coefficient and ZT factor (eigenvalues of the tensor shown) at 300 K along three different crystallographic directions. Seebeck coefficient and ZT plots can be compared for three different temperatures available through the buttons given below. Generally very high Kpoints are needed for obtaining thermoelectric properties. We assume the Kpoints obtained from above convergence were sufficient. WARNING: Constant relaxation time approximation (10^-14 s) and only electronic contribution to thermal conductivity were utilized for calculating ZT.

Detials about the work can be found at High-throughput Identification and Characterization of Two-dimensional Materials using Density functional theory

JARVIS-ML

JARVIS-ML uses data from the atomistic calculations (mainly JARVIS-DFT) to predict properties of materials in a computationally cheap way for prescreeing of materials. JARVIS-ML uses structural as well as chemical features of materials.

More details coming soon !

NIST Disclaimer

---