Fei Zhou
Last modified: Oct. 1, 2020
Contents
CSLD, or compressive sensing lattice dynamics, is a comprehensive package to study lattice dynamics from first-principles. The interatomic force constants were fitted using the compressive sensing algorithm. Some of the attractive features of CSLD are
- Accurate: up to 6th order anharmonic terms. Optimized treatment of long-range Coulomb interactions in polar semiconductors.
- Robust: fitting regularized with L1 penalty
- Efficient: using as few training structures as possible
- Versatility: phonon (harmonic) and anharmonic force constants, vibrational free energies, interatomic potentials for solid molecular dynamics simulations.
CSLD is released as an open source package under the MIT license. For issues, please report via github, or contact Fei Zhou mailto:[email protected]
Hereby we refer to all files from the main source code directory. User input commands start with "$", followed by output without leading "$":
$ ls setup.py setup.py
Prerequisite
- Python 3
- Python packages: numpy, scipy, matplotlib, ConfigParser
- Library spglib
- f2py3 associated with Python3. If it is available in another name, e.g. f2py, make it available by e.g. alias f2py3=f2py
- C++ and Fortran 90 compilers
If the prerequisites are not met, please either ask your sys admin to install them, or install your own python 3 environment, e.g. miniconda or anaconda. The codes were tested on conda installations on Linux and MacOS.
- If you have no plan to modify/contribute to the source codes, go to the code directory and install for all users
$ python3 setup.py install
- Install for yourself without admin rights
$ python3 setup.py install --user
- If you plan to develop (i.e. modify) the codes
$ python3 setup.py develop
- Develop without admin rights
$ python3 setup.py develop --user
- If compilation fails, you may need to modify the Makefile manually.
If you use CSLD in your research, please cite the following works:
- Fei Zhou, Weston Nielson, Yi Xia, and Vidvuds Ozolins, Phys. Rev. Lett. 113, 185501, (2014). http://dx.doi.org/10.1103/PhysRevLett.113.185501
- F. Zhou, W. Nielson, Y. Xia, and V. Ozolins, Phys. Rev. B 100, 184308 (2019). https://link.aps.org/doi/10.1103/PhysRevB.100.184308
- F. Zhou, B. Sadigh, D. Aberg, Y. Xia, and V. Ozolins, Phys. Rev. B 100, 184309 (2019). https://link.aps.org/doi/10.1103/PhysRevB.100.184309
More detailed documentation can be found in the Manual section next. This section provides a short tutorial To get started quickly after installation. The files can be found in the examples/ directory.
CSLD requires an input structure file of very high accuracy. See the Manual section for details to prepare a VASP POSCAR-styled file first.
All input files have been prepared in the examples/Si/ folder.
- Phonon dispersion
$ cp -r examples/Si examples/Si-test; cd examples/Si-test $ csld_main -f csld.in # or just csld_main
Then Check out the generated plots.pdf
- Third and fourth order anharmonicity with modified config file
$ csld_main -f csld.in-anharmonic
The above can be equivalently achieved by modifying the command line rather than the config file
$ csld_main --override "[fitting] solution_known=sol_2nd" \ --override "[fitting] submodel1=anh 3 4" \ --override "[fitting] solution_out=solution_all" \ --override "[training] traindat1=fcc333/SPOSCAR fcc333/dir*0.06"
Now let's see how to start from scratch, by copying over the config file control.in and symmetrizing a primitive cell Si/POSCAR (pretending that Si/POSCAR was relaxed without proper symmetry):
$ cd ..; mkdir Si-test2; cp Si/csld.in Si-test2; cd Si-test2 $ polaron_main --task primitive_no_standardize --prim ../Si/POSCAR --tol 0.001 >POSCAR
Next, prepare training data:
$ mkdir fcc333; cd fcc333; echo -e "-3 3 3\n 3 -3 3\n 3 3 -3" > sc.txt $ polaron_main --task supercell --p1 sc.txt --prim ../POSCAR > SPOSCAR
where sc.txt is a 3x3 integer scaling matrix specifying the supercell. Perturbed supercells for phonon calculations can be generated in subdirectories by:
$ polaron_main --task rand_disp_dir -R 0.01 --p1 SPOSCAR -N 1 $ ls * SPOSCAR sc.txt dir_00000-disp0.01: POSCAR
where the supercell fcc333/SPOSCAR was perturbed by R=0.01 angstrom randomly for N=1 time. For anharmonic terms, larger displacements should be attempted, e.g.:
$ polaron_main --task rand_disp_dir -R 0.06 --p1 SPOSCAR -N 1
The user should then run DFT codes (e.g. VASP) to compute the corresponding total forces. Once done, the latter can be collected using the supplied get-force.sh script:
$ get-force.sh -d dir_00*/
This is an example that requires long-range forces
- Phonon calculation without long range forces. Note the absence of LO-TO splitting in the obtained plots.pdf
$ cp -r examples/NaCl examples/NaCl-test; cd examples/NaCl-test $ csld_main --override '[phonon]nac=-1'
- With long range forces
$ csld_main
Never mind the artifact in the phonon dispersion curve at zone center. It's the non-analytic correction.
- Third and fourth order anharmonicity trained on fcc333/dir_00000-disp0.06
$ csld_main -f csld.in-anharmonic
- Export second and third-order force constants files (FORCE_CONSTANTS_2ND and FORCE_CONSTANTS_3RD) for thermal conductivity calculations in ShengBTE
$ csld_main --override '[export_potential] export_shengbte=5 5 5 2 3' \ --save_pot_step 1 --phonon -f csld.in-anharmonic
To obtain the Born effective charges and dielectric tensor required for long-range treatment, dielectric calculations should be performed with density functional perturbation theory (DFPT) for the primitive cell before csld fitting. Obtain born_charge.txt and epsilon_inf.txt by
$ polaron_main --task born --p1 PATH_TO_DFPT_CALCULATION/OUTCAR
Failure to use this script to generate symmetrized files may result in errors.
The main executable is the Python 3 script scripts/csld_main. Other executables include the scripts/polaron_main script for various auxilliary functions. Help is available via:
$ csld_main -h $ polaron_main -h
The csld_main code takes two basic input files: POSCAR and csld.in.
- The structure of the primitive cell in VASP 5 format. The file name is specified in the config file (POSCAR, see above). It's important to keep high precision in the structure file. We STRONGLY recommend using the polaron_main helper script to symmetrize your primitive cell
$ polaron_main --task primitive_no_standardize --prim your_input_POSCAR --tol 0.001 >POSCAR
csld.in, the main configuration file containing essentially all the settings.
May be specified with the -f switch (see examples above)
The format is the usual configuration file with section headers and variable names under each section. Comment starts with "#". For example, the following specifies "POSCAR" as the primitive cell structure and 0.001 as the tolerance of symmetry finding (using spglib):
[structure] # primitive cell prim = POSCAR # symmetry finding tolerance, NOT used yet! sym_tol = 1E-3
The tag (e.g. "prim") is case insensitive. Our examples come with all lower case tags.
We recommend copying over one of the config files in the test/ example and adapting to your needs.
An alternative to editing csld.in is to change settings at the command line via the "--override [HEADER] TAG=VAL" option, where HEADER is a section in csld.in, and TAG and VAL are the desired entry under [HEADER]. This is equivalent to editing the corresponding entry in csld.in. For example, to change the primitive cell to PRIM
$ csld_main --override "[structure] prim = PRIM"
- Additionally, for polar semiconductors, the Born effective charege and static dielectric tensors are required to describe the long-range interactions. Perform DFPT calculations and extract the required tensors into files born_charge.txt and epsilon_inf.txt by
$ polaron_main --task born --p1 PATH_TO_DFPT_CALCULATION/OUTCAR
Additional input files include the supercell structure(s), perturbed structures and associated total forces.
Once the supercell size is determined, the main variable in the above setup is the number of structures (-N switch) to generate.
- For phonon calculations, determine the pair-interaction cutoff distance (say 8 angstrom) and estimate the number perturbations on supercell fcc333
$ polaron_main --task nsc --prim POSCAR -R 8 --p1 fcc333/sc.txt
- For anharmonic calculations, start with a few supercell calculations with larger displacement, e.g. with the same number as phonon calculations, and incrementally add more calculations as needed. Note that the most significant anharmonic terms are typically short-ranged, so one might try smaller supercells than phonon calculations.
The csld_main script runs in the following steps, each controlled by a command-line argument of csld_main with detailed settings under a section of the csld.in config file. The most important settings are given below, too.
- Model setup, i.e. generating clusters, as well as identifying independent model parameters (implemented as computation of a symmetrization matrix C)
- command line switch "--clus_step STEP". STEP=2 means generating clusters, 3 (default) =generate & save to file clusters.out.
- command line switch "--symC_step STEP". 1=load file, 2=compute, 3 =compute & save to file Cmat.mtx
- users usually should focus on the following settings rather than changing the command line switches unless they know what they are doing!
| tag | value | description |
|---|---|---|
| max_order | int | max order of clusters, e.g. 2 for harmonic model, 4 for up to fourth-order |
| cluster_diameter | real [real...] | cluster cutoff for pair, triplet, ... e.g. 8.0 4.5 |
| fractional_distance | True|**False** | whether distance is scaled relative to lattice constant (2nd line of POSCAR) |
- Training init step, i.e. computation of the correlation matrix
- switch "--train_step STEP". 1=load file, 2=compute, 3 =compute & save, 4=skip.
| tag | value | description |
|---|---|---|
| corr_type | str | Which type of property to fit: f for force, e for energy |
| cluster_diameter | real [real...] | cluster cutoff for pair, triplet, ... e.g. 8.0 4.5 |
| traindat* | str_SC str_dir1 [str_dir2...] | POSCAR of supercell, followed by list of subdirs, e.g. fcc333/POS fcc333/dir-*. Multiple traindat entries with different supercells may be supplied to mix and match supercells of different sizes, e.g. traindat1= fcc222.... \n traindat2= fcc333.... |
- Fitting step to obtain model parameters using compressive sensing or other algorithms.
- switch "--fit_step STEP". 1=file, 2=generate, 3 =generate & save.
| tag | value | description |
|---|---|---|
| method | int | 1=FPC, 3=split Bregman 5=split Bregman + right preconditioning, 201=ridge regression |
| nsubset | int | number of subset fittings |
| uscale_list | real [real...] | displacement scale. Approximately 0.01 for phonon calculations, larger values for anharmonic fitting |
| mulist | real [real...] | list of mu (weight of L1 or L2 norm in penalty) to loop over, e.g. 1e-1 1e-3 1e-4 1e-5 1e-6 |
| submodel* | str int [int] | name of the fitting, and list of orders included, e.g. harmonic 1 2 (fitting the 1st and 2nd-order FCs only such that FCs of other other will be set to zero). Multiple entries may be entered to test different fittings in one run, e.g. submodel0=harmonic 1 2 \n submodel1=up-to-third 1 2 3. If no submodel is supplied, default to fitting all orders up to max_order. |
| solution_out | str | filename for the obtained vector of solution (independent parameters) |
| solution_in | str | filename for loading previous solution instead of fitting |
| solution_known | str | filename for previously obtained parameter phi_in. The force predicted by phi_in will be subtracted from the total force. This is useful in conjunction with submodel to fit in several steps, e.g. assuming max_order=4, first fit harmonic terms with submodel=harmonic 1 2; solution_out=sol_2nd, then fit anharmonic terms with solution_known=sol_2nd; submodel1=anh 3 4; solution_out=solution_all |
- Pairwise force-field setup step to capture the bulk of the anharmonicity and to make fitting the residual force/energy easier. Turned off by default
- switch "--ldff_step STEP". **0**=off, 2=on.
| tag | value | description |
|---|---|---|
| num_basis | int | Number of basis |
| orbit_indices | int [int] | list of clusters to include in LDFF, usually choose only the nearest-neighbor pairs |
| interpolation_pts | int int int | min, max, interval of sampling points |
| basis_2 | str | pair basis functions. See Si/csld.in-forcefield for an example |
- Phonon step
- switch "--phonon_step STEP". 0=skip, 1 = compute.
| tag | value | description |
|---|---|---|
| nac | int | Method for non-analytic correction. -1 =disabled, 0=long range treatment in PRB 100, 184309 (2019). |
| wavevector | str | If specified, plot phonon dispersion. wavevector = Auto will turn on automatic generation of special paths in reciprocal space. Manual settings e.g. [[25, [0,0,0],'Gamma', [0,0.5,0.5], 'X']] will add 25 points between zone center and X point |
| unit | str | Unit for dispersion and DOS. One of THz, meV, eV, cm |
| dos_grid | int x 3 | If specified, plot density states sampled on a grid, e.g. 10 10 10 |
| ismear | int | smearing method of DOS integration. 0=Gaussian, 1=Lorentzian, -1 =tetrahedron method |
| epsilon | real | Smearing width for Gaussian/Lorentzian |
| pdos | True|False | Whether to plot partial DOS |
| thermal_t_range | real x 3 | if specified together with dos_grid, calculate quasi-harmonic thermodynamic properties in the temperature range: begin end increment in Kelvin, e.g. 50 800 50. |
| thermal_out | str | filename for thermodynamic properties |
- Exporting force constants step.
- switch "--save_pot_step STEP". 0 = skip, 1=save.
| tag | value | description |
|---|---|---|
| export_shengbte | int x 3 int [int...] | If specified, export force constants for calculation in ShengBTE. First 3 integers designate size of supercell for pair force constants, followed by list of orders to export. e.g. exporting Hessian matrix of a 5x5x5 supercell, as well as 3rd-order FCs with 5 5 5 2 3 |
- Prediction of supercells forces or energies with known solution
- switch "--pred_step STEP". 0 =skip, 1=load training setup correlation matrix from file, 2=compute, 3=compute & save. Use "csld_main --predict" to quickly predict supercells"
- Settings in "[prediction]" with the same sets of tags as "[training]". Usually used together with "[fitting] solution_in = previous_solution"
- The list of symmetrized force constants: solution.out
- For phonon calculations, the phonon plots (plots.pdf), band dispersion (phonon-dispersion.out) and density of states (phonon-total-dos.out, phonon-partial-dos.out) data files, as well as thermodynamic properties (free energy, vibrational entropy, etc, in thermal.out) in the quasi-harmonic approximation.
- The second and third order FCs can be exported to perform phonon scattering rates and thermal conductivity calculations in ShengBTE. Assuming a 5 x 5 x 5 supercell for the harmonic FCs
$ csld_main --override '[export_potential] export_shengbte=5 5 5 2 3' \ --save_pot_step 1 --phonon
- The fourth order FCs can be exported to a modified version of ShengBTE that consider frquency shift and scattering due to four-phonon processe
$ csld_main --override '[export_potential] export_shengbte=5 5 5 2 3 4' \ --save_pot_step 1 --phonon
This is a list of Frequently Asked Questions about CSLD. Feel free to suggest new entries!
- ... select cutoffs for force constants?
- For anharmonic FCs, include at the very least the second NN shell for 3rd order and first NN shell for 4th order.
- ... decide the number of supercells needed?
- Once you made a choice on your model based on cutoffs, look for the final number of symmetrized FCs left. The number of supercell calculations should be at least N(symm)/3N(atom).
- ... use CSLD with another DFT code (Quantum ESPRESSO, etc)?
- Quite minimal: convert the structure to VASP5 format, and supply the computed forces to force.txt as a plain text table. It might be slightly trickier to figure out how to extract the dielectric constant or Born effective charge for semiconductors.
- ... "ERROR: to get corrections properly, please increase [model]dpcor_bond to approximately the cutoff distance of first neighbor shell"
- This means that you have a semiconductor system that needs dipole interaction corrections (see CSLD phonon paper PRB 2019). Set dpcor_bond under [model] as suggested.