Embedded Files
Tips for doing good Wannierization
Tips for doing good Wannierization
Obtaining good Wannierization is sometimes a bit of an art. Here are some good link:
- Beginner's guide for Wannierization (by Stepan Tsirkin): https://github.com/wannier-berri/wannier-berri/discussions/442
- How to construct high-quality Wannier functions (by QuanSheng Wu): https://www.wanniertools.org/tutorials/high-quality-wfs
Using WannierBerri for EPW
Using WannierBerri for EPW
Date: 11 Sept. 2026
Here is an example for CsSnI3. This material is studied by Dr. Ao Wang at UCLouvain.
While it is relatively easy to Wannierize the bands close to the Fermi level (for e.g. transport calculations)using 9 WFs (I-p) for valence and 3 WFs (Sn-p), it is difficult to use less Wannier functions and keeping the symmetries. Since the cost of EPW calculations scale at least square with the number of WF, here we try to show how to decrease the number of WFs.
To do this, the python WannierBerri code is useful as you can easily do symmetry adapted Wannier functions. Note that these are not MLWFs.
During the Wannier Developer workshop 2026 in Trieste, Dr. Stepan Tsirkin implemented the capability to convert ukk and amn files from WannierBerri to EPW compatible (branch wrtie_epw). If you read this in the future, this conversion is probably part of the main WannierBerri code.
Note that WannierBerri is a python code. While it can be parallelize with Ray, it is mostly usefull when computing properties but not when doing Wannerization.Therefore, for now, it is best to run in sequential.
Due to heavy IO, I recommend copying the QE scf.in output files to your desktop and run WannierBerri locally. It is much faster than running on an HPC in most cases.
Important: the example below use k-point symmetry (so only the IBZ) with scf.in just to be faster. However, at present, to convert to EPW format, you need to use full BZ, so using nscf.in. This is not a big limitation since EPW requires this as well.
I used PBE pseudo from PseudoDojo v0.5.
Here is the QE scf.in (note the nbnd)----------------------------------------------------------------&control calculation='scf', restart_mode='from_scratch', prefix='csi', nstep = 1000 wf_collect = .true. pseudo_dir = './' outdir='./' / &system ibrav= 1, celldm(1) = 11.871414629, nat= 5, ntyp= 3, ecutwfc = 80.0 occupations = 'fixed' nbnd = 40 / &electrons electron_maxstep = 200 diagonalization='david' mixing_mode = 'plain' mixing_beta = 0.7 conv_thr = 1.0d-11 /ATOMIC_SPECIES Cs 132.9 Cs.upf Sn 118.7 Sn.upf I 126.9 I.upfATOMIC_POSITIONS (crystal)Cs 0.0000000000 0.0000000000 0.0000000000Sn 0.5000000000 0.5000000000 0.5000000000I 0.5000000000 0.0000000000 0.5000000000I 0.5000000000 0.5000000000 -0.0000000000I 0.0000000000 0.5000000000 0.5000000000K_POINTS {automatic} 12 12 12 0 0 0----------------------------------------------------------------
Then run a standard QE bands to obtain the csi.bands.gnu
Then run WannierBerri with the following python script
berri.py ----------------------------------------------------------------# Author: S. Ponc\'e - From 09/09/2026 to 11/09/2026# During Wannier workshop in Trieste. # With help from S. Tsirkin, J.-M. Lihm, and Claude Opus 5from wannierberri.symmetry.projections import Projection, ProjectionsSetfrom irrep.bandstructure import BandStructureimport wannierberri as wbimport pickleimport numpy as npimport matplotlibmatplotlib.use("Agg")import matplotlib.pyplot as pltimport scipyimport ray
ray.init(num_cpus=4)
path_data = "../"
try: bandstructure = pickle.load(open("bandstructure.pkl","rb"))except (FileNotFoundError, OSError, EOFError) as err: print(f"Error : {err}") bandstructure = BandStructure.from_espresso( prefix=path_data + "/csi", IBstart=12, # skip the 12 semicore bands #Ecut=680, normalize=True, include_TR=True, irreducible=True, ) bandstructure.pickle("bandstructure.pkl")
try: w90data = wb.w90files.WannierData.from_npz("./csi_wandata", files=["mmn", "eig", "chk", "symmetrizer" ], ignore_missing_files=False)except (FileNotFoundError, OSError, EOFError) as err: print(f"Error : {err}") w90data = wb.w90files.WannierData.from_bandstructure(bandstructure, seedname="./csi_wandata", files=["mmn", "eig", "chk", "symmetrizer" ]) w90data.to_npz("./csi_wandata")
spacegroup = w90data.symmetrizer.spacegroupspacegroup.show()
projection_s = Projection(orbital='s', position_num=[0.5, 0.5, 0.5], spacegroup=spacegroup)projection_p = Projection(orbital='p', position_num=[0.5, 0.5, 0.5], spacegroup=spacegroup)
projections_set = ProjectionsSet(projections=[projection_s, projection_p])
w90data.set_projections(projections_set, bandstructure)
w90data.wannierise(init="amn", froz_min=3.0, # above the VBM (3.825 eV) froz_max=5.6, # below band 25 (7.020 eV) outer_min=-4.7, outer_max=8.9, print_progress_every=10, num_iter=201, conv_tol=1e-10, mix_ratio_z=1.0, sitesym=True, parallel=False )
system = wb.system.System_R.from_wannierdata(wandata=w90data, berry=True)path=wb.Path.from_nodes(system, nodes=[[0.000000, 0.000000 , 0.000000], [0.000000, 0.500000 , 0.000000], [0.500000, 0.500000 , 0.000000], [0.000000, 0.000000 , 0.000000], [0.500000, 0.500000 , 0.500000], [0.500000, 0.500000 , 0.000000], ], labels=["G","X","M","G","R","M"], length=200 ) # length [ Ang] ~= 2*pi/dk
bands_path=wb.evaluate_k_path(system, path=path)
A = np.loadtxt("../csi.bands.gnu")bohr_ang = scipy.constants.physical_constants['Bohr radius'][0] / 1e-10alatt = 11.871414629* bohr_angA[:,0]*= 2*np.pi/alattA[:,1] = A[:,1]plt.scatter(A[:,0], A[:,1], c="black", s=4)#plt.plot(A[:,0], A[:,1], c="black")
bands_path.plot_path_fat(path, quantity=None, Eshift=0, Emin=-10, Emax=50, iband=None, mode="fatband", fatfactor=20, cut_k=False, linecolor="red", close_fig=False, show_fig=False, label=f"WB" )
plt.ylim(-5, 12)#plt.ylim(3, 5.6) # Zoomplt.hlines([3.0, 5.6], 0, A[-1,0], linestyles="dashed", label="frozen window", color="black")plt.hlines(3.8253, 0, A[-1,0], linestyles="dashed", label="Fermi level", color="green")plt.legend()plt.savefig("csi_bands.pdf")----------------------------------------------------------------
We obtain the following results with a total of 4 WFs (Sn-p and Sn-s):----------------------------------------------------------------------------------------------------wannier centers and spreads---------------------------------------------------------------------------------------------------- 3.141041039293 3.141041039293 3.141041039293 | 6.652125391906 3.141041039293 3.141041039293 3.141041039293 | 7.022677077242 3.141041039293 3.141041039293 3.141041039293 | 7.022677077242 3.141041039293 3.141041039293 3.141041039293 | 7.022677077242
Here is an example for CsSnI3. This material is studied by Dr. Ao Wang at UCLouvain.
While it is relatively easy to Wannierize the bands close to the Fermi level (for e.g. transport calculations)using 9 WFs (I-p) for valence and 3 WFs (Sn-p), it is difficult to use less Wannier functions and keeping the symmetries. Since the cost of EPW calculations scale at least square with the number of WF, here we try to show how to decrease the number of WFs.
To do this, the python WannierBerri code is useful as you can easily do symmetry adapted Wannier functions. Note that these are not MLWFs.
During the Wannier Developer workshop 2026 in Trieste, Dr. Stepan Tsirkin implemented the capability to convert ukk and amn files from WannierBerri to EPW compatible (branch wrtie_epw). If you read this in the future, this conversion is probably part of the main WannierBerri code.
Note that WannierBerri is a python code. While it can be parallelize with Ray, it is mostly usefull when computing properties but not when doing Wannerization.Therefore, for now, it is best to run in sequential.
Due to heavy IO, I recommend copying the QE scf.in output files to your desktop and run WannierBerri locally. It is much faster than running on an HPC in most cases.
Important: the example below use k-point symmetry (so only the IBZ) with scf.in just to be faster. However, at present, to convert to EPW format, you need to use full BZ, so using nscf.in. This is not a big limitation since EPW requires this as well.
I used PBE pseudo from PseudoDojo v0.5.
Here is the QE scf.in (note the nbnd)----------------------------------------------------------------&control calculation='scf', restart_mode='from_scratch', prefix='csi', nstep = 1000 wf_collect = .true. pseudo_dir = './' outdir='./' / &system ibrav= 1, celldm(1) = 11.871414629, nat= 5, ntyp= 3, ecutwfc = 80.0 occupations = 'fixed' nbnd = 40 / &electrons electron_maxstep = 200 diagonalization='david' mixing_mode = 'plain' mixing_beta = 0.7 conv_thr = 1.0d-11 /ATOMIC_SPECIES Cs 132.9 Cs.upf Sn 118.7 Sn.upf I 126.9 I.upfATOMIC_POSITIONS (crystal)Cs 0.0000000000 0.0000000000 0.0000000000Sn 0.5000000000 0.5000000000 0.5000000000I 0.5000000000 0.0000000000 0.5000000000I 0.5000000000 0.5000000000 -0.0000000000I 0.0000000000 0.5000000000 0.5000000000K_POINTS {automatic} 12 12 12 0 0 0----------------------------------------------------------------
Then run a standard QE bands to obtain the csi.bands.gnu
Then run WannierBerri with the following python script
berri.py ----------------------------------------------------------------# Author: S. Ponc\'e - From 09/09/2026 to 11/09/2026# During Wannier workshop in Trieste. # With help from S. Tsirkin, J.-M. Lihm, and Claude Opus 5from wannierberri.symmetry.projections import Projection, ProjectionsSetfrom irrep.bandstructure import BandStructureimport wannierberri as wbimport pickleimport numpy as npimport matplotlibmatplotlib.use("Agg")import matplotlib.pyplot as pltimport scipyimport ray
ray.init(num_cpus=4)
path_data = "../"
try: bandstructure = pickle.load(open("bandstructure.pkl","rb"))except (FileNotFoundError, OSError, EOFError) as err: print(f"Error : {err}") bandstructure = BandStructure.from_espresso( prefix=path_data + "/csi", IBstart=12, # skip the 12 semicore bands #Ecut=680, normalize=True, include_TR=True, irreducible=True, ) bandstructure.pickle("bandstructure.pkl")
try: w90data = wb.w90files.WannierData.from_npz("./csi_wandata", files=["mmn", "eig", "chk", "symmetrizer" ], ignore_missing_files=False)except (FileNotFoundError, OSError, EOFError) as err: print(f"Error : {err}") w90data = wb.w90files.WannierData.from_bandstructure(bandstructure, seedname="./csi_wandata", files=["mmn", "eig", "chk", "symmetrizer" ]) w90data.to_npz("./csi_wandata")
spacegroup = w90data.symmetrizer.spacegroupspacegroup.show()
projection_s = Projection(orbital='s', position_num=[0.5, 0.5, 0.5], spacegroup=spacegroup)projection_p = Projection(orbital='p', position_num=[0.5, 0.5, 0.5], spacegroup=spacegroup)
projections_set = ProjectionsSet(projections=[projection_s, projection_p])
w90data.set_projections(projections_set, bandstructure)
w90data.wannierise(init="amn", froz_min=3.0, # above the VBM (3.825 eV) froz_max=5.6, # below band 25 (7.020 eV) outer_min=-4.7, outer_max=8.9, print_progress_every=10, num_iter=201, conv_tol=1e-10, mix_ratio_z=1.0, sitesym=True, parallel=False )
system = wb.system.System_R.from_wannierdata(wandata=w90data, berry=True)path=wb.Path.from_nodes(system, nodes=[[0.000000, 0.000000 , 0.000000], [0.000000, 0.500000 , 0.000000], [0.500000, 0.500000 , 0.000000], [0.000000, 0.000000 , 0.000000], [0.500000, 0.500000 , 0.500000], [0.500000, 0.500000 , 0.000000], ], labels=["G","X","M","G","R","M"], length=200 ) # length [ Ang] ~= 2*pi/dk
bands_path=wb.evaluate_k_path(system, path=path)
A = np.loadtxt("../csi.bands.gnu")bohr_ang = scipy.constants.physical_constants['Bohr radius'][0] / 1e-10alatt = 11.871414629* bohr_angA[:,0]*= 2*np.pi/alattA[:,1] = A[:,1]plt.scatter(A[:,0], A[:,1], c="black", s=4)#plt.plot(A[:,0], A[:,1], c="black")
bands_path.plot_path_fat(path, quantity=None, Eshift=0, Emin=-10, Emax=50, iband=None, mode="fatband", fatfactor=20, cut_k=False, linecolor="red", close_fig=False, show_fig=False, label=f"WB" )
plt.ylim(-5, 12)#plt.ylim(3, 5.6) # Zoomplt.hlines([3.0, 5.6], 0, A[-1,0], linestyles="dashed", label="frozen window", color="black")plt.hlines(3.8253, 0, A[-1,0], linestyles="dashed", label="Fermi level", color="green")plt.legend()plt.savefig("csi_bands.pdf")----------------------------------------------------------------
We obtain the following results with a total of 4 WFs (Sn-p and Sn-s):----------------------------------------------------------------------------------------------------wannier centers and spreads---------------------------------------------------------------------------------------------------- 3.141041039293 3.141041039293 3.141041039293 | 6.652125391906 3.141041039293 3.141041039293 3.141041039293 | 7.022677077242 3.141041039293 3.141041039293 3.141041039293 | 7.022677077242 3.141041039293 3.141041039293 3.141041039293 | 7.022677077242
While the valence band does not reproduce the full valence band maximum, it is perfect for the relevant energy region.
Then you can convert the results for EPW using ----------------------------------------------------------------# S. Ponc\'e - 10/09/2026import numpy as npfrom wannierberri.w90files.chk import CheckPointfrom wannierberri.w90files.mmn import MMN
chk = CheckPoint.from_npz("csi_wandata.chk.npz")mmn = MMN.from_npz("csi_wandata.mmn.npz")
chk.write_epw_ukk("csi.ukk")mmn.write_epw("csi.mmn")----------------------------------------------------------------
Update: 07/10/2026
Dr. Junfeng Qiao has tried a similar approach with Wannier.jl
He performed an analysis using Claude Opus 5.5.
The main conclusion is that 4 WFs can be achieved but require a 12x12x12 k-point grid while 12 WFs only requires 6x6x6 k-point grids. It is therefore a tradeoff.
In case where a relatively dense k-point grid is needed (due to the convergence of other properties for example), having less WFs is therefore interesting.
Here is the analysis:
Page updated
Google Sites
Report abuse