Mat Edi Mobility

mat-edi-mobility

Goal

To compute the point-defect-limited carrier mobility $\mu(T)$ of a semiconductor and the underlying electron-defect scattering matrix elements $M(\mathbf{k}i, \mathbf{k}f) = \langle \psi{\mathbf{k}i} | \Delta V | \psi{\mathbf{k}f} \rangle$ from first principles, using the supercell difference-potential method of the Quantum ESPRESSO + EDI pipeline. Here $\Delta V = V{\text{defect}} - V{\text{pristine}}$ is the Kohn-Sham perturbation potential of an isolated point defect, extracted as the difference between the potentials of a defect-containing supercell and a pristine supercell. The matrix elements are computed on a coarse k-grid, transformed to a maximally localized Wannier basis $M(\mathbf{R}, \mathbf{R}')$, and interpolated onto dense fine grids. The state-resolved scattering rate follows from Fermi's golden rule,

$$\frac{1}{\tau_{n\mathbf{k}}} = \frac{2\pi}{\hbar}, n_{\text{d}}, \frac{1}{N_{\mathbf{k}}} \sum_{m,\mathbf{k}'} |M_{n\mathbf{k}, m\mathbf{k}'}|^2, \delta(\varepsilon_{n\mathbf{k}} - \varepsilon_{m\mathbf{k}'}),$$

and the mobility from the linearized Boltzmann transport equation in the self-energy (SERTA) and momentum (MRTA) relaxation-time approximations. The three first-class outputs are the direct matrix elements $M$ on the coarse grid, the Wannier-interpolated $M$ on a fine k-path, and the final mobility $\mu(T)$.

Background

Defect-limited transport requires $M(\mathbf{k}i, \mathbf{k}f)$ on k-grids far denser than any tractable supercell calculation can supply. EDI solves this the way EPW solves the electron-phonon problem: compute $M$ directly on a coarse grid commensurate with the supercell, rotate into a Wannier basis where $M(\mathbf{R}, \mathbf{R}')$ decays rapidly with $|\mathbf{R}|$ and $|\mathbf{R}'|$, and interpolate to $300\times300$ or denser at negligible cost. The scattering rate scales linearly with the defect concentration $n{\text{d}}$ (dilute, incoherent-defect limit), so **the mobility scales as $1/n{\text{d}}$ and the concentration must always be reported alongside $\mu$**. For 2D materials the pristine and defect supercells are computed independently, so their electrostatic potentials carry an arbitrary offset that must be removed by vacuum-level alignment (pot_align = 'vacuum').

This skill is the defect-scattering companion to mat-epw-mobility, which computes the phonon-limited mobility. At finite temperature the two channels combine by Matthiessen's rule, $1/\mu_{\text{total}} = 1/\mu_{\text{phonon}} + 1/\mu_{\text{defect}}$. Defect energetics (formation energies, charge states) that set which defects dominate come from mat-defect-energy; the AMSET route in mat-dft-electronic-transport is an alternative phonon-limited channel.

External binary: building Quantum ESPRESSO 7.5 + EDI

EDI is an in-tree plugin of Quantum ESPRESSO, not a standalone code: it links QE's compiled static libraries and must be built against a matching QE version.

1. Obtain QE 7.5 (GitLab tag archive; no ReleasePack is required — make w90 auto-fetches the pinned Wannier90 submodule even from a tarball) and configure it.

2. Clone EDI into the QE root and build:

   # Requires: Quantum ESPRESSO 7.5 + EDI (external MPI build)
   cd <QE-root>
   make pw
   make w90
   git clone https://github.com/yuanyue-liu-group/EDI edi-code
   cd edi-code && make
   

   This produces src/edi.x (MPI) and src/extract_pot.x (serial).

3. gfortran patch (required for GCC builds): ed_coarse.f90 contains lines up to 158 characters, so add long-line support to edi-code/src/makefile after the include line:

   F90FLAGS += -ffree-line-length-none
   FFLAGS   += -ffree-line-length-none
   

4. extract_pot.x segfault patch (required): extract_pot.f90 calls clean_pw(.TRUE.) at line 77, before any read_file, which segfaults on a fresh process. Comment it out (or delete that line) and rebuild:

   ! CALL clean_pw(.TRUE.)   ! upstream bug: runs before any read_file, segfaults
   

Pseudopotentials for the worked example are PseudoDojo NC-SR v0.5 PBE stringent (https://www.pseudo-dojo.org/pseudos/nc-sr-05_pbe_stringent_upf.tgz); the Mo and S UPFs are used as Mo.upf and S.upf. QE 7.5 must be built with Wannier90.

Instructions

[!IMPORTANT] Quantum ESPRESSO (pw.x) and EDI (edi.x, extract_pot.x) are external compiled MPI binaries, exactly as VASP is in mat-dft-vasp. This skill targets a source build of QE 7.5 + EDI v2.0 (repo main @ 41fed72) with PseudoDojo NC-SR v0.5 PBE stringent pseudopotentials. The # Env: base-agent annotations apply only to the Python parser scripts. Reference input decks for every step are in resources/inputs/. Replace <ranks> with your MPI rank count; mpirun -np and srun -n are interchangeable. Set pseudo_dir in each deck to your UPF directory (the decks use ./pseudo/).

The pipeline is a DAG. The two supercell SCFs (step 3) are independent of the primitive cell and of each other and can run in parallel. Steps 5 and 6 must run in order (6 reads the Wannier archive written by 5).

1. Primitive-cell SCF

Compute the ground-state charge density of the pristine primitive cell. assume_isolated = '2D' truncates the out-of-plane Coulomb tail and must be consistent across every calculation. Deck: resources/inputs/scf_primitive.in.

# Requires: Quantum ESPRESSO 7.5 + EDI (external MPI build)
mpirun -np <ranks> pw.x -nk <pools> < scf_primitive.in > scf_primitive.out

Confirm convergence has been achieved.

2. Primitive-cell NSCF on the explicit coarse grid

Produce the Bloch states EDI folds into Wannier functions. The grid must be an explicit K_POINTS crystal list (never automatic) and its dimensions must equal coarse_nk1/nk2/nk3 in the EDI input (12x12x1 here). Generate the list with the helper, then splice it into the NSCF deck (resources/inputs/nscf_primitive.in already contains the 144-point 12x12x1 card):

# Env: base-agent
python .agents/skills/mat-edi-mobility/scripts/gen_kgrid.py 12 12 1

# Requires: Quantum ESPRESSO 7.5 + EDI (external MPI build)
mpirun -np <ranks> pw.x -nk <pools> < nscf_primitive.in > nscf_primitive.out

nbnd (17 here) must cover the disentanglement window (the 5 kept bands 13-17). Note: the upstream example ships an 18x18x1 NSCF that is inconsistent with its coarse_nk1 = 12; this deck fixes that to 12x12x1.

3. Pristine and defect supercell SCFs

Compute the Kohn-Sham potentials whose difference is $\Delta V$. Both are Gamma-only SCFs of a 6x6x1 supercell (108 atoms pristine; 107 with one S vacancy). The supercell lattice vectors must be exact integer multiples of the primitive cell and host-atom positions must fold back to the primitive sites, or $\Delta V$ is corrupted. Relax the defect supercell before this final SCF. Decks: resources/inputs/scf_pristine_super.in, resources/inputs/scf_defect_super.in.

# Requires: Quantum ESPRESSO 7.5 + EDI (external MPI build)
mpirun -np <ranks> pw.x < scf_pristine_super.in > scf_pristine_super.out
mpirun -np <ranks> pw.x < scf_defect_super.in > scf_defect_super.out

4. Extract the difference potential

Run the serial extract_pot.x to write the pristine and defect local KS potentials as cube files (V_p.cube, V_d.cube) on the supercell real-space grid. Deck: resources/inputs/extract_pot.in.

# Requires: Quantum ESPRESSO 7.5 + EDI (external MPI build)
mpirun -np 1 extract_pot.x < extract_pot.in > extract_pot.out

extract_pot.x is serial; run it with a single rank.

5. First EDI pass: direct matrix elements + Wannierization

Compute $M(\mathbf{k}_i, \mathbf{k}_f)$ directly on the coarse 12x12x1 grid, build $M(\mathbf{R}, \mathbf{R}')$, run the Wannier90 minimization, and write the Wannier archive mos2_edmatw_2d.bin that the transport pass reuses. Deck: resources/inputs/edi_setup.in (edwread = .false., wannierize = .true., do_transport = .false.).

# Requires: Quantum ESPRESSO 7.5 + EDI (external MPI build)
mpirun -np <ranks> edi.x -nk <ranks> -i edi_setup.in > edi_setup.out

Key namelist variables:

• nbndsub = 5, proj(1) = 'Mo:d', bands_skipped = 'exclude_bands = 1-12' — the 5-band Mo-d Wannier manifold; material-specific.
• dis_win_min/max, dis_froz_min/max — outer and frozen disentanglement windows (eV); set them around the transport bands.
• wdata(1) = 'guiding_centres = .true.' — required in long-vacuum 2D cells or the Wannier centres fold to a wrong z-image.
• coarse_nk* must equal the step-2 NSCF grid; fine_nk* sets the interpolation grid.
• edmat_interp_from_file = .true. with filki_interp/filkf_interp also writes the interpolated $M$ along the validation path (mos2_edmat_interp.dat), used in 6b.

Because it also computes the coarse-grid Bloch $M$, this pass writes the direct matrix elements to mos2_edmat_bloch.dat (columns: iki ikf kix kiy kiz kfx kfy kfz ibnd jbnd |M|^2 Re(M) Im(M) |M_loc|^2 |M_nl|^2).

-nk <ranks> (pools = rank count) is recommended for EDI performance.

6. Second EDI pass: fine-grid interpolation + BTE transport

Interpolate $M$ onto the fine grid and solve the BTE for $\mu(T)$. This pass reuses the Wannier archive (edwread = .true., wannierize = .false., do_transport = .true.). Deck: resources/inputs/edi_transport.in.

# Requires: Quantum ESPRESSO 7.5 + EDI (external MPI build)
mpirun -np <ranks> edi.x -nk <ranks> -i edi_transport.in > edi_transport.out

Key transport variables:

• fine_nk1 = fine_nk2 = 300 — production grid; 48 is a fast smoke test.
• transport_win_min/max — narrow energy window (eV) around the carrier band edge; only pocket states scatter.
• carrier_conc (cm^-2 in 2D) sets the Fermi level by bisection; defect_conc (cm^-2 in 2D) scales the rate. Mobility scales as $1/$defect_conc — report it.
• delta_method = 'gaussian', delta_sigma = 0.01 — energy-conserving delta; also 'triangular' (2D optimized) and 'adaptive' (velocity-dependent smearing).
• temps(i), nstemp — temperature list.

Outputs: mos2_transport.dat (T mu_SERTA_xx mu_MRTA_xx mu_SERTA_yy mu_MRTA_yy in cm^2/Vs) and mos2_inv_tau.dat (state-resolved ik ibnd E inv_tau_SERTA inv_tau_MRTA tau_SERTA tau_MRTA). Parse them:

# Env: base-agent
python .agents/skills/mat-edi-mobility/scripts/parse_transport.py mos2_transport.dat --output-dir results/
python .agents/skills/mat-edi-mobility/scripts/parse_inv_tau.py mos2_inv_tau.dat --output-dir results/

Fine-grid convergence is cheap to sweep: because edwread = .true. reuses mos2_edmatw_2d.bin, each rerun redoes only the interpolation + BTE (seconds to a few minutes, e.g. ~6 s at 48x48 up to ~3 min at 300x300 for this example). The outputs are prefix-named (mos2_transport.dat), so copy them to a per-grid name between reruns. Converge fine_nk* before trusting $\mu$: a 48x48 smoke grid runs 21-23% low here.

6b. Validate: direct vs Wannier-interpolated M on a k-path

Confirm the Wannier interpolation reproduces the directly computed $M$ (this is the interpolation's correctness gate, analogous to the DFPT-vs-EPW $|g|$ benchmark in mat-epw-mobility). Both are evaluated on the SAME k-path: one initial point at K (resources/inputs/ki.dat) and a 300-point Gamma-K-M-Gamma final path (resources/inputs/kf.dat). Deck: resources/inputs/edi_direct.in, which enables both edmat_interp_from_file and edmat_direct_from_file on the same path.

Direct mode needs NSCF wavefunctions at every path k-point — the coarse 12x12x1 NSCF only contains 8 of the 300 path points. Run it in two steps:

1. Launch the direct pass once. It detects the missing k-points, auto-writes nscf_custom.in listing exactly the ki + kf points it needs, and aborts.
2. Run that NSCF into a separate save tree (e.g. primitive_path/dout/; 1.5 min on 32 tasks for the 301-point path here), point edi_outdir at it (already set to ../primitive_path/dout/ in the deck), and rerun the direct pass (3m40s at 32 ranks x 4 cores for this example — direct mode is memory-heavier; see Constraints).

# Requires: Quantum ESPRESSO 7.5 + EDI (external MPI build)
# 1. First attempt writes nscf_custom.in and stops:
mpirun -np <ranks> edi.x -nk <ranks> -i edi_direct.in > edi_direct.out
# 2. Run the path NSCF into primitive_path/, then rerun the direct pass:
mpirun -np <ranks> pw.x < nscf_custom.in > nscf_custom.out    # stage into primitive_path/dout/
mpirun -np <ranks> edi.x -nk <ranks> -i edi_direct.in > edi_direct.out

This writes mos2_edmat_direct.dat (direct $M$, absolute NSCF band indices) and mos2_edmat_interp.dat (interpolated $M$, Wannier-subspace indices 1-5). Compare:

# Env: base-agent
python .agents/skills/mat-edi-mobility/scripts/parse_edmat.py mos2_edmat_interp.dat --output-dir results/
python .agents/skills/mat-edi-mobility/scripts/compare_edmat.py mos2_edmat_direct.dat mos2_edmat_interp.dat --band-offset 12 --band-sum --output-dir results/

--band-offset 12 maps the two files' band conventions (direct = interp + number of excluded bands). --band-sum compares the gauge-invariant band-summed $\sum_{mn} |M_{mn}|^2$ per k-point — the meaningful metric, since individual band pairs mix under the arbitrary Wannier gauge along band crossings (drop the flag for the pairwise statistics). A small relative RMS confirms the interpolation. Finally, benchmark $\mu(T)$ against the upstream reference with compare_reference.py (see the example).

Examples

See examples/mos2-s-vacancy/ for the full end-to-end run on a sulfur vacancy in monolayer MoS2, including the coarse-grid direct $M$, the Wannier-interpolated $M$ with the direct-vs-interpolated comparison, the mobility-vs-temperature table validated against the upstream reference outputs, and a literature cross-check against ACS Nano 18, 8511 (2024).

Constraints

• External codes: requires an MPI build of Quantum ESPRESSO 7.5 with Wannier90 and the EDI plugin built in-tree (edi.x, extract_pot.x); the Python parsers run in base-agent. GCC builds require -ffree-line-length-none (see Background). Version pins reflect what was targeted, not a claim that other versions are broken.
• In-tree build only: EDI links QE 7.5 static libraries; the QE version must match exactly. Our QE 7.4.1 build is not reusable.
• Point defects only: EDI v2.0 handles point defects (vacancies, substitutions, interstitials). Extended defects (surfaces, grain boundaries) are not yet supported.
• SERTA / MRTA, not iterative BTE: mobility is relaxation-time-based; there is no self-consistent iterative BTE solution.
• Mobility scales with defect concentration: $\mu \propto 1/$defect_conc in the dilute limit — always report the concentration alongside $\mu$.
• 2D systems need assume_isolated = '2D' (vacuum-separated cell, c = 24 A here) in every pw.x run and pot_align = 'vacuum' in EDI, consistently.
• Coarse-grid match: the primitive NSCF grid must equal coarse_nk1/nk2/nk3, and the NSCF must use an explicit K_POINTS crystal list, or the Wannier folding is inconsistent.
• Supercell consistency: supercell lattice vectors must be exact integer multiples of the primitive cell, host atoms must fold to primitive sites, and the FFT grid must be commensurate, or $\Delta V$ is corrupted.
• Memory scales with supercell volume x ranks-per-node: in the setup, transport, and direct modes every MPI rank holds the full supercell real-space potentials ($V_d$, $V_p$, $V_{\text{colin}}$) — about 2 GB/rank for this 6x6 MoS2 example (240x240x300 FFT grid), and direct mode holds an extra copy. On ~2 GB/core clusters, under-subscribe the node: run setup/transport at half ranks (--ntasks=64 --cpus-per-task=2, edi.x -nk 64) and direct mode at a quarter (--ntasks=32 --cpus-per-task=4, edi.x -nk 32). Symptom of getting it wrong: the job stays RUNNING with the output frozen right after the Full double-FT banner while ranks are silently OOM-killed and survivors deadlock in the next MPI collective. CPU burn does not prove progress — check output-file mtime and live task count.
• Material-specific inputs: ecutwfc, nbnd, nbndsub, projections, exclude_bands, and the disentanglement/transport windows are set for MoS2 and must be adapted to your material and band manifold.

References

• Z. Xiao, R. Guo, C. Zhang, Y. Liu, "Point Defect Limited Carrier Mobility in 2D Transition Metal Dichalcogenides", ACS Nano 18, 8511 (2024). DOI
• I.-T. Lu, J.-J. Zhou, M. Bernardi, "Efficient ab initio calculations of electron-defect scattering and defect-limited carrier mobility", Phys. Rev. Materials 3, 033804 (2019). DOI
• I.-T. Lu, J. Park, J.-J. Zhou, M. Bernardi, "Ab initio electron-defect interactions using Wannier functions", npj Comput. Mater. 6, 17 (2020). DOI
• A. A. Mostofi, et al., "An updated version of wannier90: A tool for obtaining maximally-localised Wannier functions", Comput. Phys. Commun. 185, 2309 (2014). DOI
• P. Giannozzi, et al., "Quantum ESPRESSO toward the exascale", J. Chem. Phys. 152, 154105 (2020). DOI
• The EDI code paper is in preparation; cite the repository: github.com/yuanyue-liu-group/EDI.

Author: Chenmu Zhang Contact: GitHub @cz2014