dft-workflows

Overview

DFT is the workhorse of computational materials science: approximate quantum mechanics accurate enough for structures, energetics, and electronic properties of most solids. This skill covers building a reliable DFT workflow — functional choice, convergence discipline, and the standard calculations (relaxation, DOS, bands, defects, surfaces) with the checks that separate trustworthy results from expensive noise.

When to use

• Predicting or rationalizing crystal structures, phase stability, and reaction energies
• Computing electronic structure: band gaps, DOS, charge transfer, bonding analysis
• Studying defects, dopants, surfaces, and interfaces via supercell models
• Screening materials (stability, band alignment, adsorption energies) before experiments
• Setting up high-throughput calculations with consistent, documented settings

Core concepts

• Functionals and their biases: LDA/GGA (PBE) underbind and underestimate gaps; hybrids (HSE06) fix gaps at ~10× cost; +U corrects localized d/f electrons; vdW corrections (D3, rVV10) matter for layered and molecular crystals. No functional is universally best.

• Basis and k-points: plane-wave cutoff (ENCUT) and k-mesh density are convergence parameters, not truths — converge them per system. Metals need denser meshes than insulators.

• Energy above hull: the formation energy relative to the convex hull of competing phases; < ~25 meV/atom is often considered (meta)stable and synthesizable.

• Band gap problem: standard GGA gaps are systematically too small; compare trends across materials rather than absolute values unless using hybrids or GW.

• Charged defects: formation energies depend on the Fermi level and chemical potentials; finite-size corrections (Freysoldt/FNV) are mandatory for charged supercells.

• Thermodynamics from DFT: 0 K energies need vibrational (phonon) and configurational entropy corrections for finite-temperature phase diagrams.

• Pseudopotentials/PAW: core electrons are replaced by effective potentials — verify the valence configuration includes semicore states for transition metals, which affect energetics noticeably.

• Smearing: Fermi-Dirac or Gaussian smearing aids SCF convergence for metals, but smearing entropy contaminates free energies — extrapolate σ→0 or use the tetrahedron method for final energies.

• Hubbard U: DFT+U corrects self-interaction for localized d/f states — U is material-specific, not universal; determine it by linear response or benchmark fitting, and always report the value used.

Practical workflow

1. Converge before you compute

# Typical convergence test: vary ENCUT and k-mesh, watch total energy per atom
# Converge to < 1-5 meV/atom before any production run

1. Fix the structure; scan ENCUT in steps, then k-mesh density — converge each to your target tolerance.
2. Choose the functional for the property: PBE(+U/+D3) for structures and trends; HSE06/GW for gaps and band alignment.
3. Document the full setting set (functional, cutoff, k-mesh, smearing, convergence criteria) — reproducibility requires all of it.

2. Standard calculation sequence

1. Relaxation: optimize lattice + ions (ISIF=3 in VASP terms) until forces < 0.01 eV/Å.
2. Static run: single-point at relaxed geometry with a denser k-mesh for accurate energies and DOS.
3. Bands: non-self-consistent run along high-symmetry k-paths (use standard paths for the lattice type).
4. Phonons: finite-displacement or DFPT to confirm dynamical stability (no imaginary modes) and get vibrational free energy.

3. Defects and surfaces

1. Build a supercell; relax; check that defect-defect distance exceeds ~10 Å (test size convergence).
2. Compute formation energies across charge states and Fermi-level positions; apply charge corrections.
3. For surfaces: use symmetric slabs or dipole corrections; converge slab thickness and vacuum gap (~15 Å).

4. High-throughput discipline

1. Lock one setting set for the whole campaign; never compare energies across different functionals or cutoffs.
2. Automate with a workflow manager (AiiDA, atomate, ASE) and store full provenance.
3. Validate a subset against experiment or higher-level theory before trusting the screen.

5. Run ab initio molecular dynamics when statics are not enough

1. Use AIMD for finite-temperature structures, diffusion, or phase transitions — equilibrate in NVT, then collect statistics in NVE or NVT.
2. Use ~1 fs timesteps (0.5 fs with hydrogen); check energy drift in NVE — drift means the timestep is too large or SCF convergence too loose.
3. Extract radial distribution functions, mean-square displacements (diffusion coefficients), and time-averaged structures — extend the run until these stabilize.

6. Quick-reference checklist

• ENCUT and k-mesh converged to < 5 meV/atom for this system
• Functional chosen for the property (structures vs gaps vs barriers)
• Spin polarization enabled with sensible initial moments where relevant
• Solvent/dispersion corrections applied where physically needed
• Forces < 0.01 eV/Å on the relaxed structure
• Charged-defect corrections (FNV/Freysoldt) applied
• Phonons checked for imaginary modes (dynamical stability)
• Full settings (functional, cutoff, k-mesh, U values) recorded for reproducibility

Common pitfalls

• Comparing unconverged or mixed-setting energies: a 50 meV/atom error swamps most phase-stability conclusions.
• Forgetting spin: magnetic systems need spin-polarized calculations with sensible initial moments; non-magnetic defaults give wrong ground states.
• Trusting GGA band gaps: always caveat them; use hybrids or GW when the gap value itself matters.
• Charged defects without corrections: raw supercell energies for charged defects are meaningless — apply FNV/Freysoldt corrections.
• Ignoring zero-point and thermal effects: 0 K DFT energies can invert the true finite-temperature stability ordering.
• Overinterpreting small differences: energy differences below your convergence tolerance are noise, not physics.
• Metal settings on insulators and vice versa: insufficient k-point density or wrong smearing for metals gives noisy, unreliable energetics — match the electronic settings to the electronic structure.
• Missing dispersion in soft matter: PBE without vdW corrections gets interlayer and intermolecular distances badly wrong — add D3 or rVV10 where non-covalent binding matters.