Skip to content

1.5 SCF, cutoff and k-mesh convergence

The inputs on this page are educational templates that have not been executed for this course. Use an authorized installation, verify version-specific options and establish your own convergence.

1.5.1 Three questions, not one

Electronic self-consistency asks whether the solver finds a stable density at fixed numerical settings. Basis convergence asks whether increasing ENCUT changes the answer. Sampling convergence asks whether increasing the Brillouin-zone mesh changes it. Passing the first does not establish the other two. Tightening EDIFF cannot repair a deficient basis or mesh. Electronic minimization; ENCUT.

Choose the observable before the tolerance. Possible teaching goals are an energy difference stable within 1 meV/atom, a maximum force stable within 0.005 eV/Å, or a lattice parameter stable within a declared length tolerance. These are proposed goals, not universal standards or achieved results. Phase comparisons require convergence of the difference between phases, not just one phase's energy; near degeneracy may require much tighter choices.

1.5.2 A controlled three-pass experiment

A — Electronic threshold. Keep geometry, potentials, mesh, cutoff and occupations fixed. Compare EDIFF values 1E-5, 1E-6 and 1E-7 eV. Verify that each run actually reaches its threshold and record the observable and iteration count. The EDIFF condition involves total/free-energy and band-energy changes; it is not a maximum-force criterion. Tighten further if the target answer still changes appreciably. EDIFF.

B — Plane-wave cutoff. Use a reasonably dense mesh and chosen electronic threshold. Let \(C\) be the largest ENMAX across species and consider \(C,1.15C,1.30C,1.45C\), rounded upward to convenient eV values. This is a proposed sequence, not a VASP requirement. A concrete 300/350/400/450/500 eV Si sequence is appropriate only where the chosen dataset supports these points. Keep the potential and geometry unchanged; otherwise basis and structural effects become entangled. ENCUT; POTCAR.

C — k-point mesh. At a suitably converged cutoff, try Gamma-centered \(4^3,6^3,8^3,10^3,12^3\) meshes. Keep SIGMA and energy definition fixed. These are trial points, not a promised stopping point. Recheck cutoff at the selected mesh, and test at least one denser mesh beyond the apparent plateau because errors interact. Record both the full input mesh and symmetry-reduced count from output. Irreducible counts alone are misleading across different symmetries. Supercells often need fewer subdivisions for comparable reciprocal resolution, but the property must still be tested. KPOINTS.

1.5.3 Record and decide

Leave results blank until computed:

run_id,encut_eV,kmesh,centering,sigma_eV,ediff_eV,energy_label,energy_eV_cell,atoms,max_force_eV_A,pressure_kbar,scf_converged

For two atoms, a trial/reference difference in meV/atom is

\[ \Delta E_{\mathrm{meV/atom}}=1000\frac{E_{\mathrm{trial}}-E_{\mathrm{reference}}}{2}. \]

Use the densest successful point as a provisional reference and identify it. A plateau against an unconverged reference is not proof. Plot values as well as differences: a single small adjacent difference may be accidental. Measure performance only after validating the calculations, so cost is optimized at a fixed accuracy.

Gaussian broadening also needs a check. Repeat selected points with a smaller SIGMA, such as 0.025 eV, and compare the same energy definition. Do not use high-order Methfessel–Paxton smearing by habit for an insulator. For metals, mesh and broadening should be tested together; this Si recipe is not a metal default. ISMEAR; Smearing technique.

A perfectly symmetric Si cell may have negligible forces even with poor settings. To test force accuracy, make a separate slightly displaced structure, use the exact same displacement for all points, and check intended symmetry handling. It is a diagnostic geometry, not a relaxed candidate.

1.5.4 Exercise and further study

Produce two convergence plots and explain the property, tolerance, reference, selected cutoff/mesh and extra verification point. If you cannot run, submit a blank protocol marked “planned, not executed”; do not invent a smooth numerical curve. The official bulk tutorial links sampling to total energies; the independent NSC 2023 convergence workshop supplies practical context. This exercise is an original protocol and does not reproduce their numerical data.


Previous:First bulk-silicon single point · Next:Atomic and cell relaxation