Skip to content

3.5 SCF, grids and basis: separate tests

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.

3.5.1 Choose an error budget

Define tolerable error for a conformer difference, bond length or force before settings. No one input is automatically converged for every property. Keep geometry, cell, charge/spin, functional and potentials fixed. Use a displaced geometry for force checks so near-zero minimum forces do not hide error.

3.5.2 Tighten the electronic solve first

At fixed basis/grid compare EPS_SCF=1E-6, 1E-7, 1E-8. EPS_SCF is an algorithmic stopping criterion, not a declared total-energy error in hartree. Measure energy/force differences. Hitting MAX_SCF only exhausts allowance; keep failure detection enabled. SCF reference.

The simple occupied-orbital OT setup is not a universal metal recipe and must not be casually combined with fractional occupations. For metals/small gaps, diagonalization with mixing, extra orbitals and suitable smearing is a common branch, with k sampling where needed. Advanced OT versions add capabilities, so do not claim every implementation forbids metals. OT reference.

3.5.3 Converge both grid controls

CUTOFF governs the finest grid; REL_CUTOFF affects assignment of Gaussian products among multigrid levels. Improving one can leave the other limiting. Grid errors need not decrease monotonically: lowest energy in a scan is not the decision criterion. Official cutoff tutorial.

  1. At adequate EPS_SCF and REL_CUTOFF=80 Ry, scan CUTOFF=300,400,500,650,800 Ry.
  2. At a provisional cutoff scan REL_CUTOFF=40,60,80,100 Ry.
  3. Try a stronger pair, e.g. 1000/120 Ry, as provisional reference; extend if observable still changes materially.
  4. Compare identical force components/energy differences, then another relevant geometry before MD/reactions.
  5. Leave margin beyond accidental agreement; measure time only for correctly converged cases.

These are a proposed lab plan, not obtained convergence. Use independent c0500-r0080-style directories and CSV:

cutoff_Ry,rel_cutoff_Ry,eps_scf,energy_Ha,max_force_change_Ha_per_bohr,wall_s,converged

Enter observed values only, leaving failed/missing runs marked, not zero.

3.5.4 Enlarge Gaussian basis separately

At tested grids compare available larger GTH-compatible TZVP/TZV2P MOLOPT entries for all elements, verifying names while keeping potentials/valence partition fixed. Recheck grids after basis changes. Evaluate the target observable, not just raw absolute energy. Noncovalent energies introduce basis superposition error and functional/dispersion suitability. Converged PBE is not universally accurate; changing dispersion to match experiment changes the model and must be declared.

3.5.5 Exercise

Produce an error budget separating electronic, grid and orbital-basis sensitivity. State acceptance rule, provisional reference and another-geometry test. No fabricated convergence table substitutes for this evidence.

Previous：Complete isolated-water calculation · Next：Geometry optimization and boundary conditions