It's common for people to have trouble converging DFT calculations for correlated electron systems. A lot of times you have to try a few different tricks before one finally works.
- Check your initial guess; does it make sense chemically?
- Level shifting
- Smearing
- Converge minimal basis, then project to larger basis (easy in pyscf)
- Vary solver (dependent on package; broyden, newton, different DIIS techniques, ...)
- Converge Hartree-Fock, then use as input to DFT
- Increase number of k-points
- Remove approximate linear dependence if the basis overlap matrix has a large condition number
- Increase tolerances (TOLINTEG in CRYSTAL, direct_scf_tol in pyscf)
| Symptom | Common causes |
|---|---|
| Initial energy large and fluctuates wildly | Initial guess is poor; try improving it or using smearing to get a better starting point |
| Calculation almost converges but dE gets stuck at around 1e-5 | Tolerances can be too low, not enough k-points, damping is too large, approximate linear dependence, solver not performing well |