SCF Convergence Suggestions

Generally, Hartree-Fock wave functions for simple organic molecules converge in fewer than 10 iterations, while complex calculations involving higher-level methods or open shells may take a few extra iterations. Molecules which include transition metals generally converge more slowly, however. Make sure your job has really converged and did not simply end because it reached the maximum number of SCF iterations, a number set in the SCF tab.

If a job gives poor SCF convergence, you can try either modifying the convergence methods used or improving the initial guess. To modify the convergence methods, try any or all of the following settings:

  • Set nofail=1 in the gen section to apply automatic recovery measures when bad SCF convergence is detected. See Automatic SCF recovery with nofail below for details.

  • Change the Accuracy level setting in the SCF tab to Ultrafine. This corresponds to the keyword setting iacc=1, which causes the job to use denser pseudospectral grids and tighter cutoffs, and generally increases computational costs by a factor of two to three. Note, however, that as of Jaguar v7.0 this action is no longer necessarily because if Jaguar detects that the SCF is not converging, it automatically sets iacc=1.

  • Select GVB-DIIS from the Convergence scheme option menu in the SCF tab. Generally, DIIS is the better choice, but the GVB-DIIS convergence scheme sometimes leads to convergence when DIIS does not. The GVB-DIIS scheme works by ensuring that each new set of occupied orbitals has maximum overlap with the previous set. The advantage of this is that oscillations due to frequently changing occupations are damped. The disadvantage is that if the initial guess occupies the wrong orbitals, the SCF can converge to an excited state.

  • Set the SCF level shift in the SCF tab to 0.5 or 1.0. The higher the setting, the more the energies of the virtual orbitals are increased before diagonalization, and the more the mixing of the real and virtual orbitals is reduced. High SCF level shifts can slow convergence by several iterations, but can often help otherwise intractable cases to converge. Because jobs with SCF level shifts are slightly more likely to converge to excited states, you may also want to restart these jobs without any SCF level shift. As of Jaguar 7.0, a level shift of 1.0 is used automatically if Jaguar detects that the SCF is not converging.

  • If the calculation is a DFT job, use finer DFT grids. You can adjust this setting from the Grid density option menu in the Theory tab. This setting also increases the computational cost.

  • As a last resort, try setting iacscf to 1 or 4 (see Table 1 for descriptions of these settings). This keyword can help when pseudospectral error is particularly high, but it works by removing more eigenvectors from the overlap matrix. This effectively reduces the size of the basis set, which means that if you want to compare energies for two different molecules, you need to use the same value for iacscf. You might need to increase the setting of maxit to 100 or more when using this keyword.

Automatic SCF recovery with nofail

The nofail=1 setting applies automatic recovery measures when SCF convergence fails. The recovery strategy depends on the nature of the failure.

If the SCF runs out of iterations but the energy has been decreasing steadily, the procedure simply increases maxit. If the SCF is not converging smoothly, the procedure applies more aggressive measures. These include restarting in a reduced basis set (by raising the linear dependency threshold cut20), applying a larger level shift (vshift), and switching the convergence scheme (iconv). If the SCF converges in the reduced basis, the calculation restarts with the original basis using the converged wave function as an initial guess.

If the energy oscillates or the DIIS error is large, the problem is more likely related to pseudospectral error. In this case, the procedure switches to higher-accuracy grids and tighter cutoffs (iacc=1), and may add diffuse basis functions. If these measures are insufficient, the procedure reduces the basis set and may switch to fully analytic integrals (nops=1) to avoid pseudospectral error entirely. A single SCF may encounter both types of failure and switch between recovery strategies as convergence behavior changes.

If convergence cannot be achieved in the original basis or with the original nops setting, nofail=1 allows the calculation to proceed with a reduced basis or nops=1 rather than failing. This can affect the quality of the resulting energy and wave function, so you should verify that results obtained with nofail=1 are consistent with those from fully converged calculations when high accuracy is required.

With nofail=2, the same recovery strategies are attempted, but the calculation fails if convergence cannot be achieved with the original settings.