<div dir="ltr">Hi, I've done additional tests to try to figure out what may be going wrong. It seems that, for the open-shell systems I'm interested in, the performance of the chosen preconditioner is critical. Here's what I understand: the system is in a region of his phase space wherein after some MD steps, the guess wavefunction obtained by the ASPC method starts to depart from the BO surface and the use of the computationally cheap preconditioners (full_kinetic, full_single_inverse) do not help speed up the SCF convergence. Regularly resetting the extrapolation method by restarting the simulation is a workaround. But, besides being a rather tedious approach, I don't think there is an ideal fixed number of steps after which to restart the simulation: the issue always tends to show up again, later. With the full_all preconditioner, the simulation run rather smoothly ! As expected, the conserved quantity tend to be pretty well conserved over 1.5ps+, when using a timestep of 0.5fs or even of 1fs. The huge problem now is that a MD step is taking 80s in average, with ~80% of the time dedicated to preconditioning. Is there a way to decrease the time spent in preconditioning ? I'm using currently cp2k-2.5.1, the mixed_precision algorithm cannot be combined with the full_all preconditioner. Is there an other way to do this? Are there changes made to preconditioning in the development version? Best regards, Max <div dir="ltr"> </div></div>