Moderators: Global Moderator, Moderator
Dear All,
During ionic optimization with IBRION = 2, we noticed very small increases in the total energy at some ionic steps.
Should we always expect the total energy to decrease (i.e., move downhill) at every ionic step?
Many thanks in advance,
Ibrahim
Hello Ibrahim,
Structure optimization via IBRION=2 performs a conjugate-gradient optimization. While I am not familiar with all implementation details of this particular algorithm in VASP it is generally not guaranteed that every conjugate-gradient step will always decrease the energy. At a given position (i.e. box geometry and atom coordinates) the algorithm will determine the updated position by using a search direction dependent on the current forces and information from previous steps. In this search direction the conjugate-gradient algorithm expects the potential energy to decrease. However, it cannot know that this is actually the case because the potential energy surface is unknown and could contain many little minima and maxima. So it is possible, perfectly normal and not unlikely that small upward deviations in potential energy happen during an overall descending structure optimization run. How often this occurs is influenced by the step size POTIM and how well the forces are converged ("noise" in the potential energy landscape).
All the best,
Andreas Singraber
Dear Andreas Singraber,
Thanks for the detailed reply.
Can you please also clarify why we see that electronic minimization algorithms in VASP (the blocked Davidson (ALGO = Normal) and RMM-DIIS algorithms) do not guarantee monotonic downhill convergence of the electronic energy, with energy increases occurring at some electronic steps?
Many thanks in advance,
Ibrahim
Dear Ibrahim,
sorry for the late reply! The exact reason for "uphill" movement during the electronic minimization is likely very dependent on the actual case. But in general neither the Davidson, nor the RMM-DIIS algorithm guarantee a monotonic downhill convergence. In contrast to the conjugate-gradient method they do not follow the gradient of the energy but rather try to minimize a residual. For details you may need to study the theory behind those two algorithms (RMM-DIIS and Blocked-Davidson). However, to short-cut directly to the answer of your question: Yes, it is perfectly normal to observe smaller uphill movements (bumps) while converging to the desired ground-state energy.
All the best,
Andreas Singraber