用dmol程序优化纳米线的几何结构，经过十几步的优化后，能量变化逐步减小，眼看就要优化好了，可是突然又会有一个很大的能量变化，然后又重新开始优化。如果允许它一直算下去，体系能量就会周期性的出现突然增大的现象（没有考察结构是不是也回到了初始状态），整个优化过程周而复始的进行，丝毫没有要收敛的迹象，请问有什么方法可以解决这个问题吗？谢谢！

[ Last edited by zjuer on 2008-5-27 at 12:56 ]

这一段来自dmol帮助文件，希望对你有用
Performance tips
DMol3 uses a localized numeric basis set and very efficient density fitting technique, making the code particularly suitable for performing accurate DFT calculations for extended molecular, solid, and surface systems. Better computational performance can be achieved using the following guidelines:

If the system has low symmetry, e.g., Cs or C2, it is better to perform calculations without symmetry (C1).
Use a small confinement on the atomic basis set by choosing Medium, or even Coarse, quality.
In the case of structure optimization, it may be better to maintain a Fine numerical integration grid, use Medium SCF tolerance, and use Coarse orbital cutoff quality, initially. After the structure is converged, only a few cycles may be needed to converge the structure with a better orbital cutoff quality.
Real space cutoff can be set as low as 2.5 Å for Coarse quality calculations, typically without a significant loss in total or binding energy. The performance gain is particularly significant for solid calculations with k-points. Using too small a cutoff value may cause a failure to converge.
Slow SCF convergence may become a bottleneck for systems that have degenerate levels close to the Fermi level. Typically, these are systems including metal atoms or systems in which dissociation processes are being studied. Using the smearing option with a small value for the smearing factor (0.001 Hartree) may be very beneficial here.
A potential way to improve convergence for coarse k-point sets without introducing thermal smearing is to switch off the tetrahedra integration algorithm with the defeat_tetrahedra keyword.
The performance of geometry optimization calculations may suffer if the system has near-linear bends. Eliminating these bends will typically shorten optimization cycles.
You should use sufficiently accurate integration grid and SCF tolerance to calculate accurate gradients for use in optimizations and TS searches.

谢谢楼上的兄弟，我试试看。

l.pl......p,.o,omuinyby

我通常都是先用“use smearing”优化，然后再将优化好的构型在去掉“use smearing”的情况下优化，还算幸运，构型一般都能收敛。