Problem: During geometry optimization in Z-Matrix coordinates, the atomic configuration makes a large unrealistic jump.
Possible cause 1: the triplet of atoms to which the current atom is related by the Z-matrix is (almost) co-linear. When, in a geometry step, the triplet passes through co-linearity, the dihedral angle for the current atom should make a discontinuous jump of 180 degrees. This is not checked in the program and the dihedral angle may not get corrected, resulting in a geometric jump of the atom (and hence of all atoms related to it by the Z-matrix).
Cure: Check the triplets of atoms, used in your Z-matrix to define the dihedral angles. If one of them is almost colinear, then redefine the Z-matrix or use Cartesian optimization.
Possible cause 2: the connectivity of the Z-matrix does not reflect some important bonds. Especially when the molecule contains (a large number of) rings, this badly affects the stability of the geometry update step. The reason is basically that computed Cartesian forces are transformed into changes of the curvilinear internal coordinates. The transformation between the two systems of coordinates is non-linear, but mathematically assumed to be linear. This is only a good approximation for small steps.
Cure: Redefine the Z-matrix and/or (if the geometry steps are very large) set a smaller upper bound on the maximum step (key GEOMETRY, subkey step).Constraints are violated
Problem: constraints are violated: coordinates that were specified as frozen turn out to change during the optimization or coordinates that should remain the same start to differ after a few geometry update cycles.
Possible cause: there is an internal conflict between different demands, usually: symmetry versus constraints. The problem arises easily when a constrained optimization is requested for a molecule with some symmetry while the coordinates were defined with a Z-matrix structure that does not properly reflect the symmetry. Usually the deviations from the requested constraints are small. If they are really large, there might be a bug and you should contact ADF support.
Cure: redefine the Z-matrix and/or use Cartesian optimization (if the constraints are expressible in Cartesians).