I have only seen this once before, and it was when using multi-level SE... it might be related to SNORM and the inability to correctly calculate the normal vector for multilevel SE. You might be able to punch out the normal vector for a non-se version of the run and then incorporate that into the SE version of the run (using the SNORM bulk data entry), or use PARAM,SNORM,0.0 (which might not give correct answers). SNORM is used to determine the appropriate normal direction for CQUAD4, CTRIA3, CQUADR, and CTRIAR elements. If there isn't any curvature (or 'angle') between adjacent elements, then SNORM wouldn't make much difference.
