Geometry-Corrected Quadratic Optimization Algorithm\nfor NDDO-Descendant Semiempirical Models

The long-held assumption\nthat the optimization of parameters for\nNDDO-descendant semiempirical methods may be performed without precise\ngeometry optimization is assessed in detail; the relevant equations\nfor the analytical evaluation of the geometry-corrected derivatives\nof molecular properties that account for changes in the optimum geometry\nare then presented. The first and second derivatives calculated from\nour implementation of MNDO are used for a limited reparameterization\nof 1,113 CHNO molecules taken from the PM7 training set, demonstrating\nan improvement over the PARAM program used in the optimization of\nparameters for the PMx methods.