Essential defects of present-day semiempirical methods (CNDO/%, CNDO/S, MIND0/3, MNDO) can be identified on two levels. First, the formalism shared by nearly all these methods treats parametric expressions, supposed to refer to Lowdin orthogonalized orbitals, as transferable. At best, this idea is approximately valid for the B functions for strong bonds. Second, the expressions used for these approximately transferable functions are based on inadequate arguments. As a result, the errors in many terms of the energy expression amount to several electron volts. The mechanism of error compensation through the parametrization is investigated in detail and shown to be surprisingly flexible, at least for some combinations of errors. Some cases of systematic absence of compensation are identified, and for weak interactions, e.g., those responsible for the water dimer, the semiempirical expressions appear to be so irrelevant that a detailed analysis of the causes of their failure is no longer possible. Minimal requirements for a correct semiempirical approach are given in the MORBIT rules. The two-center one-electron integral p is to be replaced by a function introduced by Mulliken; an efficient approximation for this expression is proposed.
Defining the ProblemIn the past 20 years several semiempirical methods have been designed which can reproduce chosen experimental results for a wide range of molecules by means of a simple computational scheme. As examples 1 mention M I N D O /~ [l], which is suitable for the calculation of equilibrium geometries and heats of formation, and CNDO/S [2], useful in the study of uv spectra. Whenever such a theory is successful in its chosen domain the term "theory of structure" tends to come up, though its meaning is seldom specified. In the present paper I shall use this term exclusively for those methods which not only give good results for, e.g., bondlengths and energies (although such a method qualifies as a useful algorithm) but also explain why the energy assumes a given minimum value when the R are assigned their equilibrium values {Re}. This means that not only E ( R , ) should be correct, but also that the separate terms in the energy expression must be correct as functions of {R}. What I intend to show is that: (a) several popular semiempirical methods cannot be accepted as theories of structure; (b) any useful algorithm which is not a theory of structure ceases to be useful in specified, frequently occurring, circumstances. Now an analysis of the terms in a semiempirical energy expression is subject to some uncertainty because most of these terms are not clearly defined. This is because semiempirical methods try to condense as much information as possible into the few energy terms which are maintained. The formalism corresponds to