Progressive systematic underestimation of reaction energies by the B3LYP model as the number of C-C bonds increases: why organic chemists should use multiple DFT models for calculations involving polycarbon hydrocarbons.

[reaction: see text] Computational studies of three different reaction types involving hydrocarbons (homolytic C-C bond breaking of alkanes, progressive insertions of triplet methylene into C-H bonds of ethane, and [2+2] cyclizations of methyl-substituted alkenes to form polymethylcyclobutanes) show that the B3LYP model consistently underestimates the reaction energy, even when extremely large basis sets are employed. The error is systematic and cumulative, such that the reaction energies of reactions involving hydrocarbons with more than 4-6 C-C bonds are predicted quite poorly. Energies are underestimated for slightly and highly methyl-substituted cyclic and acyclic hydrocarbons, so the errors do not arise from structural issues such as steric repulsion or ring strain energy. We trace the error associated with the B3LYP approach to its consistent underestimation of the C-C bond energy. Other DFT models show this problem to lesser extents, while the MP2 method avoids it. As a consequence, we discourage the use of the B3LYP model for reaction energy calculations for organic compounds containing more than four carbon atoms. We advocate use of a collection of pure and hybrid DFT models (and ab initio models where possible) to provide computational "error bars".