Performance of Gradient-Corrected and Hybrid Density Functional Theory: Role of the Underlying Local Density Approximation and the Gradient Correction

We have analyzed the performance of several widely used density functional theory procedures, namely B-P86, B-PW91, B-LYP, B3-P86, B3-PW91, and B3-LYP, for the E3 set of thermochemical properties. Each of these procedures employs a local density approximation (LDA) functional and a gradient correction for the correlation energy. We and that the VWN3 LDA functional in B-P86, B-PW91, B3-P86, and B3-PW91 leads to extremely large deviations from benchmark values for heats of formation (as large as -455.6 kJ mo(-1) for the B-PW91(VWN3) value for azulene!) and that VWN3 also gives significant errors in the calculated ionization energies and electron affinities. The PW91 gradient correction generally performs much better than P86 for heats of formation, and we propose that this is because P86 severely violates a uniform scaling condition that PW91 almost satisfies. Thus, of the procedures that we have examined, we recommend the use of the VWN5 or PW92 forms of LDA, preferably in combination with the PW91 gradient correction. Our results confirm previous findings that VWN3 is a more suitable LDA than VWNS for B3-LYP, and we attribute this to fortuitous error cancellation between understabilization of molecules by LYP and overstabilization by VWN3.