Basis Set Effects
For the frequently-used B3LYP-D3BJ and PBE-D3BJ density functional methods, we considered both the def2-SVP and def2-TZVP basis sets. In both cases, the triple zeta basis set significantly improved correlation with the DLPNO-CCSD(T)/cc-pVTZ baseline, for example, the median R2 scores improved from 0.868±0.064 to 0.920±0.025 for B3LYP-D3BJ and from 0.835±0.025 to 0.885±0.018 for PBE-D3BJ. There were comparable improvements in median Spearman rank correlation and reduced mean absolute relative errors, all statistically significant (i.e. p-values far below 0.001). The increased basis sets also roughly doubled the CPU time required.
While the PBE method is still significantly faster than B3LYP, the newer B97-3c proves to be faster than either with comparable accuracy (i.e., roughly intermediate to the TZ results for B3LYP-D3BJ and PBE-D3BJ). Additionally, the time required for B3LYP-D3BJ/def2-TZVP calculations is only slightly less than RI-MP2/cc-pVTZ results, which exhibit significantly improved accuracy relative to DLPNO-CCSD(T)/cc-pVTZ (i.e., median R2 = 0.964±0.006 and median MARE of 0.115 ±0.011 kcal/mol for RI-MP2).
Thus increasing basis set size for these density functional methods, at least from double zeta to triple zeta, does improve accuracy, albeit at a significant computational cost. In general, the B97-3c method provides accuracy comparable to popular dispersion-corrected DFT methods such as B3LYP-D3BJ with faster performance, and RI-MP2 provides greater accuracy at a very similar speed.
Dispersion Corrections
Since the bonding is consistent across multiple conformers, the ranking of small energy differences is known to be dominated by non-bonding interactions.\cite{Martin_2013,Gruzman_2009} Density functional methods are known to incorrectly account for dispersion interactions, which has led to a variety of empirical corrections.\cite{Grimme_2011,Becke_2005,Johnson_2005,Johnson_2006,Caldeweyher_2017,Caldeweyher_2019,Grimme_2011a,Johnson_2009} Comparing un-corrected PBE, B3LYP, and ωB97X single-point energies to DLPNO-CCSD(T) illustrate a significant effect. The uncorrected median R2 values drop by ~0.12, and the median Spearman correlations drop by ~0.08. For example, the median R2 of B3LYP / TZ drops from 0.920±0.012 to 0.706±0.050 without the D3BJ dispersion correction.