(169l) The Tradeoff between Chemical Accuracy and Computational Cost: An Assessment of Thermochemical Prediction with Density Functional Theory

Density functional theory (DFT) methods are widely used for computing thermochemical properties such as non-covalent interactions, barrier heights, and reaction energies. The choice of a particular model chemistry, which includes selecting an appropriate exchange-correlation (XC) functional and basis set, is important for achieving chemical accuracy. There exist several classes of XC functionals, each experiencing some degree of tradeoff between chemical accuracy and computational complexity. Additionally, increasing the size of the basis set generally improves the accuracy of computed molecular properties at an additional computational cost. While there have been numerous benchmarking studies for density functionals, the majority only focus on chemical accuracy. We argue that it is necessary to also consider the compute wall time when selecting an appropriate model chemistry suitable for e.g. high-throughput computational screening or big data assembly.

In this work, we benchmark XC functional and basis set pairings using datasets covering a wide range of thermochemical prediction tasks, including non-covalent interactions (DES15K), barrier heights (BH9), reaction energies (BSE49), and an all-around thermochemistry benchmark (GMTKN55). We investigate the use of some empirical correction methods which address known deficiencies in DFT, such as incomplete basis set error, electron self-interaction error, and the underestimation of Van der Waals interactions. We show that smaller basis sets paired with the DFT-C correction are capable of achieving near chemical accuracy on a surprising number of property prediction calculations. We find that while XC functionals with high computational complexity consistently perform better, XC functionals with relatively lower complexity can perform similarly when paired with the appropriate empirical corrections. We note trends of chemical accuracy of density functionals and give several recommendations on model chemistry choice with respect to necessary compute time and type of chemistry.