Automatic Generation of Accurate and Cost-efficient Auxiliary Basis Sets

TL;DR

This paper presents methods to automatically generate auxiliary basis sets for density fitting in quantum chemistry, reducing their size and computational cost while maintaining high accuracy across various molecules and levels of theory.

Contribution

The authors introduce two truncation techniques—singular value decomposition contraction and high-angular momentum function removal—to produce smaller, cost-effective auxiliary basis sets without sacrificing accuracy.

Findings

Accurate energies achieved with significantly reduced basis set sizes.

Reduction schemes decrease basis set size scaling factor from ~10 to ~3-6.

Generated auxiliary basis sets are highly transferable across methods and molecules.

Abstract

We have recently discussed an algorithm to automatically generate auxiliary basis sets (ABSs) of the standard form for density fitting (DF) or resolution-of-the-identity (RI) calculations in a given atomic orbital basis set (OBS) of any form [J. Chem. Theory Comput. 2021, 17, 6886]. In this work, we study two ways to reduce the cost of such automatically generated ABSs without sacrificing their accuracy. We contract the ABS with a singular value decomposition proposed by K\'allay [J. Chem. Phys. 2014, 141, 244113], used here in a somewhat different setting. We also drop the high-angular momentum functions from the ABS, as they are unnecessary for global fitting methods. Studying the effect of these two types of truncations on Hartree--Fock (HF) and second-order M{\o}ller--Plesset perturbation theory (MP2) calculations on a chemically diverse set of first- and second-row molecules within the RI/DF approach, we show that accurate total and atomization energies can be achieved by a combination of the two approaches with significant reductions in the size of the ABS. While the original approach yields ABSs whose number of functions $N_{\text{bf}}^{\text{ABS}}$ scales with the number of functions in the OBS, $N_{\text{OBS}}^{\text{bf}}$, as $N_{\text{ABS}}^{\text{bf}}=\gamma N_{\text{OBS}}^{\text{bf}}$ with the prefactor $\gamma\approx\mathcal{O}(10)$, the reduction schemes of this work afford results of essentially the same quality as the original unpruned and uncontracted ABS with $\gamma\approx5\text{-}6$, while an accuracy that may suffice for routine applications is achievable with a further reduced ABS with $\gamma\approx3\text{-}4$. The observed errors are similar at HF and MP2 levels of theory, suggesting that the generated ABSs are highly transferable, and can also be applied to model challenging properties with high-level methods.