Acceleration of Self-Consistent Field Calculations Using Basis Set Projection and Many-Body Expansion as Initial Guess Methods

In Self-Consistent Field (SCF) calculations, the choice
of initial
guess plays a key role in determining the time-to-solution by influencing
the number of iterations required for convergence. However, focusing
solely on reducing iterations may overlook the computational cost
associated with improving the accuracy of initial guesses. This study
critically evaluates the effectiveness of two initial guess methodsbasis
set projection (BSP) and many-body expansion (MBE) on Hartree–Fock
and hybrid Density Functional Theory (B3LYP and MN15) methods. We
also introduce a new approach, a hybrid of MBE and BSP. Our assessment
considers both the number of SCF iterations and the total computational
wall-times. The results demonstrate that BSP, MBE, and the hybrid
method could significantly outperform the conventional superposition
of atomic densities (SAD) technique. With these methods, reductions
in total wall-time, including the time spent generating initial guesses,
by up to 21.9, 27.6, and 21.6% could be observed with HF, B3LYP, and
MN15, respectively, when tested on systems containing up to 14,386
basis functions. Furthermore, we also examine the influence of these
initial guess schemes on difficult-to-converge metalloprotein and
triplet electronic states. Speedups could be observed with non-SAD
approaches although in the case of triplet electronic states, higher
convergence failures could be observed.