Enabling accurate and large-scale explicitly correlated CCSD(T) computations via a reduced-cost and parallel implementation

Parallel algorithms to accelerate explicitly correlated second-order Moller–Plesset (MP2) and coupled-cluster singles and doubles with perturbative triples [CCSD(T)] calculations and benchmarks on extended molecular systems are reported. A hybrid Open Multi-Processing (OpenMP)/Message Passing Interface (MPI) parallel approach is used to distribute the computational load among processor cores and compute nodes. The intermediates at both the MP2 and the CCSD(T) levels are expressed in a density fitting formalism, using only three-index quantities to decrease the amount of data to be stored and communicated. To further reduce compute time, the frozen natural orbital, the natural auxiliary function, and the natural auxiliary basis schemes are implemented in a hybrid parallel manner. The combination of these three approximations and our recent size-consistent explicitly correlated triples correction with the new hybrid parallelization offers a unique accuracy-over-cost performance among explicitly correlated CC methods. Our comprehensive benchmarks demonstrate excellent parallel scaling of the cost-determining operations up to hundreds of processor cores. As demonstrated on the non-covalent interaction energy of the corannulene dimer, highly- accurate explicitly correlated CCSD(T) calculations can be carried out for systems of 60-atoms and 2500 orbitals, which were beyond computational limits without local correlation approximations. This enables various applications, such as benchmarking of or, for certain size ranges, replacing local CCSD(T) or density functional methods as well as the further advancement of robust thermochemistry protocols designed for larger molecules of ca. 20–50-atoms.