Linear-scaling local natural orbital CCSD(T) approach for open-shell systems: algorithm, benchmarks, and large-scale applications (2311.10048v1)

Abstract: The extension of the highly-optimized local natural orbital (LNO) CCSD(T) method is presented for high-spin open-shell molecules. The techniques enabling the outstanding efficiency of the closed-shell LNO-CCSD(T) variant are adopted, including the iteration- and redundancy-free MP2 and (T) formulations, as well as the integral-direct, memory- and disk use economic, and OpenMP-parallel algorithms. For large molecules, the efficiency of our open-shell LNO-CCSD(T) method approaches that of its closed-shell parent method due to a novel approximation for higher-order long-range spin-polarization effects. The accuracy of open-shell LNO-CCSD(T) is extensively tested for radicals and reactions thereof, ionization processes, as well as spin-state splittings and transition-metal compounds. At the size range, where the canonical CCSD(T) reference is accessible (up to 20-30 atoms) the average open-shell LNO-CCSD(T) correlation energies are found to be 99.9-99.95% accurate, which translates into average absolute deviations of a few tenth of a kcal/mol in the investigated energy differences already with the default settings. This enables the accurate modeling of large systems with complex electronic structure, as illustrated on open-shell organic radicals and transition metal complexes of up to 179 atoms, as well as on challenging biochemical systems, including up to 601 atoms and 11,000 basis functions. While the protein models involve difficulties for local approximations, such as the spin states of a bounded iron ion or an extremely delocalized singly occupied orbital, the corresponding single-node LNO-CCSD(T) computations were feasible in a matter of days with 10s to a 100 GB of memory use. Therefore, the new LNO-CCSD(T) implementation enables highly-accurate computations for open-shell systems of unprecedented size and complexity with widely accessible hardware.