Robert E. Thomas, Qiming Sun, Ali Alavi, & George H. Booth, (2015) DOI: 10.1021/acs.jctc.5b00917
Contributed by Jan Jensen
This work is licensed under a Creative Commons Attribution 4.0
Contributed by Jan Jensen
The multiconfigurational self-consistent field (MCSCF), especially complete active space SCF (CASSCF) method remains the only systematically improveable option for a host of molecules and processes with strong non-dynamical correlation. However, the requirement of a full configuration interaction (FCI) calculation for the active space results in a computational cost that scales exponentially with the size of the active space. There are therefore many interesting problems that remain out of reach for CASSCF. But perhaps not for long.
This papers shows that the FCI steps of the CASSCF calculation can be replaced by FCI quantum Monte Carlo (FCIQMC) calculations, which scales significantly better than FCI. In the FCIQMC approach new determinants ($j$) are generated from old determinants ($i$) stochastically and included if the probability of this generation step exceeds a randomly chosen number between 0 and 1. The authors show that ca 25,000 sampling iterations yield energies within 0.1 milli-Hartrees of conventional CASSCF. As a result this approach could be used for a complete active space with 24 electrons in 24 orbitals with "only modest computational resources"!
This papers shows that the FCI steps of the CASSCF calculation can be replaced by FCI quantum Monte Carlo (FCIQMC) calculations, which scales significantly better than FCI. In the FCIQMC approach new determinants ($j$) are generated from old determinants ($i$) stochastically and included if the probability of this generation step exceeds a randomly chosen number between 0 and 1. The authors show that ca 25,000 sampling iterations yield energies within 0.1 milli-Hartrees of conventional CASSCF. As a result this approach could be used for a complete active space with 24 electrons in 24 orbitals with "only modest computational resources"!
This work is licensed under a Creative Commons Attribution 4.0