Login / Signup

A direct-compatible formulation of the coupled perturbed complete active space self-consistent field equations on graphical processing units.

James W SnyderB Scott FalesEdward G HohensteinBenjamin G LevineTodd J Martinez
Published in: The Journal of chemical physics (2018)
We recently developed an algorithm to compute response properties for the state-averaged complete active space self-consistent field method (SA-CASSCF) that capitalized on sparsity in the atomic orbital basis. Our original algorithm was limited to treating small to moderate sized active spaces, but the recent development of graphical processing unit (GPU) based direct-configuration interaction algorithms provides an opportunity to extend this to large active spaces. We present here a direct-compatible version of the coupled perturbed equations, enabling us to compute response properties for systems treated with arbitrary active spaces (subject to available memory and computation time). This work demonstrates that the computationally demanding portions of the SA-CASSCF method can be formulated in terms of seven fundamental operations, including Coulomb and exchange matrix builds and their derivatives, as well as, generalized one- and two-particle density matrix and σ vector constructions. As in our previous work, this algorithm exhibits low computational scaling and is accelerated by the use of GPUs, making possible optimizations and nonadiabatic dynamics on systems with O(1000) basis functions and O(100) atoms, respectively.
Keyphrases
  • machine learning
  • deep learning
  • drug delivery
  • neural network
  • finite element