Parallel Multi-Coordinate Descent Methods for Full Configuration Interaction

TL;DR

This paper introduces a parallel coordinate descent algorithm for full configuration interaction calculations, enabling efficient electronic structure computations on multi-core systems with high parallel efficiency.

Contribution

The paper presents mCDFCI, a novel multi-threaded algorithm that reformulates FCI as an unconstrained minimization problem and employs a deterministic compression strategy for parallel determinant selection.

Findings

Successfully computed the chromium dimer energy with over 2 billion determinants.

Achieved up to 79.3% parallel efficiency on 128 cores.

Provided the nitrogen dimer binding curve with high accuracy.

Abstract

We develop a multi-threaded parallel coordinate descent full configuration interaction algorithm (mCDFCI), for the electronic structure ground-state calculation in the configuration interaction framework. The FCI problem is reformulated as an unconstrained minimization problem, and tackled by a modified block coordinate descent method with a deterministic compression strategy. mCDFCI is designed to prioritize determinants based on their importance, with block updates enabling efficient parallelization on shared-memory, multi-core computing infrastructure. We demonstrate the efficiency of the algorithm by computing an accurate benchmark energy for the chromium dimer in the Ahlrichs SV basis (48e, 42o), which explicitly includes $2.07 \times 10^9$ variational determinants. We also provide the binding curve of the nitrogen dimer under the cc-pVQZ basis set (14e, 110o). Benchmarks show up to $79.3\%$ parallel efficiency on 128 cores.