Double-Hybrid, but Not Double-Cost: GPU-Accelerated DHDFT for the COMPAS-3 Data Set of Polybenzenoid Hydrocarbons

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Materials Science

Background:

• Double Hybrid Density Functional Theory (DHDFT) offers high accuracy for molecular electronic structure.
• Traditional DHDFT calculations are computationally expensive, limiting their application to smaller systems.
• GPU acceleration is a promising avenue to overcome these computational bottlenecks.

Purpose of the Study:

• To implement and demonstrate the first GPU-accelerated DHDFT method.
• To assess the computational efficiency and scalability of the GPU-accelerated DHDFT approach.
• To benchmark various density functionals for isomerization energies of polybenzenoid hydrocarbons.

Main Methods:

• Implementation of DHDFT with major computational steps optimized for GPUs.
• Utilized the resolution-of-identity (RI) approximation for efficient electron repulsion integral (ERI) calculations.
• Performed large-scale calculations on the COMPAS-3x dataset using the Perlmutter supercomputer.

Main Results:

• Achieved efficient GPU utilization, enabling calculations with basis functions up to g angular momentum.
• Demonstrated feasibility of large-scale DHDFT calculations (∼39,000 isomers) within practical computational limits (900 node-hours).
• Found that the PT2 component adds minimal cost for medium-sized molecules, making DHDFT comparable to hybrid DFT.
• Benchmarking revealed SVWN5 (LDA) outperforms GGAs/MGGA without dispersion corrections; M06-L-D4 (MGGA) is superior with corrections.

Conclusions:

• GPU acceleration significantly reduces the computational cost of DHDFT, making it practical for large chemical systems.
• The developed method enables accurate electronic structure calculations for complex organic molecules.
• The study provides valuable insights into the performance of different density functionals for isomerization energy predictions.