Sparse tensor based nuclear gradients for periodic Hartree-Fock and low-scaling correlated wave function methods in...

Area of Science:

• Computational chemistry
• Quantum chemistry
• Materials science

Background:

• Double-hybrid density functionals offer high accuracy for electronic structure calculations.
• Traditional methods like Hartree-Fock exact exchange (HFX) and second-order Møller-Plesset (MP2) are computationally expensive.
• This limits their application to large and periodic systems.

Purpose of the Study:

• To develop and implement low-scaling methods for HFX, SOS-MP2, and direct random phase approximation (dRPA) energy gradients.
• To enable accurate double-hybrid calculations for large and periodic condensed phase systems.
• To improve the efficiency and scalability of electronic structure calculations.

Main Methods:

• Developed low-scaling methods for HFX, SOS-MP2, and dRPA energy gradients.
• Implemented these methods in the CP2K software package.
• Utilized resolution-of-the-identity approximation with sparse tensor contractions via DBT and DBM libraries for GPU acceleration.

Main Results:

• Achieved favorable sub-cubic scaling with system size for the implemented methods.
• Demonstrated good strong scaling performance on large supercomputers.
• Observed GPU acceleration up to a factor of 3, showcasing efficient parallelization.

Conclusions:

• The developed low-scaling methods significantly enhance the feasibility of double-hybrid calculations for large and periodic systems.
• These advancements pave the way for more routine application of accurate computational methods in condensed phase research.
• The efficient implementation on hundreds of GPU nodes demonstrates the scalability of the approach.