Modular implementation of the linear- and cubic-scaling orbital minimization methods in electronic structure codes...

Area of Science:

• Computational physics and chemistry
• Materials science
• Quantum mechanics

Background:

• Electronic structure calculations are crucial for understanding material properties.
• Existing methods often face performance limitations with increasing system size.
• Efficient and scalable computational approaches are needed for modern materials discovery.

Purpose of the Study:

• To present a modular code design for efficient and massively parallel electronic structure solvers.
• To implement and evaluate both cubic- and linear-scaling algorithms.
• To demonstrate the performance benefits of linear-scaling methods for large systems.

Main Methods:

• Modular implementation of the orbital minimization method within the SIESTA code.
• Leveraging external libraries: Distributed Block Compressed Sparse Row (DBCSR) for sparse matrices and Scalable Linear Algebra Package (ScaLAPACK) for dense matrices.
• Utilizing MatrixSwitch and libOMM from the Electronic Structure Library for matrix format flexibility and energy minimization.

Main Results:

• Demonstrated the performance of various cubic-scaling algorithms.
• Showcased the parallel performance of newly developed linear-scaling solvers.
• Confirmed the superiority of linear-scaling solvers over cubic-scaling ones for insulating systems of several hundred atoms.

Conclusions:

• The modular approach enables efficient and massively parallel solvers for electronic structure calculations.
• Linear-scaling solvers offer significant advantages for large insulating systems.
• This work advances the capability of computational methods in materials science and quantum chemistry.