FEREBUS: Highly parallelized engine for kriging training.

Area of Science:

• Computational chemistry and materials science.
• Development of novel computational methods and algorithms.

Background:

• FFLUX is a new force field utilizing quantum topological atoms, multipolar electrostatics, and IQA energy terms.
• The FEREBUS program calculates hyperparameters for kriging machine learning models.
• Optimization of the log-likelihood function is computationally intensive.

Purpose of the Study:

• To present a parallelization strategy for the FEREBUS program.
• To optimize the calculation of kriging hyperparameters using Particle Swarm Optimization (PSO) and Differential Evolution (DE).
• To significantly reduce the computational time required for model generation.

Main Methods:

• Implementation of MPI parallelization to distribute particles/vectors across processes.
• Utilization of OpenMP for parallelizing the log-likelihood calculation, including matrix operations.
• Application of PSO and DE algorithms for optimizing the concentrated log-likelihood function.

Main Results:

• Achieved a speed-up of 61 times by scaling from a single core to 90 cores.
• Reduced computational time from 2871 seconds (single core) to 41 seconds (90 cores).
• Demonstrated significant computational savings of approximately 98% of single-core time.

Conclusions:

• The presented parallelization strategy for FEREBUS is highly effective in accelerating hyperparameter optimization.
• This approach significantly enhances the efficiency of generating machine learning models based on the FLUX force field.
• The parallelized FEREBUS program offers substantial computational time savings for complex scientific modeling tasks.