Massively parallelization strategy for material simulation using high-dimensional neural network potential

Area of Science:

• Computational Materials Science
• Materials Simulation
• Machine Learning in Materials

Background:

• Potential energy surface (PES) calculations are computationally intensive and limit modern material simulations.
• High-dimensional neural networks (HDNNs) offer a promising approach for fast and accurate PES computations.
• The calculation of structural descriptors, crucial for HDNNs, represents a significant computational bottleneck.

Purpose of the Study:

• To introduce a massive parallelization strategy for optimizing the computation of power-type structural descriptors.
• To address the computational cost associated with HDNN-based PES calculations.
• To enhance the efficiency of material simulations by accelerating descriptor computation.

Main Methods:

• Developed a three-level massive parallelization strategy for structural descriptor computation.
• Parallelization is applied sequentially over atoms, structural descriptors, and n-body functions.
• The strategy is implemented and tested using a boron crystal system.

Main Results:

• Demonstrated significant parallelization efficiency, reaching up to 100% at the atom level.
• Achieved 58% and 34% parallelization efficiency at the structural descriptor and n-body function levels, respectively.
• The proposed method effectively reduces the computational bottleneck in HDNN calculations.

Conclusions:

• The introduced massive parallelization strategy substantially accelerates the computation of structural descriptors for HDNNs.
• This optimization is critical for enabling faster and more accurate potential energy surface calculations in materials science.
• The method provides a scalable solution for computationally demanding material simulations.