GPU-Accelerated Neural Network Potential Energy Surfaces for Diffusion Monte Carlo.

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Molecular Dynamics

Background:

• Diffusion Monte Carlo (DMC) is crucial for studying molecular vibrations, but potential energy calculations are computationally intensive.
• Conventional methods struggle with complex molecular systems.

Purpose of the Study:

• To develop a computationally efficient method for DMC calculations.
• To accelerate the evaluation of potential energy surfaces (PES) in molecular simulations.

Main Methods:

• Training neural network potential energy surfaces (NN PES) using data from small-scale DMC calculations.
• Utilizing graphics processing units (GPUs) for highly parallelizable NN PES evaluations.
• Incorporating permutation symmetry into NN descriptors for specific molecules.

Main Results:

• NN PES provide accurate zero-point energies and wave functions comparable to traditional methods.
• Significant reduction in computational requirements for DMC simulations.
• Demonstrated success for H2O, CH5+, and (H2O)2 systems.

Conclusions:

• Neural network potentials offer a computationally advantageous alternative for DMC simulations.
• The developed approach accelerates the study of molecular vibrational landscapes.
• Permutation symmetry is important for accurate NN PES in certain molecular systems.