Interatomic force from neural network based variational quantum Monte Carlo.

Area of Science:

• Computational physics and chemistry
• Materials science
• Quantum mechanics

Background:

• Accurate ab initio calculations are crucial for advancements in physics, chemistry, biology, and materials science.
• Machine learning, particularly neural networks, has recently accelerated computational techniques in these fields.
• Existing research primarily focuses on neural networks for calculating system energy.

Purpose of the Study:

• To investigate the application of neural network wavefunction methods for calculating interatomic forces.
• To implement and evaluate common force estimators within the variational quantum Monte Carlo (VMC) framework using neural networks.
• To provide guidelines for future applications of neural network wavefunctions in molecular and materials simulations.

Main Methods:

• Implementation and testing of several force estimators in variational quantum Monte Carlo (VMC).
• Utilizing neural network wavefunctions as an ansatz for electronic structure calculations.
• Analysis of the relationship between force error, neural network quality, force term contributions, and computational cost.

Main Results:

• Neural network wavefunctions demonstrate an improvement in calculating interatomic forces compared to traditional VMC.
• The study quantifies the impact of neural network quality on force accuracy.
• Analysis provides insights into the contribution of different force terms and their computational expense.

Conclusions:

• Neural network wavefunction methods show significant promise for simulating the structures and dynamics of molecules and materials.
• The findings suggest these methods can generate valuable training data for developing accurate force fields.
• This work paves the way for more efficient and accurate computational studies in condensed matter physics and quantum chemistry.