Machine Learning Diffusion Monte Carlo Forces.

Area of Science:

• Computational chemistry
• Quantum mechanics
• Materials science

Background:

• Diffusion Monte Carlo (DMC) offers high accuracy for electronic properties but is computationally expensive for force calculations.
• Calculating forces with DMC is a bottleneck for ab initio molecular dynamics and geometry optimizations.

Purpose of the Study:

• To develop a machine learning approach for accurately predicting "DMC forces" from DMC energy calculations.
• To enable efficient molecular dynamics simulations and geometry optimizations using high-accuracy DMC potential energy surfaces.

Main Methods:

• Utilized Behler-Parrinello Neural Networks to learn DMC forces from DMC energy data.
• Compared machine learning models trained on various datasets (DFT energies with/without forces, DMC energies without forces).
• Performed rigorous comparisons of potential energy surfaces, dynamics, and optimization predictions.

Main Results:

• Machine-learned DMC dynamics accurately reproduced average bond lengths and angles for C2, H2O, and CH3Cl within a few percent of experimental values.
• Achieved results at approximately one hundredth of the computational cost of traditional methods.
• Demonstrated the feasibility of learning forces from noisy energy data without explicit force information.

Conclusions:

• Machine learning provides a viable and cost-effective method for performing dynamics simulations on high-accuracy DMC potential energy surfaces.
• This technique facilitates the generation of DMC-quality molecular geometries, overcoming current algorithmic limitations.