Active learning for robust, high-complexity reactive atomistic simulations

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Physical chemistry

Background:

• Machine learned reactive force fields (MLRFs) use polynomial expansions for reactive simulations.
• Flexible MLRF models can lead to numerous parameters, complicating system analysis.
• Standard iterative fitting methods struggle to create effective training sets for complex systems.

Purpose of the Study:

• To develop an active learning approach for semi-automated generation of informative training sets.
• To enable robust parameterization of machine learned force fields.
• To improve the accuracy of simulations for reactive materials.

Main Methods:

• An active learning strategy utilizing cluster analysis and Shannon information theory.
• Development of a model using linear combinations of Chebyshev polynomials.
• Explicitly describing up to four-body interactions for C/O systems under extreme conditions.

Main Results:

• Demonstrated a flexible training database management approach.
• Developed MLRFs showing excellent agreement with Kohn-Sham density functional theory.
• Accurate prediction of structure, dynamics, and speciation in C/O systems.

Conclusions:

• The active learning method facilitates the creation of robust MLRFs.
• This approach enhances the reliability of simulations for reactive materials.
• The developed models accurately capture complex chemical behaviors.