Producing chemically accurate atomic Gaussian process regression models by active learning for molecular simulation.

Area of Science:

• Computational chemistry
• Materials science

Background:

• Machine learning (ML) is vital for developing sophisticated molecular force fields.
• Chemically accurate ML potentials are essential as force field applications expand.
• Gaussian process regression (GPR) is a powerful ML technique for potential development.

Purpose of the Study:

• To demonstrate a method for developing chemically accurate GPR models.
• To apply active learning for efficient model development across complex molecules.
• To extend previous work on active learning for force field accuracy.

Main Methods:

• Utilizing a per-atom active learning approach.
• Developing Gaussian process regression models.
• Testing models on an increasingly complex set of molecules, including peptide-capped glycine.

Main Results:

• The active learning technique significantly reduces computational time (CPU time).
• The method successfully generates chemically accurate ML potentials.
• Demonstrated progression towards more accurate models with less computational cost.

Conclusions:

• The per-atom active learning approach enables efficient development of accurate ML potentials.
• This method is effective for complex molecular systems like peptide-capped glycine.
• The study highlights the potential of active learning in advancing force field development.