Ensemble learning of coarse-grained molecular dynamics force fields with a kernel approach.

Area of Science:

• Computational Chemistry
• Machine Learning in Molecular Modeling
• Biophysics

Background:

• Gradient-domain machine learning (GDML) accurately learns molecular potentials and force fields using kernel ridge regression.
• Coarse-grained (CG) models simplify molecular systems but learning them from all-atom data presents computational challenges.
• Direct application of GDML to CG modeling is hindered by memory constraints due to data averaging.

Purpose of the Study:

• To develop a data-efficient and memory-saving method for learning effective CG models using GDML.
• To apply GDML to learn a CG force field from all-atom simulation data.
• To reconstruct the free energy landscape of a CG biomolecular system.

Main Methods:

• A novel 2-layer training scheme combining ensemble learning and stratified sampling for GDML.
• Learning CG force fields based on the thermodynamic consistency principle by minimizing force errors.
• Application to the alanine dipeptide system to reconstruct its CG free energy landscape.

Main Results:

• The proposed GDML training scheme enables efficient learning of CG models.
• The method achieves smaller free energy errors than neural networks on small training datasets.
• High accuracy comparable to neural networks is achieved with sufficiently large training sets.

Conclusions:

• The novel GDML training scheme effectively addresses memory and data efficiency challenges in CG model learning.
• This approach provides a powerful alternative for developing accurate CG force fields.
• The method demonstrates superior performance on small datasets compared to neural networks, highlighting its utility in resource-limited scenarios.