Mapping Still Matters: Coarse-Graining with Machine Learning Potentials.

Area of Science:

• Computational chemistry
• Molecular modeling
• Machine learning in science

Background:

• Coarse-grained (CG) modeling extends the reach of molecular simulations to larger scales.
• The accuracy of classical CG models heavily relies on the chosen mapping strategy.
• Machine learning potentials (MLPs) offer a new avenue for developing accurate CG models.

Purpose of the Study:

• To investigate the impact of mapping choices on representations learned by equivariant MLPs.
• To identify potential pitfalls in CG model development using MLPs.
• To provide guidance for creating transferable CG models.

Main Methods:

• Systematic study of liquid hexane, amino acids, and polyalanine.
• Utilizing equivariant machine learning potentials (MLPs).
• Analyzing the influence of mapping on learned representations.

Main Results:

• Overlapping bonded and nonbonded interaction length scales can cause unphysical bond permutations.
• Failure to encode species or maintain stereochemistry introduces unphysical symmetries.
• Equivariant MLPs are sensitive to the details of the CG mapping.

Conclusions:

• The selection of CG mapping significantly impacts MLP performance and model transferability.
• Careful consideration of species encoding and stereochemistry is essential for accurate CG models.
• Findings offer practical guidelines for developing robust and transferable CG models using MLPs.