Operator forces for coarse-grained molecular dynamics.

Area of Science:

• Computational Chemistry
• Molecular Dynamics Simulations
• Machine Learning in Chemistry

Background:

• Coarse-grained (CG) molecular dynamics (MD) simulations enhance computational efficiency by representing groups of atoms as single beads.
• Machine-learned coarse-graining (MLCG) offers a powerful method for developing accurate CG force fields.
• Traditional MLCG calibration relies on force matching, requiring extensive atomistic simulation data, including forces, which is often unavailable for existing datasets.

Purpose of the Study:

• To develop a novel kernel-based approach for calibrating MLCG force fields.
• To overcome the limitations of traditional force matching, particularly in low-data regimes or when reference forces are absent.
• To reduce local distortions introduced by previous noise-based kernel methods while maintaining global accuracy.

Main Methods:

• Introduction of general kernels based on normalizing flows for MLCG force field construction.
• Adaptation of force matching to utilize only configurational samples, eliminating the need for explicit force labels.
• Demonstration of the method's efficacy on small protein systems.

Main Results:

• Flow-based kernels significantly reduce local distortions compared to noise-based kernels.
• The proposed method preserves global conformational accuracy in CG molecular dynamics.
• High-quality CG forces can be generated using only configurational data, demonstrating the utility of flow-based kernels.

Conclusions:

• Normalizing flow-based kernels represent a significant advancement in machine-learned coarse-graining.
• This approach enables accurate force field generation from limited or force-absent atomistic data.
• The method enhances the applicability of CG simulations for large-scale molecular modeling.