Learning composition-transferable coarse-grained models: Designing external potential ensembles to maximize...

Area of Science:

• Computational chemistry and molecular modeling.
• Statistical mechanics and thermodynamics.
• Materials science and condensed matter physics.

Background:

• Bottom-up coarse-graining of molecular models faces challenges in thermodynamic faithfulness and transferability across different conditions.
• Current efforts often focus on refining the functional form of coarse-grained potentials.
• The choice of simulation ensemble is critical for model development but often overlooked.

Purpose of the Study:

• To investigate the impact of simulation ensemble choice on the predictive power of coarse-grained models.
• To demonstrate that specific ensembles can enhance thermodynamic faithfulness and transferability.
• To introduce a method for systematically designing informative ensembles using the Fisher information metric.

Main Methods:

• Coarse-graining molecular models from ensembles with applied perturbations, specifically spatially varying external potentials.
• Applying the strategy to atomistic models of water and methanol, and a binary mixture of Gaussian spheres.
• Quantifying ensemble informativeness using the Fisher information metric to design optimal bias potentials.

Main Results:

• Near-quantitative capture of activity coefficients across the entire composition range for tested systems.
• Demonstrated success without explicitly targeting activity coefficients during the coarse-graining process.
• Validation of the hypothesis that ensembles with applied thermodynamic potentials are more "thermodynamically informative".

Conclusions:

• Smarter selection of simulation ensembles, particularly those with applied thermodynamic potentials, significantly improves coarse-grained model accuracy and transferability.
• The Fisher information metric provides a principled approach to designing optimal ensembles for learning thermodynamically faithful models.
• This work offers a new paradigm for developing robust and predictive coarse-grained molecular models.