Transfer learning of memory kernels for transferable coarse-graining of polymer dynamics

Area of Science:

• Computational chemistry and materials science
• Polymer physics and dynamics
• Machine learning applications in molecular modeling

Background:

• Coarse-grained (CG) modeling is crucial for simulating large-scale polymer systems.
• Reproducing dynamic properties accurately across different conditions remains a challenge for CG models.
• The memory kernel in generalized Langevin dynamics dictates system dynamics across all time scales.

Purpose of the Study:

• To develop and validate transfer learning methods for constructing accurate and transferable CG memory kernels.
• To enhance the predictive power of CG models for polymer solutions under varying thermodynamic conditions.
• To enable efficient CG model development requiring minimal training data.

Main Methods:

• Proposed transfer learning techniques for memory kernels in implicit-solvent CG polymer solutions.
• Utilized Gaussian process regression as the core component for learning memory kernels.
• Integrated model order reduction (proper orthogonal decomposition) and active learning for efficiency.

Main Results:

• Demonstrated the accuracy and efficiency of the proposed transfer learning methods on two polymer solution systems.
• Developed transferable memory kernels enabling accurate out-of-sample predictions, including extrapolated parameter domains.
• Showcased the ability of the resulting CG models to reproduce polymer dynamics across diverse conditions (temperature, viscosity, concentration, length).

Conclusions:

• The developed transfer learning framework significantly improves the transferability and predictive accuracy of CG models for polymer dynamics.
• This approach facilitates the creation of robust CG models capable of simulating polymer solutions under a wide range of conditions.
• The methods offer a practical and data-efficient pathway for advancing CG modeling in polymer science.