Transfer Learning of Memory Kernels in Coarse-grained Modeling

Zhan Ma, Shu Wang, Minhee Kim, Kaibo Liu, Chun-Long Chen, Wenxiao PanarXiv 2021 · OpenReview

TL;DR

This paper introduces a novel transfer learning framework for coarse-grained (CG) molecular dynamics simulations, specifically focusing on polymer solutions. By leveraging Gaussian Process Regression (GPR) combined with model order reduction (POD) and active learning, they enable CG models to accurately predict polymer dynamics across a wide range of conditions (temperature, concentration, viscosity) and even extrapolate to unseen parameter regimes, significantly reducing computational cost.

What Problem Does This Paper Address?

Coarse-grained (CG) molecular dynamics simulations offer a computationally efficient way to study large molecular systems like polymers and biomolecules. However, a major challenge is ensuring that these simplified CG models accurately capture the dynamic properties (like diffusion and relaxation times) of the original, more detailed atomistic systems. This accuracy is particularly difficult to achieve when the CG model needs to be applied under varying conditions (e.g., different temperatures or solvent properties) or for different system configurations (e.g., varying polymer concentrations or lengths). The "memory kernel" in the Generalized Langevin Equation (GLE), which describes the dissipative forces in CG models, is crucial for capturing these dynamics across all time scales. However, constructing transferable memory kernels that work across these different conditions has been a significant hurdle.

Methods

The core of this work lies in developing methods for "transfer learning" of these memory kernels. The authors propose two main approaches:

  1. GPR Transfer Learning: This method directly uses Gaussian Process Regression (GPR) to learn the memory kernel as a function of time and various system parameters (temperature, concentration, viscosity). GPR is chosen for its ability to provide predictions with quantified uncertainty, allowing for both interpolation and extrapolation.
  2. ROM-GPR Transfer Learning: This approach enhances the GPR method by integrating it with Model Order Reduction (MOR), specifically using Proper Orthogonal Decomposition (POD). The memory kernel is first decomposed into dominant temporal and parameter modes. GPR is then applied to learn these dominant parameter modes, significantly reducing the number of data points and computational cost required for training.

Furthermore, they incorporate an Active Learning strategy. This technique intelligently guides the sampling of new training data by focusing on regions where the GPR model's uncertainty is highest. This ensures that new data points provide the most information gain, leading to more accurate models with fewer training data points.

Datasets

The paper uses data generated from atomistic molecular dynamics (MD) simulations of two example systems:

  • Star Polymer Solutions: Atomistic simulations of star polymers with varying concentrations and temperatures.
  • Peptoid Polymer Solutions: Atomistic simulations of peptoid polymers with varying temperatures and lengths (number of repeat units).

The memory kernels derived from these atomistic simulations serve as the "ground truth" for training and evaluating the transfer learning methods. The authors do not introduce new benchmark datasets but rather utilize existing simulation methodologies to generate the necessary data for their proposed framework.

Key Insights and Ideas

  • Memory Kernel is Key: The paper strongly emphasizes that the memory kernel is the critical component for CG models to accurately reproduce dynamic properties across time scales.
  • GPR for Transferability: Gaussian Process Regression proves to be a powerful tool for learning the complex, multi-parameter dependence of memory kernels, enabling transferability.
  • POD for Efficiency: Decomposing the memory kernel into dominant modes via POD dramatically reduces the dimensionality of the problem, making GPR training computationally feasible for complex systems.
  • Active Learning for Data Efficiency: Intelligently selecting training data points based on uncertainty quantification is crucial for minimizing computational effort and achieving high accuracy, especially in high-dimensional parameter spaces.
  • Extrapolation is Possible: The framework demonstrates the ability to predict memory kernels and thus CG dynamics in parameter regimes not seen during training, a significant advantage for predictive modeling.

Key Results

  • Accuracy: The ROM-GPR transfer learning method, especially when coupled with active learning, achieved high accuracy in predicting memory kernels. For star polymers, relative errors were as low as 2.5% for interpolation and increased to around 6.8% for extrapolation. For peptoid polymers, errors were around 2.8% for interpolation and 8.6% for extrapolation.
  • Computational Speedup: The ROM-GPR method offered substantial computational savings compared to GPR alone. For star polymers, constructing the model and predicting the kernel was significantly faster (e.g., 0.1127s vs. 1086s for model construction).
  • Reproducing Dynamics: CG models built with the transferred memory kernels accurately reproduced atomistic dynamics, including velocity autocorrelation functions (VACF), diffusion coefficients, and mean squared displacements (MSD), across various thermodynamic conditions and system parameters. This included capturing different diffusion regimes (super-, sub-, and normal diffusion).
  • Multi-parameter Transferability: The methods were successfully applied to systems with multiple varying parameters (temperature, concentration, solvent viscosity, polymer length), demonstrating robust transferability.

Limitations and Future Work

  • Computational Cost of Atomistic Data: While the transfer learning methods are efficient, generating the initial atomistic data for training the memory kernels can still be computationally expensive.
  • Parameter Space Exploration: The active learning strategy is effective, but the initial sampling and the definition of the parameter space still require careful consideration.
  • Future Work: The authors suggest integrating their transferable memory kernel approach with existing methods for transferable CG potentials to ensure both structural and dynamic properties are preserved. They also note the general applicability of their methodology to a broader range of parameters and soft matter systems.

Summary updated

Abstract

The present work concerns the transferability of coarse-grained (CG) modeling in reproducing the dynamic properties of the reference atomistic systems across a range of parameters. In particular, we focus on implicit-solvent CG modeling of polymer solutions. The CG model is based on the generalized Langevin equation, where the memory kernel plays the critical role in determining the dynamics in all time scales. Thus, we propose methods for transfer learning of memory kernels. The key ingredient of our methods is Gaussian process regression. By integration with the model order reduction via proper orthogonal decomposition and the active learning technique, the transfer learning can be practically efficient and requires minimum training data. Through two example polymer solution systems, we demonstrate the accuracy and efficiency of the proposed transfer learning methods in the construction of transferable memory kernels. The transferability allows for out-of-sample predictions, even in the extrapolated domain of parameters. Built on the transferable memory kernels, the CG models can reproduce the dynamic properties of polymers in all time scales at different thermodynamic conditions (such as temperature and solvent viscosity) and for different systems with varying concentrations and lengths of polymers.
BibTeX
@misc{ma2021transfer,
  author = {Zhan Ma and Shu Wang and Minhee Kim and Kaibo Liu and Chun-Long Chen and Wenxiao Pan},
  title = {{Transfer Learning of Memory Kernels in Coarse-grained Modeling}},
  year = {2021},
  eprint = {2103.10578},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.10578},
}
Figure