Knowledge distillation of noisy force labels for improved coarse-grained force fields.

Area of Science:

• Computational chemistry
• Materials science
• Molecular dynamics

Background:

• Molecular dynamics (MD) simulations are crucial for atomistic material behavior but computationally intensive.
• Coarse-grained (CG) models simplify systems to reduce computational cost but sacrifice atomic detail and introduce mapping noise.
• CG models' inherent entropic contributions prevent direct fitting to all-atom (AA) energies, limiting training data to noisy forces.

Purpose of the Study:

• To develop a knowledge distillation framework for training more accurate and stable CG models.
• To address the challenges of mapping noise and limited training data in CG model development.
• To improve the efficiency and accuracy of MD simulations using enhanced CG force fields.

Main Methods:

• A knowledge distillation framework was implemented using a teacher-student neural network approach.
• The initial CG neural network potential (teacher) was trained on denoised AA-mapped forces.
• Refined CG models (students) were trained using force and energy predictions from the teacher model in single- and ensemble-training setups.

Main Results:

• The framework successfully denoises AA-mapped forces and distills knowledge to train CG models.
• Validation on a deep eutectic solvent showed improved two-, three-, and many-body properties compared to traditional CG models.
• Training the student model on ensemble teacher-predicted forces and per-bead energies significantly enhanced CG force field quality and stability.

Conclusions:

• Knowledge distillation offers a robust method for improving CG model accuracy and stability.
• The developed framework overcomes limitations of traditional CG model training, particularly for complex systems.
• This approach paves the way for more reliable and efficient large-scale molecular dynamics simulations.