Accelerating self-consistent field theoretic simulations for disordered systems with deep learning.

Area of Science:

• Polymer science and computational materials science.
• Application of machine learning in theoretical physics.

Background:

• Polymer thermodynamics and self-assembly are crucial for advanced materials and drug delivery.
• Self-consistent field theory (SCFT) is a powerful tool for predicting polymer behavior.
• Standard SCFT methods face computational challenges with complex systems like anisotropic and worm-like chains.

Purpose of the Study:

• To develop a computationally efficient machine learning approach for SCFT simulations.
• To predict polymer density fields directly from potential fields, bypassing computationally expensive propagator calculations.
• To enhance the speed and scalability of large-scale SCFT simulations.

Main Methods:

• Integration of various neural network models into the SCFT framework.
• Direct prediction of density fields from potential fields using machine learning.
• Comparative analysis of different neural network architectures for performance evaluation.
• Focus on Gaussian chain models forming disordered, microphase-separated structures.

Main Results:

• A robust and computationally efficient machine learning model for SCFT was developed.
• The model achieves speedups of over 3x for similar system sizes and up to 100x for larger systems.
• Demonstrated the feasibility of using deep learning to accelerate SCFT simulations.
• The developed methods show potential for extension to more complex and computationally demanding models.

Conclusions:

• Machine learning offers a viable strategy to significantly improve the efficiency of SCFT simulations.
• The developed approach alleviates computational bottlenecks in predicting polymer thermodynamics and self-assembly.
• This work paves the way for more extensive and complex simulations in polymer science.