Accelerating polymer self-consistent field simulation and inverse DSA-lithography with deep neural networks.

Area of Science:

• Polymer Science
• Computational Chemistry
• Materials Science

Background:

• Self-consistent field theory (SCFT) is vital for studying block copolymer (BCP) self-assembly.
• SCFT simulations are computationally expensive, hindering applications requiring extensive forward simulations.

Purpose of the Study:

• To accelerate SCFT simulations using a deep learning-based approach.
• To reduce the computational cost of BCP self-assembly studies.
• To enhance the efficiency of inverse design problems in BCP self-assembly.

Main Methods:

• A deep neural network (DNN) was trained to predict equilibrium polymer structures from early SCFT iteration outputs.
• The DNN model replaced the computationally intensive forward simulation in inverse design methods.
• The method was validated on 2D and 3D bulk systems and applied to inverse directed self-assembly-lithography.

Main Results:

• The DNN accurately predicts equilibrium states from limited SCFT iterations, significantly reducing simulation time.
• The number of initial SCFT iterations can be optimized for speed-accuracy trade-offs.
• Training set size impacts DNN performance, providing guidance for cost-effective dataset generation.
• The DNN-accelerated inverse design method improved efficiency by 100-fold, eliminating SCFT simulations.

Conclusions:

• Deep learning offers a powerful strategy to accelerate SCFT simulations for BCP self-assembly.
• This method substantially reduces computational demands for both forward and inverse design problems.
• The approach facilitates more efficient exploration of complex polymer self-assembly phenomena and material design.