Template Design for Complex Block Copolymer Patterns Using a Machine Learning Method.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Directed self-assembly (DSA) is crucial for creating nanoscale patterns.
• Designing guiding templates for DSA is a complex inverse design problem.
• Current methods often rely on computationally expensive forward simulations.

Purpose of the Study:

• To develop a machine learning-based method for the inverse design of guiding templates for DSA.
• To predict DSA templates solely using machine learning, eliminating the need for forward simulations.
• To evaluate the performance and generalization ability of various neural network architectures.

Main Methods:

• Formulated the inverse design problem as a multi-label classification task.
• Trained various neural network (NN) models, including convolutional neural networks (CNNs) with residual blocks.
• Utilized simulated pattern samples from self-consistent field theory (SCFT) calculations for training.
• Applied augmentation techniques tailored for morphology prediction to enhance NN performance.

Main Results:

• Achieved a significant improvement in exact match accuracy for template prediction, increasing from 59.8% (baseline) to 97.1% (best model).
• Demonstrated that machine learning models can predict templates without requiring forward simulations.
• The best-performing NN model showed excellent generalization capabilities for human-designed DSA patterns.

Conclusions:

• Machine learning, specifically deep neural networks, offers a powerful and efficient solution for the inverse design of DSA guiding templates.
• The developed method significantly outperforms traditional approaches by eliminating the need for extensive simulations.
• The study highlights the potential of AI in accelerating materials design and fabrication processes.