Deep Learning to Speed up the Development of Structure-Property Relations For Hexagonal Boron Nitride and Graphene

Area of Science:

• Materials Science
• Computational Materials Science
• Artificial Intelligence in Materials

Background:

• Structure-property maps are crucial for materials discovery but often rely on computationally intensive simulations.
• Deep learning offers a potential alternative to reduce computational costs in predicting material properties.

Purpose of the Study:

• To explore the use of deep learning agents, specifically Convolutional Neural Networks (CNNs) and Multilayer Perceptrons (MLPs), for predicting structure-property relations in materials.
• To investigate the efficiency of transfer learning for adapting these models to new 2D materials.

Main Methods:

• Simulated hexagonal boron nitride (h-BN) microstructures with varying radiation and temperature damage were used.
• Low-dimensional physical descriptors were developed to represent microstructural defects.
• Deep learning models were trained to predict residual strength from atomic positions.
• Transfer learning was employed to adapt models to other 2D materials.

Main Results:

• Purpose-specific microstructure representation using physical descriptors enabled accurate predictions at low computational cost.
• Deep learning agents achieved high prediction accuracy (≈95% R²).
• Transfer learning required significantly less data (≈10% or less) compared to training from scratch (23-45%) for new materials.

Conclusions:

• Deep learning, particularly with optimized microstructure representations, can accelerate materials discovery by reducing simulation dependence.
• Transfer learning is a highly efficient approach for adapting predictive models to new 2D materials, offering substantial data savings.