Predicting elastic and plastic properties of small iron polycrystals by machine learning

Area of Science:

• Materials Science
• Computational Materials Science
• Solid Mechanics

Background:

• Crystalline material deformation exhibits complex system behavior.
• Small samples show stochastic responses and significant variation in mechanical properties.
• Understanding this variability is crucial for materials design and application.

Purpose of the Study:

• To investigate the predictability of mechanical properties (shear moduli and yield stresses) in small iron polycrystals.
• To establish a correlation between initial polycrystalline structure and mechanical response.
• To assess the efficacy of machine learning in predicting material behavior.

Main Methods:

• Utilized Voronoi tessellation to generate a large set of small, cube-shaped iron polycrystals.
• Employed molecular dynamics simulations to obtain stress-strain curves for each sample.
• Trained a convolutional neural network to map microstructural features to mechanical properties.

Main Results:

• The convolutional neural network successfully predicted the shear modulus with higher accuracy than the yield stress.
• Demonstrated a quantifiable relationship between the initial polycrystalline structure and mechanical properties.
• Highlighted the sensitivity of the material's response to initial structural perturbations.

Conclusions:

• Machine learning, specifically convolutional neural networks, can predict mechanical properties of crystalline materials from their structure.
• Shear modulus prediction is more robust than yield stress prediction due to underlying physical mechanisms.
• Further research into the sensitivity of deformation to initial conditions is warranted for complex materials.