Machine learning predicted inelasticity in defective two-dimensional transition metal dichalcogenides using SHAP...

Area of Science:

• Materials Science
• Computational Materials Science
• Nanotechnology

Background:

• Crystallographic defects in 2D transition metal dichalcogenides (TMDCs) offer property tuning but can degrade mechanical reliability.
• Understanding the impact of defects, particularly vacancies, on mechanical properties is crucial for reliable applications.
• Machine learning (ML) shows promise for modeling materials and uncovering structure-property relationships, though it's an emerging field.

Purpose of the Study:

• To analyze the mechanical properties of pristine and defected 2D TMDCs using ML and deep learning.
• To predict key mechanical indicators like failure stress and strain to failure.
• To investigate the influence of chirality and strain on the mechanical behavior of 2D TMDCs.

Main Methods:

• Extensive molecular dynamics simulations were conducted on 2D TMDCs with various crystallographic defects.
• XGBoost and densely connected neural network (DenseNet) algorithms were employed for predictive modeling.
• Shapley value analysis was used to enhance the interpretability of the ML models.

Main Results:

• Accurate, state-of-the-art predictions of mechanical properties were achieved using both XGBoost and DenseNet models.
• The study successfully correlated material structure, including defects and chirality, with mechanical performance.
• Comparative evaluation highlighted the predictive capabilities and interpretability of the employed ML techniques.

Conclusions:

• ML and deep learning are effective tools for analyzing the mechanical properties of defected 2D TMDCs.
• The findings provide a pathway for tuning material properties by controlling defects for specific applications.
• Improved model interpretability through Shapley values aids in understanding structure-property relationships for reliable material design.