Accelerating the design of lattice structures using machine learning.

Area of Science:

• Engineering
• Materials Science
• Computational Science

Background:

• Lattice structures offer design versatility but rapid, optimized mechanical property design is challenging.
• Increasing design variables lead to intractable design spaces, necessitating efficient computational methods.
• Existing machine learning (ML) approaches face challenges in model interpretation and efficient training data curation.

Purpose of the Study:

• To develop an interpretable ML framework for accelerated lattice structure design.
• To identify key design variables influencing mechanical properties using ML and explainability techniques.
• To enhance the efficiency of training data curation for optimization tasks in lattice design.

Main Methods:

• Combined ML-based surrogate modeling with Shapley Additive Explanation (SHAP) for variable interpretation.
• Utilized active learning methods, specifically Bayesian optimization, for efficient design space exploration.
• Developed an intelligent system integrating ML for design variable discovery and acceleration.

Main Results:

• ML-based surrogate models demonstrated high prediction accuracy (R² > 0.95).
• SHAP analysis effectively identified design variables impacting lattice structure performance.
• Active learning reduced simulation requirements by 5x compared to grid-based search.

Conclusions:

• The integration of ML and SHAP provides a powerful tool for interpreting design variables in lattice structures.
• Active learning strategies significantly improve the efficiency of the design and optimization process.
• Intelligent design systems leveraging ML are crucial for accelerating the development of tailored lattice structures.