Physics-Guided Generative Inverse Design of Thermally Anisotropic Microstructures via FEM-Validated Conditional GANs

Area of Science:

• Materials Science
• Computational Science
• Engineering

Background:

• Inverse design problems in materials science are complex due to nonlinear relationships between structure and properties.
• Existing data-driven generative models for microstructure design face challenges in maintaining physical consistency and scalability.

Purpose of the Study:

• To develop a physics-guided generative inverse design framework for creating microstructures with specific physical properties.
• To address the limitations of current methods in ensuring physical consistency and practical application.

Main Methods:

• Combined finite element method (FEM) simulations with a conditional Wasserstein generative adversarial network (GAN).
• Enforced physical laws through FEM-based data generation and closed-loop FEM re-simulation.
• Utilized FEniCS, an open-source Python FEM platform, for automated, scalable parallel execution.

Main Results:

• Generated microstructures with prescribed directional thermal conductivities, achieving relative errors typically below 5-10%.
• Demonstrated the framework's ability to preserve key geometric characteristics without explicit constraints.
• Successfully applied to the inverse design of thermally anisotropic microstructures with parallel elliptical pores.

Conclusions:

• Physics-guided generative modeling offers a practical and flexible approach for inverse microstructure design.
• The framework effectively bridges data-driven methods with physical laws for complex engineering problems.
• Validated FEM simulations confirm the accuracy and reliability of the generated microstructures.