Fast inverse design of microstructures via generative invariance networks

Area of Science:

• Materials Science
• Computational Engineering
• Photovoltaics

Background:

• Efficient design of material microstructures with targeted properties is crucial across many engineering fields.
• Traditional microstructure-sensitive design is often iterative and time-consuming.
• Rapid generation of microstructures with user-specified properties can significantly enhance design workflows.

Purpose of the Study:

• To reformulate microstructure design using a constrained generative adversarial network (GAN) model.
• To develop a method for rapidly generating two-phase morphologies for photovoltaic applications with user-defined performance characteristics.
• To incorporate physics-based constraints into a data-driven inverse design framework.

Main Methods:

• Utilized a constrained generative adversarial network (GAN) model to encode invariance constraints.
• Developed differentiable, deep learning-based surrogates for physics models to map microstructures to photovoltaic properties.
• Proposed a multi-fidelity surrogate approach to reduce data labeling costs by fivefold.

Main Results:

• Successfully generated microstructures with user-defined short-circuit current density and fill factor combinations.
• Achieved rapid microstructure generation in as little as 190 milliseconds.
• Demonstrated the ability to incorporate complex, non-differentiable constraints into the design process.

Conclusions:

• The proposed physics-aware, data-driven framework accelerates the inverse design of microstructures with desired properties.
• This approach transforms the iterative nature of microstructure-sensitive design.
• Enables faster development of advanced materials for photovoltaic and other applications.