Fast inverse design of microstructures via generative invariance networks

The problem of the efficient design material microstructures exhibiting desired properties spans a variety engineering and science applications. ability to rapidly generate that exhibit user-specified property distributions can transform iterative process traditional microstructure-sensitive design. We reformulate microstructure using constrained generative adversarial network (GAN) model. This approach explicitly encodes invariance constraints within GANs two-phase morphologies for photovoltaic applications obeying specifications: specifically, user-defined short-circuit current density fill factor combinations. Such be represented by differentiable, deep learning-based surrogates full physics models mapping properties. Furthermore, we propose multi-fidelity surrogate reduces expensive label requirements five. Our framework enables incorporation or non-differentiable fast generation (in 190 ms) with proposed physics-aware data-driven methods inverse problems used considerably accelerate field Physics-aware are tailored Multi-fidelity data create inexpensive yet accurate machine learning evaluating physics-based such frameworks.