Fast inverse design of microstructures via generative invariance networks

Abstract

The problem of efficient design of material microstructures exhibiting desired properties spans a variety of engineering and science applications. An ability to rapidly generate microstructures that exhibit user-specified property distributions transforms the iterative process of traditional microstructure-sensitive design. We reformulate the microstructure design process as a constrained Generative Adversarial Network (GAN). This approach explicitly encodes invariance constraints within a GAN to generate two-phase morphologies for photovoltaic applications obeying design specifications: specifically, various short circuit current density and fill-factor combinations. Such invariance constraints can be represented by deep learning-based surrogates of full physics models mapping microstructure to photovoltaic properties. To circumvent data generation bottlenecks, we utilize a multi-fidelity surrogate that reduces the requirements of expensive labels by 5X. Our approach enables fast generation of microstructures (in ≈190ms) with user-defined properties. Such physics-aware data-driven methods for inverse design problems are expected to democratize and accelerate the field of microstructure-sensitive design.

Modeling the Solid Electrolyte Interphase: Machine Learning as a Game Changer?

“…Beyond the chemical rules learnt from data, prior expert knowledge can be utilized for the generation of practical battery systems. [329,330] Although most of the current application of generative models at the atomistic scale focused on small molecules, few practical cases of solid state materials have been demonstrated. [331][332][333] At the microstructure scale, generative models for solid materials have been very successful even for complex heterogeneous systems and limited datasets.…”

“…[331][332][333] At the microstructure scale, generative models for solid materials have been very successful even for complex heterogeneous systems and limited datasets. [325,[328][329][330][334][335][336][337] Generative models can be used for inverting the design process of better battery interphases by going from properties to structure-and composition-conditional generative models that learn correlations in structural space conditioned to the properties. [15] Thus, favorable SEI forming electrolyte compositions can be probabilistically generated.…”

Rechargeable Batteries of the Future—The State of the Art from a BATTERY 2030+ Perspective

“…[109,115] Beyond constraints learnt from data, additional known physical laws and constraints can be imposed on the generation. [116,117] Generation of new structures at the atomic scale can also be defined as a Markov decision process and deep reinforcement learning [118,119] working with a physics simulation environment can be used to train a structure generator. Finally, if the underlying correlations (that define the physical viability of the structure) are simple enough to be defined as heuristic rules, a logic-based generative scheme can be defined using them.…”

“…[108,121,122] On the other hand, at the microstructural scale, generative models of solid materials have been very successful even for complex heterogeneous systems and limited datasets. [109,110,[115][116][117][123][124][125] Outstanding developments in convolutional networks in image processing and generation have been key to that success.…”

A framework for data-driven solution and parameter estimation of PDEs using conditional generative adversarial networks

“…The concept of deep convolutional neural networks (CNN) has been employed for both the generator and the discriminator of GAN to capture highly heterogeneous spatial features [46][47][48] and their evolution in time [49,50]. Additionally, an augmentation of discrete labeled data has been used to control the generator's output, termed conditional GAN or cGAN [51][52][53][54]. Still, all the aforementioned frameworks cannot be used directly to address the data-driven solution of PDEs since they are limited to only categorical conditioning, which is not sufficient for the parametrization of spatially heterogeneous input fields.…”