Research

    Porous structures play a significant role in both synthetic fuel production and the hydrogen economy by facilitating key processes such as catalysis, adsorption, and separation. Technologies like electrolyzers, fuel cells, reformers, chemical reactors, and energy storage systems rely on the development of highly efficient multiphase micro- and nano-scale porous materials (Fig. 1). 

    The characterization of those structures requires advanced techniques as computer tomography (CT) and focused ion beam – scanning electron microscopy (FIB-SEM) (Fig. 2). Moreover, understanding the correlation between performance and microstructural features, and subsequently optimizing these materials, requires the novel fabrication techniques, advanced heat transfer designs, and comprehensive performance evaluations through measurements and simulations. 

    Machine learning has the potential to advance our understanding of various physical processes. However, grasping the intricate relationship between microstructures of energy devices, their performance, and degradation is challenging due to structural complexity and experimental limitations. To address this, we're employing machine learning techniques, including convolutional neural networks (CNNs). These methods enhance experimental measurements, such as segmenting and improving resolution for complex 3D data obtained from FIB-SEM and CT scans. We've also developed approaches for reconstructing 3D models directly from 2D cross-section images using generative adversarial networks (GANs) (Fig. 3A) and fabricating artificial structural models by conditional GANs and diffusion models (Fig. 3B). Leveraging CNNs, we can predict complex microstructural properties directly from 2D images (Fig. 3C).

  Traditionally, machine learning models require substantial training data, yet such information isn't always available. Considering, for example, the prediction of material degradation, wherein the expected lifespans extend over multiple years, yet empirical data are scarce. To address this challenge, we've introduced physically constrained unsupervised image translation (UNIT) networks (Fig. 3D) and long-short term memory networks (LSTMs) to replicate degradation phenomena. Our focus lies in refining machine learning techniques by integrating numerical simulations with limited experimental data and employing physically informed neural networks, adhering to the principles of grey-box models.