Application of Image-Based Deep Learning to Functional Materials R&D
|
황진하 홍익대 Deep learning has been continuously evolving towards unprecedented applications and methodologies, due to synergistic integration of big data, algorithm enhancement, and ultra-fast parallel computing environment. Deep learning ignites significantly substantial academic and industrial interests, in association with domain knowledge across science and technology. Deep learning can be exploited to two primary tasks, i.e., classification and regression. The complementary approaches of neural networks and deep learning have been expanded initially to applications such as object recognition in image modes, speech recognition, natural language processing, information retrieval, and event detection in videos. Recently, materials community has been reporting novel applications in simulation and analytics sectors. The current work places its main emphases on image-based deep learning encountered in materials domain R&D sections: in particular composite materials belonging to next-generation energy resources and 3D printing in terms of quantitative microstructure and defect detection/quantification. In this work, the electrode composites and loess-based materials are chosen as a model system, in association with the future fuel cell and 3D printing technologies, respectively. Deep learning algorithms are applied to automatic microstructure characterization in solid oxide fuel cells (SOFCs) and detection/quantification of cracks in the loess/water composites. The synergic integration of deep learning and traditional analytical methodology will be demonstrated with the aim to obtaining insights into the facile approaches on analytical issues indispensible in materials R&D domains. |