Machine learning identifies scale-free properties in disordered materials

This paper explores the use of machine learning to predict and design wave-matter interactions in disordered structures, focusing on identifying scale-free properties for waves. The authors develop disorder-to-localization (D2L) and localization-to-disorder (L2D) convolutional neural networks (CNNs) to predict wave localization and generate disordered structures, respectively. The study reveals that the generated structures exhibit scale invariance with heavy-tailed distributions, resulting in significantly improved robustness to accidental defects. The research highlights the crucial role of neural network architecture in determining the properties of machine-learning-generated structures.

Sunkyu Yu, Xianji Piao, Namkyoo Park