Tractable descriptors for digital aggregate generation

Abstract

This paper introduces a deep learning-based approach for generating 2D aggregate shapes using a small set of tractable and physically meaningful parameters. In contrast to existing methods that often rely on quantities without clear physical interpretations, the approach presented here leverages scale invariant moments to capture essential shape characteristics. Additionally, we introduce the idea of using scale-invariant state descriptors, such as area, to control the size of the generated shapes. The neural network is trained to generate shapes corresponding to these parameters, and its ability to learn the relationship between shape constants and state descriptors without explicit data augmentation is demonstrated. The framework thus presented provides a foundation for developing microstructure generators that offer enhanced interpretability by relying on parameters that provide meaningful insights into the description of material morphology composed of non-trivial shapes.