Realistic aggregate based on rough textures with deep learning

Model construction is a critical element in simulating the mesoscale structure of concrete. This study introduces an advanced deep learning approach that excels in generating highly realistic aggregates with diverse and heterogeneous surface roughness. The methodology leverages a synergy between Gaussian Process Regression (GPR) and Convolutional Neural Networks (CNNs) to imbue textures with remarkable variations, all originating from a single source texture. Initially, aggregates are randomly generated through the cell fracture technique, and subsequently, the Catmull–Clark subdivision algorithm is applied to refine and smoothen their surfaces. The integration of displacement mapping, in conjunction with the generated texture, results in the creation of intricately textured rough surfaces. Significantly, this research presents an innovative technique for introducing a wide spectrum of surface roughness variations to aggregates, greatly enhancing our ability to construct highly accurate mesoscale models for concrete. This groundbreaking development not only enriches the domain of concrete modeling but also clears a path for elevated levels of precision and sophistication in the realm of concrete structure analysis.