Inducing a Realistic Surface Roughness onto 3D Mesh Data Using Conditional Generative Adversarial Network (cGAN)

In the age of machine learning, data-driven approaches with hybrid data (a mixture of real images and simulation images) are getting increasingly popular. One major issue with creating a realistic simulation for surface engineering is that the surface of the mesh model used in the simulation is smooth. Often, this mesh does not contain information on surface texture; thus, simulating an object based on these meshes may not represent an actual surface texture of a real component. This article presents a novel technique for introducing surface roughness onto a smooth mesh object to facilitate engineering simulation by using a conditional Generative-Adversarial Network (cGAN) that is trained on real height maps to generate random 2D height maps that represents a realistic texture of a typical upskin and downskin surface of an additively manufactured (AM) part. This approach extracted the past scans of AM components from the Focus Variation microscopy. The 3D surface deviation is extracted as height maps and used as the training data for the generative network. This paper will also discuss the structural similarities between the synthetic and real data using standard descriptors for surface texture characterisation, such as \(S_{a}\), \(S_{q}\) and \(S_{dq}\).