Generating Realistic Nanorough Surfaces Using an N-Gram-Graph Augmented Deep Convolutional Generative Adversarial Network

Modeling and simulation of roughness generation and functionality can be aided by synthesized nanorough surfaces that mimic real experimental ones. We can save time and resources in optimizing the linkages in the process-surface-functionality triangle if the synthesized samples are generated in a computationally inexpensive manner. Existing nanorough surface generation techniques presuppose that the structural feature space to be employed in the generation process may be identified. In many circumstances, however, this assumption cannot be safely confirmed. As a result, a data-driven approach appears to be a viable option worth considering. Generating synthesized nanorough surfaces in the context of a multi-physics simulation requires (1) identifying the structural feature space so that the generation of new nanorough surfaces is possible and (2) the generation process to be property-preserving. In this work, we present methods for integrating new nanorough surfaces similar to a preset sample of surfaces into multi-physics simulations in a computationally inexpensive fashion. We look at how a Generative Adversarial Network (GAN)-based strategy, given a nanorough surface data set, may learn to produce nanorough surface samples that are statistically equivalent to the ones belonging to the training data set. We also look at how combining the GAN framework with a variety of nanorough similarity measures might improve the realisticity of the synthesized nanorough surfaces. We showcase via multiple experiments that our framework is able to produce sufficiently realistic nanorough surfaces, in many cases indistinguishable from real ones.