DeepFlames: Neural network-driven self-assembly of flame particles into hierarchical structures

Abstract

Abstract The spontaneous assembly of materials from elementary building blocks is one of the most intriguing natural phenomena. Conventional modeling relies physical approaches to examine such processes. In this paper, a framework is proposed to offer an alternative paradigm, via the use of deep learning, and specifically the use of generative adversarial models as well as a combination of natural language processing and transformer neural nets to create hierarchical assemblies of building blocks. We study the assembly of elementary flame particles into hierarchical materials with features across scales, illustrating the Universality–Diversity Principle (UDP), and create novel material using additive manufacturing.

Graphical abstract