A semi-supervised approach to architected materials design using graph neural networks. (November 2020)

• Record Type: Journal Article
• Authors: Guo, Kai; Buehler, Markus J.
• Is Part Of: Extreme mechanics letters. Volume 41(2021)
• Publication Date: 2020-11
• Subjects: Architected material -- Artificial intelligence -- Machine learning -- Graph neural network -- Semi-supervised learning -- Materiomics -- Multiscale mechanics -- Mechanics -- Truss -- Structures; Mechanics -- Periodicals; Mechanics, Applied -- Periodicals; Mechanics; Electronic journals; Periodicals; 531.05
• DOI: 10.1016/j.eml.2020.101029
• Languages: English
• ISSNs: 2352-4316

Abstract

Abstract: Recent breakthroughs in artificial intelligence (AI) afford opportunities for new paradigms for material design and optimization. For modeling-driven design approaches, the optimization of mechanical properties, in general, requires boundary value problems (BVPs) to be solved. Machine learning (ML) models, trained on high-throughput labeled data obtained from the solution of BVPs, are able to explore and exploit vast design spaces. Nevertheless, prior to the implementation of those methods, applied load and displacement constraints have to be known beforehand. Here, we present a semi-supervised approach to design topological structures of architected materials based on the load levels of only 1% of nodes, along with the connectivity and mechanical properties of the architected materials. Graph neural networks (GNNs), which can learn graph embeddings and show outstanding performance on semi-supervised classification tasks using small amount of data, have been used to predict the distribution of the load levels of the remaining 99%. The integration of the network with an algorithm to redistribute truss thickness enables us to perform multiscale design of architected materials under various geometries and loads. This work reveals the potential of a novel paradigm to design architected materials via semi-supervised learning, and inspires applications such as using sparse sensors in truss designs in additive manufacturing, architectures and civil infrastructure under complex loading conditions.