Accelerating Heterogeneous Multiscale Simulations of Advanced Materials Properties with Graph‐Based Clustering. Issue 2 (10th December 2020)

• Record Type: Journal Article
• Authors: Vassaux, Maxime; Gopalakrishnan, Krishnakumar; Sinclair, Robert C.; Richardson, Robin. A.; Coveney, Peter V.
• Is Part Of: Advanced theory and simulations. Volume 4:Issue 2(2021)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2020-12-10
• Subjects: clustering -- graph theory -- model reduction -- multiscale modeling -- splines -- unsupervised learning; Science -- Simulation methods -- Periodicals; Science -- Methodology -- Periodicals; Engineering -- Simulation methods -- Periodicals; Engineering -- Methodology -- Periodicals; 507.21
• DOI: 10.1002/adts.202000234
• Languages: English
• ISSNs: 2513-0390

Abstract

Abstract: Heterogeneous multiscale methods (HMM) capable of simulating asynchronously multiple scales concurrently are now tractable with the advent of exascale supercomputers. However, naive implementations display a large number of redundancies and are very costly. The macroscale model typically requires computations of a large number of very similar microscale simulations. In hierarchical methods, this is barely an issue as phenomenological constitutive models are inexpensive. However, when microscale simulations require, for example, high‐dimensional molecular dynamics (MD) or finite element (FE) simulations, redundancy must be avoided. A clustering algorithm suited for HMM workflows is proposed that automatically sorts and eliminates redundant microscale simulations. The algorithm features a combination of splines to render a low‐dimension representation of the parameter configurations of microscale simulations and a graph network representation based on their similarity. The algorithm enables the clustering of similar parameter configurations into a single one in order to reduce to a minimum the number of microscale simulations required. An implementation of the algorithm in the context of an HMM application coupling FE and MD to predict the chemically specific mechanical behavior of polymer‐graphene nanocomposites. The algorithm furnishes a threefold reduction of the computational effort with limited loss of accuracy. Abstract : Multiscale simulations methods have the potential to predict the properties of novel complex materials. Properties could be predicted from the chemical structure and used by engineers from the aeronautical or automotive industry avoiding expensive certification. However, multiscale simulations remain untractable, even for the largest supercomputers. The novel reduction algorithm detects redundancies within multiscale simulations and enables substantial acceleration.