Model-free data-driven identification algorithm enhanced by local manifold learning
Journal
Computational Mechanics
Journal Volume
71
Journal Issue
4
Start Page
637
End Page
655
ISSN
01787675
Date Issued
2023-04
Author(s)
Abstract
Reliable and consistent material data identification is essential to the data-driven computational mechanics paradigm. This paper presents a generalized data-driven identification (DDI) approach to constructing material databases of high quality. We integrate the locally convex reconstruction method into DDI to formulate the local-convexity DDI (LCDDI) method. The LCDDI method can learn the local structures of material data and thus produce structure-informed optimal material data points for solving stresses with given strains. The effectiveness of the LCDDI method at addressing large acquisitions of material data with a complex heterogeneous strain field is demonstrated through two numerical experiments: a perforated elastic plate and a center-holed elasto-plastic plate. Convergence studies show that results using the LCDDI method are dramatically improved. We further explain how the LCDDI method manages to accurately identify mechanical stress fields and a high-fidelity material database under the condition of imbalanced distribution of elastic and plastic deformation. Discussion of the LCDDI method in regards to the oversampling issue, the capability of full-domain analysis, and importance sampling is given. Finally, we conclude that the LCDDI method can extract a vast amount of material data points with improved quality from full-field strain measurements, and can serve as a more reliable technique for material data acquisition.
Subjects
Data-driven computing
Local-convexity data-driven identification
Manifold learning
Material data acquisition
Material data structure
Material database quality
Publisher
Springer Science and Business Media Deutschland GmbH
Type
journal article