Session: 12-04-02: Drucker Medal Symposium

Paper Number: 146634

146634 - Machine Learning-Augmented Parametrically Upscaled Damage Model for Microstructural Damage Sensing in Piezoelectric Composites 

This paper will discuss a digital twin for damage sensing in piezocomposite structures. It is based on a two-way (bottom-up and top-down) multiscale modeling framework coupling deformation, damage, and electric fields of structures that are characterized by piezoelectric composite microstructures. The bottom-up or hierarchical multiscale modeling is achieved by developing a Parametrically Upscaled Coupled Constitutive Damage Model (PUCCDM). The PUCCDM is a thermodynamically consistent, upscaled multi-physics constitutive damage model coupling evolving mechanical deformation and electrical fields with a second-order damage tensor, with explicit dependence on its underlying material microstructure. The microstructure for these structures consists of unidirectional piezoelectric fibers distributed nonuniformly in a passive epoxy matrix that can undergo different damage mechanisms such as interfacial debonding, crack kinking, propagation into the matrix, etc. The PUCCDM incorporates the nonuniform microstructural morphology in its coefficients through Representative Aggregated Microstructural Parameters or RAMPs, representing lower-scale descriptors of piezoelectric composite microstructural morphology. The key microstructural morphology features that govern the macro-scale behavior are delineated by (a) the fiber volume fraction and (b) the fiber spatial distribution. Optimal expressions for RAMPs are determined through principal component analysis of the two-point correlation functions. Micro-electromechanical analysis of microstructure-based statistically equivalent representative volume elements (M-SERVEs) is conducted under different loading conditions to generate a microstructural response database (MRDB). The PUCCDM coefficients in terms of the RAMPs are determined using machine learning tools operating on the MRDB. The developed PUCCDM is used for structural scale analysis of composite structures for understanding concurrent damage and failure at multiple scales. From analysis using PUCCDM it is observed that due to the electromechanical coupling in the piezocomposites, damage states in the structures are strongly correlated with its electrical response. Therefore, an electric signal can be used as a proxy indicator of the damage state.

This leads to the top-down modeling framework that can quantitatively predict microstructural damage mechanisms from the measurement of a macroscopic electric signal and its corresponding RAMPs. An advanced machine learning model (ConvLSTM) based on the combination of a convolutional neural network (CNN) and a recurrent neural network (RNN) is developed for this purpose. This augmented neural network is advantageous over conventional machine learning models due to its capability to treat path-dependent nonlocal material response data. The trained machine learning model shows good damage prediction capabilities and good generalization characteristics with respect to validation and test data. Thus, the digital twin model can be used as a quantitative global damage indicator in near-real time rather than a location-specific qualitative damage measure.

Presenting Author: Somnath Ghosh Johns Hopkins University

Presenting Author Biography: Professor Somnath Ghosh is the Michael G. Callas Professor in Civil & Systems Engineering, and a Professor of Mechanical Engineering, and Materials Science & Engineering at Johns Hopkins University. He is a world leader in the field of Mechanics of Materials at multiple scales. He is the founding director of the Center for Integrated Structure-Materials Modeling and Simulations (CISMMS). He is the co-PI/co-director of the NASA Space Technology Research Institute for Model-based Qualification and Certification of Additive Manufacturing (2023-2028). He was the PI/director of Air Force Center of Excellence in Integrated Materials Modeling from 2012-2018.

Authors:

Somnath Ghosh Johns Hopkins University