Machine learning-enabled multiscale modeling platform for damage sensing digital twin in piezoelectric composite structures

Nondestructive evaluation (NDE) of aerospace structures plays a crucial role in their successful operation under harsh environments. Most NDE methods, however, lack real-time in-situ predictive capabilities of evolving damage and are conducted in a post-mortem manner. This paper proposes a damage-sensing digital twin (DT) for piezoelectric composite structures incorporating microscale morphology and mechanisms into structural damage to address this shortcoming. The DT framework consists of a two-step computational process integrating multiscale-multiphysics modeling with machine learning (ML) tools to detect damage progression in the piezoelectric composite structure using electrical signal measurements at a few surface points. A parametrically upscaled coupled constitutive damage mechanics (PUCCDM) model is developed for this DT with the representation of microstructural morphology and mechanisms in macroscopic constitutive relations. Artificial neural network operates on micro-electro-mechanical data to derive PUCCDM constitutive coefficients as functions of underlying representative aggregated microstructural parameters (RAMPs). The PUCCDM model-simulated macroscopic electromechanical and damage fields, in conjunction with RAMPs, provide a comprehensive time-dependent dataset for a convolutional long-short-term memory (ConvLSTM) network to learn microstructure-dependent electrical and damage field correlations. The DT is consequently used to successfully predict location-specific damage from electrical signals at a limited number of sensors.