Session: CT-19-01 AI, Data Engineering and Data Analysis-1

Paper Number: 151613

151613 - Data-Driven Multiscale Modelling of History-Dependent Plasticity Using Recurrent Neural Operators 

Abstract: 

Multiscale modelling in materials science, which integrates and translates information across different length, has gained attention for its ability to bridge macroscopic behaviour with microscale phenomena. Concurrent multiscale models like FE², where microscale FE models are directly coupled to a macroscale model and run concurrently, are prohibitively computationally expensive. Therefore, a key question is how to develop a computationally efficient and scalable approach to build a surrogate model of the microscale behaviour that can be integrated into a coarser component-scale model. Development of surrogates from micromechanical simulations such as crystal plasticity finite element model (CPFEM) of metallic systems that carries over texture and grain morphology, requires a comprehensive framework involving microstructural characterisation, experimental calibration, sensitivity analysis, computational homogenisation, machine learning, and experimental validation. Recent advances in computational power and the availability of data have led to a transformative shift in mechanics towards data-driven approaches. However, it can be argued that combining data-driven approaches with physical principles creates a more interpretable and robust foundation for models used in structural integrity assessment. In this work, we take this physics-inspired approach and explore a potential solution using a neural network based model that leverages large datasets to learn material behaviour in the plastic regime directly from observations and/or mesoscale synthetic experimental simulations. We develop surrogate models for the plastic stress/strain response in the component scale which provide full 3D solutions under multiaxial loading and hence can be applied into common FEA frameworks. In addition to capturing multi-axial loading behaviour, a second challenge for the surrogate is that it needs to capture the effect of load history dependency at the macroscale which is addressed by learning neural operators. Neural operators have simpler architectures and significantly smaller number of parameters to capture the history compared to recurrent neural networks that cascade neural networks for each increment. They learn a continuous operator that maps functions. For mapping functions, a neural network is used in this example for elastic/plastic-stress strain relations and different neural networks are used for each internal or state variable encapsulating the history evolution. All these networks are coupled together. An important point is that the state variables are not specified priori but leaned from the data. This allows us to introduce a parametric surrogate model for microstructure sensitive modelling and simulation at the structural scale. The model is trained using the data from CPFEM simulations of equivalent Representative Volume Elements (RVEs) that represent statistical distributions of morphological and crystallographic descriptors of microstructure. The procedure of mesoscale synthetic experimental simulations for model learning based on 3D examples are explained in detail and the developed surrogates implemented in the application level. The results discuss the architecture of neural networks and the number of state variables for different classical history-dependent examples. Further the routes to quantify the uncertainties on the RVE level are examined for this type of machine learning model. The source of uncertainty discussed here is mainly model uncertainty including selection and collection of the training data, the completeness and accuracy of the training data, and the model architecture and its limitations. Uncertainty quantification underpins important decisions specially for life assessment of safety critical components. These advances in multiscale data-driven modelling open up new possibilities for more accurate, efficient, and adaptable predictive tools in material science and mechanics. Also, it helps reducing the reliance on trial-and-error calibration processes of traditional constitutive modelling.

Presenting Author: Sina Safari University of Bristol

Presenting Author Biography: Sina Safari received a PhD in Engineering from the University of Exeter in 2023. His main research interest is data-centric structural integrity and modelling of nonlinear engineering structures. He is currently a senior research associate within the SINDRI project at the University of Bristol. He helps with the digitisation of physical entities using cross-cutting AI technology that can be used to assess the condition of components of engineering structures.

Authors:

Sina Safari University of Bristol
Paul Wilcox University of Bristol
Mahmoud Mostafavi Monash University
David Knowles Henry Royce Institute