Developing a machine-learning based data-driven constitutive model for lattice-structured materials

The continuous progress in engineering and scientific advancements has led to a growing demand for materials with superior properties and enhanced performance. To meet these demands, a sophisticated framework or methodology is essential. Such a framework should support (i) the design of advanced materials, (ii) the development of robust models to investigate the unique behaviour of these materials, and (iii) the development of cutting-edge frameworks that not only accelerate computational performance but also facilitate practical, real-world applications.

Designing advanced materials is crucial to achieving unique mechanical behaviours and performance characteristics that are not found in traditional materials. One notable area of focus in this field is lattice-structured materials, a class of advanced engineered materials known for their unique mechanical properties. These properties include light weight, exceptional energy absorption capabilities, high stiffness, and superior strength. Moreover, these mechanical characteristics can be tailored by adjusting various geometrical parameters, making lattice-structured materials highly versatile for industries such as biomedical, automotive, aerospace, and sports equipment.

To investigate the unique mechanical properties of lattice-structured materials effectively, sophisticated constitutive models are crucial. Traditional phenomenological models often fail to capture the complexities of these materials due to their intricate geometries, complex deformation mechanisms, and macroscopic behaviours exhibited by lattice-structured materials. In contrast, advanced computational approaches, such as machine learning (ML)-based data-driven constitutive models, offer superior capabilities in this regard. These methods outperform traditional phenomenological models, providing precise insights into lattice-structured materials while also enhancing computational efficiency by saving time and resources. Recent research has leveraged the power of the ML model to integrate with commercial finite element method (FEM) software which significantly boosts computational performance and practical industrial implementation

This research thesis aims to develop an ML-based data-driven constitutive model for lattice-structured materials. A hybrid data collection methodology was employed to generate a comprehensive dataset for training an artificial neural network (ANN)-based constitutive model. This model incorporated the linear elastic and hyperelastic properties of parent materials. The ANN-based model demonstrated remarkable predictive performance in capturing the mechanical behaviour of lattice-structured materials, successfully validating its performance across diverse loading tests. Notably, it maintained exceptional predictability even when tested with unseen data or beyond the training range.

Based on this, a framework was developed to integrate the ANN constitutive model with commercial FEM software. This integrated framework aimed to enhance computational performance for numerical analyses as compared to standard ABAQUS/CAE simulations. The framework was rigorously evaluated across diverse loading scenarios, ranging from macroscale simulations to general 3D loading conditions, consistently showcasing exceptional computational performance.

Furthermore, the integrated framework was applied to real-world, three-dimensional macroscale industrial case studies, including bone scaffolding, automotive bumpers, and protective helmets. In each case, the framework demonstrated impressive computational efficiency and highly accurate numerical analyses compared to standard ABAQUS/CAE results. These findings bring out the potential of the developed framework to revolutionise the design and analysis of various industry applications and also have the prospective for widespread industrial adoption.