Machine learning methods for architected materials

Architected materials have a lot of promise for structural and functional applications such as lightweight structures for blast protection, efficient heat exchangers, or battery electrodes with large specific surface area. They have an intricate microstructure which is a feature that makes them both attractive to use but also difficult to design with. In the pursuit of designing large structures, the coupling between microscale and macroscale phenomena requires running finite element simulations with billions degrees of freedom. In the first part of this work, we explore graph-based machine learning as an alternative method to obtain the constitutive relationship of periodic lattice materials. We design a Physics-constrained graph neural network which can predict the stiffness tensor of periodic lattices with arbitrary crystal symmetry. The predictions are always positive semi-definite and equivariant to rigid body transformations. Once large structures are designed, a long-standing challenge has lied in their scalable manufacture. The most common additive manufacturing methods are slow and they cannot be scaled to large specimens. In the second part of this thesis, we explore self-assembly as a fast and scalable manufacturing method of inverse opals. We study vibration-induced crystallisation of balls across a range of conditions to uncover the mechanisms of crystallisation. These findings guide us to develop vibration-free epitaxial processes which can be used to build large single crystals with minimal defects. A current limitation in experimental methods is dynamic characterisation. While interrupted in-situ X-ray tomography can provide detailed 3D information, it has not been possible to achieve a similar level of detail for dynamic processes which cannot be interrupted. In the third part of the thesis, we develop a framework based on neural rendering which enables the reconstruction of 3D information for specimens undergoing complex spatio-temporal deformation. Using tomographic information at the beginning and end of deformation, we demonstrate that as few as 2 projections at intermediate timesteps are sufficient. The thesis is a combination of experiments and computational work in exploring machine learning techniques for understanding the mechanics of architected materials.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Grega, Ivan

Advisor dc:contributor.advisor

• Deshpande, Vikram

• architected materials
• granular mechanics
• graph neural networks
• machine learning
• metamaterials
• neural rendering

Licence

• https://www.repository.cam.ac.uk/bitstreams/5efb6c3a-54ed-4b5e-903c-6a6e31d735af/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

Author Identifier

0000-0002-0163-7009

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/388349