Data-driven material homogenisation, selection and design
Lead Research Organisation:
Swansea University
Department Name: College of Engineering
Abstract
Data-driven homogenisation techniques present an innovative and highly promising approach that will enable material modelling to match the fast pace of the current advances in material design and engineering. With such modelling strategies, the challenging and often tedious task of formulating the appropriate constitutive equations for new material designs is superseded by a neural network surrogate model constructed through a learning phase on the appropriately designed Representative Volume Element (RVE) of material. As a result the computer modelling of material behaviour and design becomes more straightforward and much more efficient. This is particularly important in the context of advanced micro-structure design and novel additive layer manufacturing techniques as well as the for the understanding and prediction of material degradation and life cycle studies. This project will integrate significant recent developments in multi-scale modelling, computational homogenisation and machine learning, thus providing a powerful digital twin based on the latest advances in material engineering with the aim of accelerating the material research, development and design.
People |
ORCID iD |
| Djordje Peric (Primary Supervisor) | |
| Eugenio Muttio Zavala (Student) |
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| EP/T517987/1 | 01/10/2020 | 30/09/2025 | |||
| 2442202 | Studentship | EP/T517987/1 | 01/10/2020 | 30/09/2023 | Eugenio Muttio Zavala |
| Description | The project has not finished yet, then, the following paragraphs describe the most significant achievements obtained at this stage. The outcomes of this project have an academic impact across different disciplines within engineering, including constitutive modelling of materials, computational mechanics, machine learning and optimisation. The main focus of this project is the development of a novel data-driven strategy for material modelling capable of reproducing the complex nonlinear constitutive relations of solid materials including elastic behaviour, plastic deformation and damage mechanics. This strategy provides an alternative to the complex task of formulating constitutive equations for new materials due to the capability of learning directly from data. In this project, it is proposed a Recurrent Neural Network (RNN) architecture that has been derived from the mathematical expressions of elastoplasticity, providing an agreement with the classical material formulations and improving the physical interpretation. The network architecture has been proven to be exact in one-dimensional elastoplasticity with hardening. This outcome has been published in a conference paper for the World Congress on Computational Mechanics 2022, and an oral presentation has been given at the European Congress on Computational Methods in Applied Sciences and Engineering 2022. A training strategy based on gradient-free optimisation has been also developed confirming the exact rendering of 1D elastoplastic models, while multiaxial cases are approximated accurately. The constitutive modelling network architecture, data generation, training and results obtained are presented in a forthcoming publication that is under review at this moment. Due to the performance observed in the proposed training strategy, an additional publication describing the most important mechanisms has been prepared and will be submitted to a specialised optimisation journal in the following weeks. The proposed training strategy is characterised by a supervised gradient-free ensemble optimisation algorithm. After achieving adequate results from the network training, the next step is to use the data-driven material model as a surrogate model in computational simulations. Hence, the trained network has been implemented into a finite element code on a Gauss point level replacing a traditional library of algorithmic constitutive models. A compact expression to compute the elastoplastic modulus is proposed, which depends on the network architecture selected. The numerical simulations obtained are intended to be presented at a conference this year. |
| Exploitation Route | The most relevant findings in data-driven material modelling should be taken as a reference to extend to more areas in engineering. As shown in the forthcoming publications, the proposed approach can easily be extended to areas where experimental data is available. Computational mechanics research groups can utilise the proposed methodology and extend it to different applications. In a non-academic environment, the future objective of this project is to provide engineers with tools that rely on accurate experimental data and accelerate the specialised analysis required in engineering design. |
| Sectors | Aerospace, Defence and Marine,Construction,Energy,Manufacturing, including Industrial Biotechology,Other |
| Description | UKAEA |
| Organisation | Culham Centre for Fusion Energy |
| Country | United Kingdom |
| Sector | Academic/University |
| PI Contribution | Development of a PhD project involving outputs such as publications, presentations at conferences, and software development. |
| Collaborator Contribution | The collaboration included feedback for the project, mentoring and involvement at the research centre. |
| Impact | Publications at conference DOI: 10.23967/wccm-apcom.2022.079 A second publication with respect to data-driven material modelling is currently under review. A third publication concerning an optimisation strategy is preparing to submit. |
| Start Year | 2020 |