S18: Advanced modelling techniques: Data-driven mechanics of materials
Numerical methods are now an important tool for studying the process and predicting the response of structures made of advanced materials such as composite materials, metamaterials, printed polymers of metals. Although complex and high-fidelity physically based models exist, they are hampered by their large computation cost, which has motivated the development of model-order reduction, data-driven analyses, and surrogate models, such as:
Reduced-order model in which the reduced number of unknown variables is defined by means of proper orthogonal decomposition and in which a further order reduction
called hyper-reduction is conducted in order to reduce the computation cost of the internal variable.
Model-free methods aiming at exploiting experimental or synthetic data without relying on a constitutive law. This is also the spirit of the so-called Data-driven computational mechanics approach.
Deep material network (DMN) approach set as a homogenization method based on analytical micromechanics models defining mechanistic building blocks organized in a hierarchical topological structure and whose parameters are defined from a training step using off-line simulations.
Surrogate models built using machine learning tools, such as neural-networks, which are trained using synthetic database built from off-line simulations
Although the field has seen an increasing activity in the last couple of years, there are still many challenges to be tackled, such as:
Accounting for history dependency in a robust way and, related to this, the management of phenomenologic internal variables in existing models
Linking the structural response to micro-structures and or process properties in an efficient way.
Conducting efficient stochastic analyses
Reaching real time evaluation for the development of digital twins