iLattice: Mechanistic machine learning for interactive design of flexible lattice structures

• Additive manufacturing offers huge potential for applications of soft lattice structures, e.g., for tailored actuation, energy absorption and dissipation, vibration mitigation, ventilation, or weight reduction, e.g., in car or bicycle seats, protective pads in helmets, or shoe mid-soles and orthopedic insoles, etc.
• The potential for rapid customization and diverse applications necessitates computational tools that facilitate design of flexible lattice structures with tailored dynamic mechanical behavior.
• However, multiscale modeling is often not applicable to AM lattice structures, which lack periodicity and scale separation.
• Thus, other kinds of surrogate or meta modeling techniques are required, where in particular mechanistic ML (MML) methods are very promising for dynamic lattice surrogate models.
• Interactive design assistants should integrate such MML surrogates in order to design for multiple, possibly conflicting designs goals and consider aleatoric and epistemic uncertainties in modeling and optimization.

• Modeling, simulation, and design optimization of flexible and functional lattice structures based on nonlinear 3D beam discretizations

• Physics-aware, mechanistic machine learning for multiscale and multiphysics material modeling and surrogate modeling of dynamic systems.

The objective of this project is to develop a computational assistant for the interactive design of flexible lattice structures with highly nonlinear and dissipative dynamic mechanical behavior. Specific goals are to develop and implement:

• Accurate, efficient, and robust nonlinear surrogate models for lattice unit cells and structures using mechanistic machine learning (MML)
• Computational design optimization of lattice structures enabled by MML surrogates
• Interactive design assistant (IDA) for flexible lattice structures

Illustration of the surrogate modeling pipeline, in which data from 3D beam RUC simulations is used to train surrogate models of RUCs. These are formulated using MML as stable port-Hamiltonian NNs in terms of only few (nodal) DOFs. The RUC surrogates are then assembled into a lattice structure surrogate. This dynamic system can be efficiently integrated to obtain the 3D lattice dynamics.

• The to be developed MML technique will be generally flexible and modular in terms of the actual microstructures and dynamic behaviors to be modeled. Thus, it could also be applied within other projects, e.g., for soft robotic or multibody systems.
• Data and methods could contribute to benchmarking of surrogate modeling techniques and become part of a larger benchmark data set for dynamic surrogate modeling
• IDA modules for Bayesian and gradient-based optimization could also be adapted to other dynamic systems and support other types of design processes.