Project
KI-LaSt
In the automotive industry, lightweight design plays a central role for significantly improving the carbon footprint of vehicles. New manufacturing technologies, such as additive manufacturing, are becoming the production process of choice for new design stategies, as they enable highly complex and ultralight component geometries using innovative lightweight materials.
However, the automotive industry is facing two major problems, especially when optimizing crash-relevant lightweight components:
1) Existing simulation methods do not meet sufficient efficiency and accuracy levels when mapping the structural behavior of these components.
2) Determining the relevant parameters for designing 3D lattice structures still requires time-consuming and resource-intensive design iterations.
In the project KI-LaSt, researchers from the Institute for Automotive Engineering (ika), the Chair of Software Engineering (SE), the Institute of Structural Mechanics and Lightweight Design (SLA), all three RWTH Aachen University, and the Cognitive Modeling group from Universität Tübingen, together with the industry experts from Ford, Dynamore and AM-Metals want to tackle these challenges. The goal of the project is to automate the selection of parameters with the help of artificial intelligence in order to accelerate the development process of additive manufactured components. In addition, the project partners want to set up a CAE process chain that maps the entire, virtual, AI-based development process.
Project Details
Project term
January 1, 2021–December 31, 2021
Affiliations
RWTH Aachen University
Institute
Chair of Software Engineering
Principal Investigator
Methods
Methods:
Lattice structures are subject of current research in various specialist areas on university and industry side. From a technical point of view, despite the enormous lightweight potential that has been propagated again and again, it has so far hardly been possible to detect relevant fields of application for lattice structures or to establish lattice structures in real applications. Although lattices are used at various points in the context of topologically optimized structures for weight reduction, the substitution or targeted use of lattice properties to generate certain properties is new and the examples are mostly limited to academic examples. Within this project, lattice structures are actively used to increase structural effectiveness and improve crash safety. Investigating this in real vehicle structures and bringing it to a prototype level can be considered a great success in the field of research of lattice structures.
Finite Element Simulations:
To achieve the project goals, especially regarding an effective training of the AI, a large amount of training data is needed. Therefore, hundred-thousands of single explicit Finite-Element simulations (crash-simulations) needed to be conducted, which could only be done with HPC. With the help of the HPC we were able to simulate several hundred thousand single structures and produce a sufficient basis of trainings data for the AI.
Artificial Intelligence:
The first AI approach was to use GNNs (Graph Neural Networks) to learn to predict performance indicators (Energy Absorption and Peak Crushing Force) on small lattice structures with the capability to generalize to bigger lattice structures. Due to the well-known out of distribution generalization problem of ML algorithms, this approach was not successful.
Results
Consequently, the conceptual planning of the project went on to a second iteration with replanning the ML approach. During this time the ML part of the project did not consume any HPC resources, so we did only use a small amount of the available resources.
Discussion
Now, the new approach is to divide the problem of constructing lattice structures into a dynamic crash case and a static load case. For the static load case we use a gradient based FEM simulation that directly optimizes the lattice’s beam thicknesses via backpropagation to a NN that maps beam ids to thicknesses. For this case no HPC is needed. For the dynamic case we plan on parametrizing the lattice structure decently but with any less options than having an arbitrary structure. This way we need again the HPC for gathering data, but directly for lattice structure simulations with a realistic size instead of assuming any kind of generalization. Afterwards, we train an NN as a surrogate model to predict the performance indicators and replace the simulation with it. This way we have a fast method to run batched simulations as an environment for RL to again train an NN as optimization model.
Additional Project Information
DFG classification: 402 Mechanics and Constructive Mechanical Engineering
Software: Python3, LS-Dyna, Hyperworks Hyperstudy, Hyperworks Hyperview
Cluster: CLAIX
Publications
Bühring J. et al., Elastic axial stiffness properties of lattice structures: Analytical approach and experimental validation for bcc and f2cc,z unit cells, Mechanics of Advanced Materials and Structures, 2022; 1-17
N. Baumann, E. Kusmenko, J. Ritz, B. Rumpe, M. Weber,
Dynamic Data Management for Continuous Retraining,
Proceedings of MODELS 2022. Workshop MDE Intelligence, pp. 359 – 366, ACM, Oct. 2022.
J. C. Kirchhof, E. Kusmenko, J. Ritz, B. Rumpe, A. Moin, A. Badii, S. Günnemann, M. Challenger,
MDE for Machine Learning-Enabled Software Systems: A Case Study and Comparison of MontiAnna & ML-Quadrat,
Proceedings of MODELS 2022. Workshop MDE Intelligence, pp. 380-387, ACM, Oct. 2022.