T2 Simulate

Physics-based Process Evaluation and Decision Support for Structural Process Improvement

T2 builds validated simulations of the process that test changes before they are built and supply data that sensors cannot measure.

Illustration. A hot laminate of four tape plies is formed while grippers hold its ends; lines and dots show fibres along and across the section. The fibres barely stretch, so the plies slide against each other on the walls, where the marker lines through the stack become stepped. Colour shows temperature, which drops where the laminate touches the tools. Runs differ in their settings, such as the gripper forces (arrows). Where one gripper holds back more, it draws in less and the laminate slides towards it, so each run ends in a different part. Every run is added to the simulation data from which M2 learns.

Work packages

Results of the first funding phase, by work package.

  1. T2.1 Development of a high-fidelity process simulation CAE chain

    A hexahedral solid-shell element that predicts thickness changes in forming simulation was extended to rate-dependent material behaviour and applied to stamp forming of thermoplastic tapes. Interpolation methods for fibre orientation tensors reduce information loss along the CAE chain. The SimEnvironment generates simulation data for the M subprojects automatically.

    Pipeline diagram. Process parameters from a project partner pass through a Python interface into the SimEnvironment. There, the process definition sets up a high-fidelity forming simulation that draws on process simulation knowledge: the macroscopic simulation model, material behaviour subroutines and solver dependencies. The simulation evaluation returns the results to the partner through a Python interface. Enlarge figure
    Data generation pipeline for stamp-forming simulations with the SimEnvironment. Figure: KI-FOR 5339.
  2. T2.2 Development of low-fidelity simulation models

    Physics-informed models of the thermal behaviour can partly be trained without simulation data. Fast Python simulations of static problems and nonlinear solid mechanics support the development of learned surrogates, as do simplified forming models and dedicated datasets, including one for a simplified double-dome benchmark.

    Training loop of PI-MGNs. A graph built from the current temperature solution and the properties of the time step passes through the network, which predicts the next time steps. The finite element method computes the error of these predictions per element and test function, and the resulting loss is backpropagated to the network. Enlarge figure
    Training of physics-informed MeshGraphNets (PI-MGNs). The current solution and the properties of the time step form a graph, from which the network predicts the next time steps. The finite element method computes the error of these predictions per element and test function, and the mean squared error of all errors updates the network by backpropagation. Figure: Würth et al., CMAME 2024.
  3. T2.3 Validity assessment towards CAE chain enhancement

    A study of the material parameters of the solid-shell element identifies those that drive the global forming behaviour and underlines the relevance of rate-dependent modelling. A covariance-based validation framework assesses whether a model reproduces the relevant system dynamics. Its numerical studies show when isothermal models suffice and when thermomechanical models are needed. The simulation data are openly available.

    Four scatter plots of correlation against variance of the wrinkling measure, one per total process time from 1.6 to 6.4 seconds. Grey dots are isothermal simulations with varied material parameters; a teal cross marks the thermomechanical reference, which lies inside the grey cluster for the two shorter times and outside it for the two longer times. Enlarge figure
    Validation metrics for thermomechanical and isothermal forming simulations at four total process times. Dots are isothermal simulations with varied material parameters, crosses the thermomechanical reference. For short process times the reference lies within the isothermal cluster; for longer ones it does not, so the isothermal model is rejected there. Figure: Mitsch et al., ESAFORM 2026.
  4. T2.4 Process evaluation and decision support for modification

    A study with T1 showed that the gripper positions clearly influence the contour, curvature and thickness distribution of the formed part. Together with T1, M1 and M3, T2 provided the high-fidelity simulation basis for the Bayesian optimisation of gripper configurations.

    Three simulated formed parts seen from above, with curvature in colour and the gripper positions marked around each part. Enlarge figure
    Simulated forming results, with curvature in colour. (a) Eight grippers in a configuration found by Bayesian optimisation. (b) Improved result with an additional, locally optimised gripper. (c) The additional gripper in a less effective position. Figure: After Döhner et al. 2026.

About the subproject

When data-driven improvement runs out, the process may still be immature. Validated simulations then show why and suggest structural changes that T1 can build.

Principal investigators
Researchers
  • Johannes Mitsch
  • Tobias Würth
Host institute
Institute of Vehicle System Technology (FAST), KIT
People in T2

Publications 22