Subprojects

M2 Learn

Learning of Dynamical Process Models Based on Data and Expert Knowledge

M2 learns fast simulators of the forming process from simulation and sensor data, and adapts them to new material properties from a few examples.

Illustration. Messages from the grippers spread over the sheet mesh, then the network predicts how the sheet draws in. The green field marks the predicted draw-in. Repeating this gives a fast rollout of the forming process.

Work packages

Results of the first funding phase, by work package.

  1. M2.1 Data-driven dynamics models of production processes

    Physical systems are encoded as graphs on which message passing networks learn the dynamics. GGNS grounds predictions with point-cloud observations, ROBIN captures global deformation modes of nonlinear solids (NeurIPS 2025 spotlight), and hierarchical graph networks were applied to forming of thermoplastic composites. ASMR and AMBER adapt the mesh resolution, while M3GN and MaNGO predict whole trajectories to avoid compounding errors.

    Training and inference loops of AMBER. A message passing network predicts a sizing field on the current mesh, and a mesh generator produces the next mesh.
    Overview of the training (left) and inference (right) of AMBER, which learns adaptive mesh generation on complex geometries from an expert dataset. Figure: Freymuth et al., NeurIPS 2025.

    Publications

  2. M2.2 Action and parameter-conditioned dynamics models

    In stamp forming, the effect of a globally applied force must reach every node of the mesh at the first step, far beyond the receptive field of a shallow graph network simulator. IGNS structures message passing as port-Hamiltonian dynamics that preserve energy and momentum, keeping long rollouts physically consistent (ICLR 2026). HEPi represents grippers and a deformable sheet as a heterogeneous graph with SE(3)-equivariant message passing, the architectural basis for action-conditioned models (ICLR 2025, oral). STORM carries oscillatory message passing over to space and time, so information propagates over long ranges in spatio-temporal graphs (NeurIPS 2026).

    IGNS pipeline. An input mesh with a sphere passes through several port-Hamiltonian warm-up steps to a globally informed initial condition, followed by an information-preserving rollout over time, trained with a multi-step loss.
    IGNS. Port-Hamiltonian warm-up steps give a globally informed initial condition, followed by an information-preserving iterative rollout, trained with a multi-step loss. Figure: Hoang et al., ICLR 2026.

    Publications

  3. M2.3 Adaptation to real-world data and transfer learning

    MaNGO frames simulation as in-context learning, so a handful of trajectories of a new material condition the simulator without any gradient update (NeurIPS 2025). PEACH encodes point-cloud sequences instead of meshes. Trained only on simulated point clouds and evaluated on real observations, it adapts to unseen materials from a small context set (NeurIPS 2026).

    PEACH framework. Point-cloud sequences are encoded into a latent material description that conditions a graph network simulator, with auxiliary losses and zero-shot transfer to real observations.
    Overview of the PEACH framework. It encodes point-cloud sequences into a latent material representation that conditions a mesh-based simulator, and generalises to unseen materials and real-world data. Figure: Dahlinger et al., NeurIPS 2026.

    Publications

  4. M2.4 Simulation and physics-informed models

    PI-MGN combines message passing with a finite element residual loss that enforces the governing equations during training, and generalises to non-stationary and nonlinear problems on arbitrary meshes. The hierarchical architectures of AMBER and ROBIN are precursors for learning residual corrections to coarse classical solvers, and ongoing work uses learned models as initial guesses for Newton solvers.

    Publications

About the subproject

Our models predict the geometry and the spatially distributed state of the part. Physical knowledge keeps them accurate when data are scarce, and point-cloud observations connect them to real objects.

Principal investigators
Researchers
  • Philipp Dahlinger
  • Tai Hoang
  • Tobias Würth
Former members
  • Dr. Niklas Freymuth
Host institute
Autonomous Learning Robots (ALR), KIT
People in M2

Publications 16