Research

Every trial on a new manufacturing process is expensive. We develop AI methods that learn more from each trial, and a methodology that takes a process from immature to mature with fewer of them.

The problem

A new process first runs in the lab, then ramps up to industrial scale. Only then does it deliver the required quality, output and cost. This takes long when new materials are involved, when the process has many settings, and when no good model of it exists.

Maturing such a process means choosing its structure, its settings, its control and its sensors. Experts do this by experiment, guided by experience. Each experiment ties up equipment, uses material and may require destructive testing of the part.

We expect AI to speed up maturation and to find better solutions. Few, expensive and noisy samples also demand new methods in AI itself.

Demanding materials

Fibre-reinforced thermoplastics are stiff along the fibres and soft across them. Stacked in layers, they deform in ways that depend on temperature and speed.

Many coupled settings

Temperatures, speeds, forces and gripper positions interact, and tuning the conventional settings alone does not prevent wrinkles.

Costly, late feedback

Defects such as wrinkles show up only after forming. Every trial costs machine time and material, and some parts must be destroyed to test them.

Every change starts over

A new material or geometry usually means repeating the experiments. What was learned on one variant rarely carries over to the next.

Our approach

Methods that need few samples

We combine data-driven learning, optimisation and reinforcement learning with physical knowledge and engineering models, and we plan each experiment so that it teaches us as much as possible.

A methodology for maturation

Temporary over-instrumentation, a modular process and step-by-step improvement turn an immature process into a mature one, with tools that scale to large teams.

Across all stages

F Maturity and knowledge Maturity measures and a knowledge graph of the process data

  1. Process chain

    T1 Construction kit The demonstrator process and its experiments

  2. Data acquisition

    T2 Physics-based simulation Forming simulations of the demonstrator process M1 Over-instrumentation Which sensors to add, and which to remove

  3. Learned simulation

    M2 Learned process models Fast process models, learned from these data

  4. Optimisation and control

    M3 Exploration and optimisation Process settings, optimised between trials M4 Estimation and control State estimation and control during a trial

The demonstrator

The participating institutes built the example process in the Karlsruhe Research Factory: non-isothermal stamp forming of thermoplastic composite sheets, an immature process with a high potential for improvement. The stamp-forming laboratory of Fraunhofer ICT is used for further experiments.

Illustration. In the press, the grippers follow the laminate as it is drawn into the die. Colour shows the laminate temperature.

  1. Pick up

    A sledge on a linear axis picks up the laminate in a gripper frame. The grippers measure force, position and angle.

  2. Heat

    An infrared oven melts the thermoplastic matrix.

  3. Form

    A press with two heatable tools forms the hot laminate, which solidifies as it cools.

  4. Inspect

    A robot takes images and 3D scans of the part. The data are stored, and their metadata are catalogued in the Virtual Process Dossier.

The stamp-forming line with gripper frame, infrared oven and press on a linear axis, with an industrial robot in the background.
Gripper frame (left), infrared oven (centre) and press (right) on the linear axis.

Why it is hard

The tapes are strongly anisotropic. They deform with large strains, depend on temperature and forming speed, slide against each other, and shrink as they solidify. The press forms the parts only partly, which is enough to study quality precursors, in particular wrinkle formation.

The material

In the published forming experiments on industrial-scale equipment, the laminates were unidirectional carbon-fibre tapes in a polyamide 6 matrix, for example twelve plies in a (0/90/45/−45/0/90)s sequence. Such tapes promise light parts that can be formed automatically in short cycle times and recycled (Zeeb et al. 2025).

Results so far

In the first funding phase we developed the core components of the methodology and validated each of them on its own: in forming experiments, forming simulations, benchmarks and external process datasets.

  • Material characterisation parameterised the physics-based forming simulations, and temperature measurements from forming experiments were used to validate them. The simulations then supplied data for learning and for testing gripper configurations that were never built.

    T1 T2

  • A learned simulator predicts stamp forming of thermoplastic laminates for unseen process settings in seconds instead of hours.

    T2 M2

  • Bayesian optimisation used the validated simulation to find gripper layouts that reduce wrinkling.

    T1 T2 M1 M3

  • M4 developed particle-flow filters, calibrated sensor models, sampling-based model predictive control and, with M2, geometry-aware reinforcement learning for distributed actuators. Each was validated on simulation benchmarks.

    M2 M4

  • The Virtual Process Dossier, a knowledge-graph data catalogue, records the demonstrator's experiments, simulations and their provenance and can be queried via SPARQL.

    F

All publications

Research data

Process data are expensive to produce, so we keep them findable and reusable, following the FAIR principles. The Virtual Process Dossier describes machines, materials, settings, processes and parts with shared vocabularies and records who produced which data, when and how.

We publish data where we can, for example the process simulation data for thermoforming on Zenodo. Large simulation campaigns and model training run on the HoreKa supercomputer at KIT.