M3 Optimise

State- and Parameter-Space Exploration and Process Optimisation

M3 finds good process settings with as few trials as possible, using what can be observed between process stages.

Illustration, computed live. A Gaussian process models part quality from a few trials, and expected improvement picks the position of the highlighted gripper for the next trial. Vermilion marks the wrinkles for the setting shown; they smooth out where the gripper holds the sheet.

Work packages

Results of the first funding phase, by work package.

  1. M3.1 Simulated benchmark environment for multi-stage exploration and optimisation

    A simulation package for multi-stage systems defines subprocesses with their inputs, outputs and noise, and connects them into a directed acyclic graph. It works with scalable Bayesian optimisation packages such as BoTorch and with GPU-accelerated inference.

    Diagram of a process P made of two subprocesses in series. Each subprocess has its own inputs and outputs, and the output of the first feeds the second. Enlarge figure
    A multi-stage process as a chain of subprocesses, each with its own inputs and observable outputs. Figure: KI-FOR 5339.
  2. M3.2 Multi-stage Bayesian optimisation

    The Partially Observable Gaussian Process Network (POGPN) treats the output of each node as latent and only partially observed, so it propagates uncertainty correctly across the whole process network and models noisy intermediate observations. It is trained with variational inference and runs with BoTorch.

  3. M3.3 Open-loop trajectory exploration and optimisation

    Data from a non-isothermal glass moulding process served as a real multi-stage test case, evaluated with a new validation strategy that combines stratified sampling and bootstrapping. Joint parameter and state-space (JPSS) modelling extracts latent features from intermediate time series and consistently outperformed direct input-output models. The JPSS paper received the Best Student Paper Award in the AI for Industry track at IEEE INDIN 2024.

  4. M3.4 Robust multi-stage exploration and optimisation

    A trust-region method based on Fisher information works with any model that has a smooth, differentiable posterior. It comes with a theoretical analysis of vanishing gradients in high-dimensional Bayesian optimisation (NeurIPS 2026).

  5. M3.5 Joint state-space and parameter-space exploration and optimisation

    POGPN outperformed conventional Gaussian process networks and Bayesian optimisation on high-dimensional benchmark functions and a penicillin production simulation, with both noise-free and noisy observations. Combined with JPSS, it was tested on high-dimensional time series of a multi-stage bioethanol process.

  6. M3.6 Application to the physical composite stamp-forming process

    In a high-fidelity Abaqus forming simulation, Bayesian optimisation of gripper positions and forces reduced out-of-plane wrinkling of carbon-fibre thermoplastic laminates, and trust-region methods placed an additional gripper. A sensitivity analysis showed that gripper positions matter more than gripper forces.

    Publications

A result, jointly with M1, T1 and T2

Random gripper layout

Layout found by Bayesian optimisation

Simulated stamp forming with eight grippers. Colour shows surface curvature, where wrinkles form. The optimised layout reduces wrinkling along the edges. From Doehner et al., Composites Part A, 2026.

About the subproject

Every real trial is expensive, so each one should be as informative as possible. Bayesian optimisation proposes the next trial by weighing what the model already knows against where it is still uncertain.

Principal investigators
Researchers
  • Saksham Kiroriwal
Host institute
Fraunhofer IOSB
People in M3

Publications 8