Joint Parameter and State-Space Bayesian Optimization: Using Process Expertise to Accelerate Manufacturing Optimization
New AI framework uses expert knowledge to optimize complex manufacturing processes twice as fast.
A team of researchers from Karlsruhe Institute of Technology (KIT) and Fraunhofer IOSB has introduced POGPN-JPSS, a breakthrough AI framework designed to dramatically accelerate the optimization of complex, multi-stage manufacturing processes. The paper, submitted to CIRP CMS 2026, addresses a critical bottleneck in industrial AI: standard Bayesian Optimization (BO) treats processes as 'black boxes,' ignoring valuable intermediate observations and expert knowledge, which limits its effectiveness in high-dimensional systems like chemical or pharmaceutical production.
The proposed framework innovatively merges two concepts. First, it uses Partially Observable Gaussian Process Networks (POGPN) to model a manufacturing process as a Directed Acyclic Graph (DAG), capturing the structure between stages. Second, it integrates Joint Parameter and State-Space (JPSS) modeling. This allows the AI to incorporate low-dimensional latent features extracted by human experts from high-dimensional, time-series sensor data (the 'state-space'), rather than trying to process the raw, noisy data directly. This fusion of data-driven learning and human process expertise is the key to its performance.
The team validated POGPN-JPSS on a challenging, high-dimensional simulation of a multi-stage bioethanol production process. The results were striking: the framework found optimal process parameters to reach a desired performance threshold twice as fast and with greater reliability compared to current state-of-the-art optimization methods. For industries where process maturation cycles can take months or years and consume vast resources, this acceleration translates directly into substantial cost savings and faster time-to-market for new products. The research underscores a growing trend in industrial AI: hybrid systems that leverage human domain expertise to guide and enhance machine learning models are outperforming purely data-driven approaches in complex, real-world environments.
- POGPN-JPSS combines structured probabilistic models (POGPN) with expert-extracted features (JPSS) to optimize multi-stage processes.
- In a bioethanol production simulation, it achieved target performance 2x faster than leading alternative methods.
- The framework directly addresses the 'high-dimensional state-space' challenge by using expert knowledge to create low-dimensional latent features from sensor data.
Why It Matters
Cuts optimization time and cost for complex manufacturing (chemicals, pharma), accelerating R&D and production scaling.