Fraunhofer Institute for Mechanics of Materials IWMThe ML4P project developed a suite of tools that covers the entire ML lifecycle - from data collection to robust operation. Demonstrations on real-world systems - such as a glass bending machine, a press hardening line, and a membrane filter production line - showed how AI can optimize production processes.
Flagship Project “ML4P – Machine Learning for Production”
Completed research project
Project description
Market needs and challenges
Production facilities hold significant potential for optimization that can be tapped using machine learning (ML) methods. In practice, however, implementation often fails due to a lack of standards, insufficient tool support, and the complexity of industrial processes. Data is often fragmented, expert knowledge is not machine-readable, and the path from proof of concept to robust production remains an insurmountable hurdle for many companies.
The project’s approach
ML4P developed a tool-supported process model that structures ML projects in production into six clearly defined phases, from goal setting to long-term operation. The model is complemented by the ML4P tool suite: a service-oriented software architecture with modular components for data collection, interactive analysis, model training, and ML deployment. Three industry-relevant demonstrators - water filter production, glass bending, and press hardening - validated the results under real-world conditions.
Contribution to addressing the challenges
The ML4P toolsuite is in productive use at SUEZ in Bitterfeld. At the Fraunhofer IWM’s glass bending facility, geometric accuracy was improved from ±2 mm to ±0.5 mm. On the Fraunhofer IWU’s press hardening demonstrator, ML-based process control enables real-time adjustment of process parameters.
Fraunhofer IWM’s work packages in the project:
Fraunhofer IWM retrofitted the existing laser glass bending system with additional sensors (thermocouples, thermal camera), expanded the PLC interfaces, and established a local time-series database. The result is a fully digitized, ML-capable existing system that serves as a blueprint for retrofitting industrial systems.
A complete information model and a digital process file were created for the glass bending system.
Based on process and simulation data, ML models (Markov Random Fields, Gaussian processes) were developed and integrated into the system via the ML4P pipeline. The result is a system that suggests optimal laser programs to the operator with a geometric accuracy of ±0.5 mm instead of ±2 mm, depending on the current furnace temperature.
Based on the project results, Fraunhofer IWM can offer industrial companies the following research and development services:
1. ML readiness analysis and sensor retrofitting of existing systems
Industrial companies receive a roadmap for digitalizing their systems without costly new purchases and with a clear focus: on the data sources relevant to ML. To this end, Fraunhofer IWM analyzes existing production systems for their suitability for ML applications and develops a customized concept for retrofitting them with appropriate sensor technology and for connecting them to digital data infrastructures.
2. Creation of information models and digital process records
Fraunhofer IWM develops structured, machine-readable information models for production facilities and processes. These form the basis for consistent data utilization throughout the ML lifecycle. Companies benefit from a centralized, constantly updated knowledge base that can be used for ML projects as well as for maintenance and engineering.
3. Development and integration of ML models for incremental forming processes
Based on process and simulation data, Fraunhofer IWM develops data-efficient ML models (e.g., Gaussian processes) for complex forming processes such as glass bending, metal forming, and additive manufacturing.
4. Integration of expert knowledge into ML models (gray-box approach)
Companies receive more robust, physically consistent models that deliver reliable quality predictions even when data is scarce - which is typically the case before the start of production. To achieve this, Fraunhofer IWM applies methods for systematically integrating physical and process-related expert knowledge into ML models.
5. Development of ML-based process monitoring and control
Fraunhofer IWM designs and implements complete ML pipelines for real-time process monitoring and adaptive process control - from data acquisition and model deployment to writing optimized parameters back into the plant control system. This enables industrial companies to respond to process fluctuations in real time, reduce scrap, and reliably meet quality targets even under varying conditions.
