Transparent AI-supported process modelling in drop forging

• Development of the virtual process environment for die forging
• Determination of the influences of material and process parameters on the geometry characteristics and thus on the part quality
• Optimisation of the effective surface design of 2-stage forming processes

The aim of the research project is to improve quality characteristics in closed-die forging processes by increasing understanding of the complex collective of interactions. Data-based models in combination with global explainability methods should help to identify previously misunderstood correlations and process fluctuations so that approaches can be developed that allow process stabilisation in future generations through adapted die face design, even with lower safety factors such as flash or oversizes. The first step is to consolidate and condense the known domain knowledge regarding the relationships between process and target variables in hot forging. This serves to evaluate the model quality and to derive the explainability of the developed models. To generate sufficient process data, new and existing sensors are integrated into an inline-capable data processing and database infrastructure to enable the acquisition, processing and exchange of all relevant process data. The component geometry as an optimisation variable is extracted from 2D image data using an automated image processing method. Extensive reference data sets are then generated in series forging tests and supporting numerical process simulations with the targeted introduction and elimination of disturbance variables and with varied manipulated variables for the process under consideration. This data is used to generate an equivalent model using machine learning methods, which predicts quality characteristics based on process data. As there are various different methods in ML that can be used to predict quality characteristics, several learning algorithms must be tested and compared against each other in the project so that the algorithm that allows the highest possible prediction quality of different quality criteria based on the process variables and also makes it easier to interpret the results later is identified. A white box model is then created to identify process variables that influence component quality. Various explainability algorithms must be compared with each other. The aim is to obtain an explainable white-box model that has maximum fidelity to the original model and can be used to identify correlations between process parameters and geometric features.

The expected outcome of the research project is the development of a novel data-driven method for optimising quality characteristics in the two-stage drop forging process. The basic idea of this method is that machine learning processes can be trained in such a way that the optimum tool geometries can be achieved more precisely. Both numerical and experimental data are used for the machine learning process.