Estimating optimum process parameters in textile draping of variable part geometries - A reinforcement learning approach

Fine-tuning of manufacturing processes for optimum part quality requires many resource-intensive trial experiments in practice. To reduce the experimental effort, physics-based process simulations in conjunction with optimisation algorithms can be applied, e.g. finite-element-models and evolutionary algorithms. However, they generally require considerable numerical expertise and long computation times. Efficient optimisation of such expensive-to-evaluate models often employs surrogate-based optimisation (SBO). SBO constructs numerically inexpensive approximations of the original model, which guide the optimiser in the parameter space. This allows concentrating costly simulations on the most promising regions. While SBO significantly reduces the computational load in many cases, current SBO-strategies are inevitably problem-specific and cannot be reused in other, even similar situations. Consequently, subtle problem variations, e.g. minor geometry changes in material forming, require an entirely new optimisation and all previous numerical effort is in vain. Thus, surrogate techniques with generalised applicability are an open field of research. ... mehrMachine Learning techniques using convolutional neural networks (CNNs) are capable of ‘learning’ complex system dynamics from data. In this work, CNNs are used to extend the predictive capabilities of SBO towards variable instead of fixed manufacturing settings. Specifically, material draw-in optimisation in textile forming (‘draping’) for variable geometries is studied. Using reinforcement learning, a CNN is trained to estimate optimum positions of pressure pads during draping of a pre-specified class of box-shaped geometries. Once trained, the CNN interprets a forming result and infers beneficial pad positions. Unlike conventional SBO strategies, it can also give recommendations for variable geometries from the selected geometry class. The paper shows that, in principle, CNNs are able to extract information from a range of different forming tasks and apply it to a new, unknown situation. Since they reuse information gained from previous simulations, they are considered a viable option for future, generalised SBO-strategies.