Optimisation of manufacturing process parameters for variable component geometries using reinforcement learning

Abstract

Tailoring manufacturing processes to optimum part quality often requires numerous resource-intensive trial experiments in practice. Physics-based process simulations in combination with general-purpose optimisation algorithms allow for an a priori process optimisation and help concentrate costly trials on the most promising variants. However, considerable computation times are a significant barrier, especially for iterative optimisation. Surrogate-based optimisation often helps reduce the computational effort but surrogate models are typically case-specific and cannot adapt to different manufacturing situations. Consequently, even minor problem variations e.g. geometry adaptions invalidate the surrogate and require resampling of data and retraining of the surrogate. Reinforcement Learning aims at inferring optimal actions in variable situations. In this work, it is used to train a neural network to estimate optimal process parameters (“actions”) for variable component geometries (“situations”). The use case is fabric forming in which pressure pads are positioned to optimise the material intake. After training, the network is found to give meaningful parameter estimations even for new geometries not considered during training. Thus, it extracts reusable information from generic process samples and successfully applies it to new, non-generic components. Since data is reused rather than resampled, the approach is deemed a promising option for lean part and process development.