Optimisation of manufacturing process parameters using deep neural networks as surrogate models

TLDR

Optimising manufacturing process parameters requires a resource‑intensive search in a high‑dimensional space, and physics‑based simulations, while useful, are computationally expensive. The study applies surrogate‑based optimisation to a composite textile draping process. Surrogate‑based optimisation is employed, using a deep neural network trained on FE simulation data to predict shear angles of over 24,000 textile elements, guiding the search for optimal draping parameters. The approach reduces the number of expensive FE simulations needed and yields a better overall solution than previously known, thanks to detailed predictions that enhance model quality.

Abstract

Optimisation of manufacturing process parameters requires resource-intensive search in a high-dimensional parameter space. In some cases, physics-based simulations can replace actual experiments. But they are computationally expensive to evaluate. Surrogate-based optimisation uses a simplified model to guide the search for optimised parameter combinations, where the surrogate model is iteratively improved with new observations. This work applies surrogate-based optimisation to a composite textile draping process. Numerical experiments are conducted with a Finite Element (FE) simulation model. The surrogate model, a deep artificial neural network, is trained to predict the shear angle of more than 24,000 textile elements. Predicting detailed process results instead of a single performance scalar improves the model quality, as more relevant data from every experiment can be used for training. For the textile draping case, the approach is shown to reduce the number of resource-intensive FE simulations required to find optimised parameter configurations. It also improves on the best-known overall solution.