Learning time-series data of industrial design optimization using recurrent neural networks

In automotive digital development, 3D shape morphing techniques are used to create new designs in order to match design targets, such as aerodynamic or stylistic requirements. Control-point based shape morphing alters existing geometries either through human user interactions or through computational optimization algorithms that optimize for product performance targets. Shape morphing is typically continuous and results in potentially large data sets of time-series recordings of control point movements. In the present paper, we utilize recurrent neural networks to model such time-series recordings in order to predict future design steps based on the history of currently performed design modifications. To build a data set sufficiently large for the training of neural networks, we use target shape matching optimization as digital analogy for a human user interactive shape modification and to build data sets of control point movements in an automated fashion. Experiments show the potential of recurrent neural networks to successfully learn time-series data representing design changes and to perform single-and multi-step prediction of potential next design steps. We thus demonstrate the feasibility of recurrent neural networks for learning successful design sequences in order to predict promising next design steps in future design tasks.