Inference of Metal Additive Manufacturing Process States via Deep Learning Techniques

Anarfi, Richard
Kwapong, Benjamin
Fletcher, Kenneth
Sparks, Todd
Flood, Aaron
Joshi, Mugdha
Journal Title
Journal ISSN
Volume Title
University of Texas at Austin

Numerical simulation of metal additive processes are computationally intensive tasks. Iterative solution techniques for physics-based methods can lead to lengthy solution times and convergence problems, particularly if fluid dynamics of the melt pool are considered. Deep learning (DL) techniques offer an opportunity to infer solution results quickly. In this paper we propose a DL method based on long short term memory (LSTM), network trained on rendered images from a metal AM process simulation and CAM data. We obtained vector representations of the images by training on an autoencoder. LSTM is a memory based recurrent neural networks (RNN) that is capable of processing long sequences of data while combating temporal stability problems encountered with conventional recurrent neural networks (RNN)s. This LSTM network is used to predict images of the process given scan path and process information. This could later be used to compare with process monitoring systems as part of a quality assurance or process control schema.