Within the Additive Manufacturing area of Directed Energy Deposition (DED), single clad geometry prediction has been well covered within the literature. Currently, the two accepted methodologies of geometry prediction are physics numerical simulation, and semi-empirical regression. This work seeks to add a viable alternative through machine learning techniques. Machine learning has enjoyed many successes in the past few years due to the availability of large datasets for which these techniques scale beautifully. However, in small, high variance, tabular datasets, such as most results from physical experimentation, these techniques suffer. Presented here is a selection of machine learning methodologies which are used to extract models that perform and generalize well. Neural Networks (NNs), Gaussian Process (GP) modeling, Support-Vector Machines (SVMs), and Gradient Boosted Decision Trees (GBTs) for regression and classification are explored in this paper. These four methodologies will be applied to a small dataset containing some single clad data available in the literature and previously unpublished experimental results of this author. These techniques produce models not only with good agreement with experimental data, but also non-material specific generalizable results. Lastly, a discussion of data augmentation using Generative Adversarial Networks (GANs) with preliminary results put forth to illustrate unique, exploitable advantages capable within the machine learning paradigm.
Hierarchical layered and refined grain structure of Inconel 718 superalloy produced by rolling-assisted directed energy deposition
