A machine learning (ML) framework for prediction of metal additive manufacturing (MAM) trace characteristic and process optimization.
At present, physics numerical simulation and semi-empirical regression are the two accepted methods for predicting clad/melt pool geometry in metal additive manufacturing. This study proposes a novel approach through the use of machine learning techniques for predicting the characteristics of the clad and melt pool.
A range of machine learning methodologies that can be used to develop models that perform and generalize well were explored in this work. This study employs the following ML models for regression and classification.:
- Neural Networks (NNs)
- Gaussian Process (GP) modeling
- Support-Vector Machines (SVMs)
- Gradient Boosted Decision Trees (GBTs)
We apply these four techniques to a small dataset comprising single clad data collected from the published literature. The findings if this work show that these methods generate models that not only demonstrate good agreement with experimental data but also yield non-material specific generalizable results.
Additionally, we present a discussion on data augmentation using Generative Adversarial Networks (GANs) and preliminary results that highlight unique advantages that can be exploited within the machine learning paradigm in metal additive manufacturing.
Each model was used for regression and classification tasks to predict the clad geometrical features and generate process maps based on the predictions. These process maps can be utilized for optimizing the MAM process.
Following is an example of the results generated by neural network model:
The following modules and packages are required in order to run the associated code:
- TensorFlow
- Sckiti Learn
- Keras
- Numpy
- Matplotlib