Model Selection and Evaluation for Machine Learning: Deep Learning in Materials Processing

Abstract

Materials processing is a critical subset of manufacturing which is benefitting by implementing machine learning to create knowledge from the data mined/collected and gain a deeper understanding of manufacturing processes. In this study, we focus on aluminum high-pressure die-casting (HPDC) process, which constitutes over 60% of all cast Al components. Routinely collected process data over a year’s time of serial production are used to make predictions on mechanical properties of castings, specifically, the ultimate tensile strength (UTS). Random Forest, Support Vector Machine (SVM), and XGBoost regression algorithms were selected from the machine learning spectrum along with a Neural Network, a deep learning method. These methods were evaluated and assessed and were compared to predictions based on historical data. Machine learning, including Neural Network, regression models do improve the predictability of UTS above that of predicting the mean from prior tests. Choosing the correct models to use for the data requires an understanding of the bias-variance trade-off such that a balance is struck between the complexity of the algorithms chosen and the size of the dataset in question. These concepts are reviewed and discussed in context of HPDC.