Design, Development, and Testing of Machine Learning Models to Estimate Properties of Friction Stir Welded Joints

This paper estimates friction stir welded joints' ultimate tensile strength (UTS) and hardness using six supervised machine learning models (viz., linear regression, support vector regression, decision tree regression, random forest regression, K-nearest neighbour, and artificial neural network). Tool traverse speed, tool rotational speed, pin diameter, shoulder diameter, tool offset, and tool tilt are the six input parameters in the 200 datasets for training and testing the models. Deep learning artificial neural networks (ANN) exhibited the highest accuracy. Therefore, the ANN approach was used successfully to estimate the UTS and the hardness of friction stir welded joints. Additionally, the relationship of pin diameter, tool offset, and tool rotation speed over UTS and hardness were extracted over the collected data points. Furthermore, experimental results, such as UTS and hardness of steel-magnesium-based welded joints and model estimated results, were compared to cross-check model generalization capability. It was noted that ANN estimates and experimental results at desired processing conditions are consistent with sufficiently high accuracy.