Hybrid deep learning model to predict the ultimate tensile strength of friction stir welded joints

Accurate prediction of tensile strength in friction stir welded joints is essential to improve the reliability of welded structures. Traditional Artificial Neural Networks (ANNs) often struggle to capture the complex relationships between welding parameters, leading to limited predictive accuracy. To address this challenge, a hybrid model was developed, combining Long Short-Term Memory (LSTM) and dense layers to leverage the ability of LSTM to learn sequential patterns and dependencies between inputs such as rotational speed, feed rate, and tool tilt angle. This hybrid model was compared to a traditional ANN model using a dataset containing welding parameters and tensile strength results. Performance indicators, including correlation coefficient and Mean Square Error (MSE), showed that the LSTM + Dense hybrid model achieved a higher correlation coefficient (0.9803) and lower MSE on the validation sets, indicating improved predictive ability. Analyses of error distribution and convergence of training and validation losses also showed that the hybrid model had higher consistency and reliability with reduced overfitting, even when dealing with outliers. The LSTM + Dense model was shown to significantly outperform the traditional ANN, making it a powerful tool for predicting tensile strength in welding applications with complex and nonlinear data. Although the accuracy and reliability of the proposed hybrid model have improved, limitations remain, including the need for larger datasets to enhance generalization and potential challenges in real-time implementation.