Prediction of Weld Strain Using Artificial Neural Network: A Modeling Approach

• Isaac O. UKRAKPO

  Author

• Osemwegie IKPONMWOSA-EWEKA

  Author

This study employed an Artificial Neural Network (ANN) to predict weld strain and related mechanical properties in welding operations. A central composite design (CCD) generated 20 experimental data points comprising three input parameters current, voltage, and gas flow rate and their corresponding output response (strain across welds). The data were normalized between 0.1 and 1.0 prior to network training to prevent weight variation issues. Multiple training algorithms and hidden neuron configurations were evaluated to identify the optimal network architecture, with selection criteria based on mean square error (MSE) and coefficient of determination (R²). The Levenberg-Marquardt backpropagation training algorithm with 10 hidden neurons demonstrated superior performance, yielding a training MSE of 0.00023 and cross-validation MSE of 0.00011. The optimal network architecture (3-10-1) utilized tan-sigmoid and purelin transfer functions for the hidden and output layers, respectively. Network training with 60% of the data, 25% for validation, and 15% for testing achieved a performance error of 2.06e-05 at epoch 14, significantly below the target error of 0.01. The trained network exhibited strong predictive capability with a correlation coefficient (R²) of 0.9862 between predicted and observed values, demonstrating excellent reliability. These findings confirm that the optimized ANN approach provides a robust, precise, and superior predictive tool for optimizing welding parameters studied in this research to achieve desired mechanical properties, offering significant advantages over traditional empirical modeling methods.

• 42_Ukrakpo and Ikponmwosa-Eweka (2026)

Copyright (c) 2026 Isaac O. UKRAKPO, Osemwegie IKPONMWOSA-EWEKA (Author)

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.