MLR based data augmentation for artificial neural network prediction of failure load in adhesive bonded GFRP single lap joints

Accurate prediction of the failure load of adhesive-bonded composite joints is essential for the reliable design of lightweight engineering structures. However, developing robust predictive models remains challenging when only limited experimental data are available. This study proposes a hybrid MLR–ANN framework in which Multiple Linear Regression (MLR) is employed as a data augmentation tool to support Artificial Neural Network (ANN) development for predicting the failure load of adhesive-bonded glass fiber-reinforced polymer (GFRP) single-lap joints. Experimental data were obtained using a Taguchi L9 orthogonal array considering overlap length, adhesive thickness, and adherend tensile strength as the principal design variables. An MLR model was first established from the experimental observations and subsequently used to generate additional data within the investigated design space, producing a complete full-factorial dataset of 27 samples for hybrid MLR-ANN development. Based on the MLR-augmented dataset, the proposed hybrid MLR–ANN framework employed an ANN with three input neurons, five hidden neurons and one output neuron, trained using the Levenberg–Marquardt algorithm. Evaluation using the original experimental data demonstrated that the proposed hybrid MLR–ANN framework achieved improved predictive accuracy over the standalone MLR model, reducing the RMSE from 0.370 to 0.276 kN, the MAE from 0.273 to 0.126 kN, and the MAPE from 5.44 to 2.14%, while increasing the correlation coefficient from 89.37 to 98.2%. These results demonstrate that the proposed MLR-based data augmentation strategy can effectively support ANN development when experimental datasets are limited.