Hybrid GA-ANN modelling for microhardness prediction in FSPed Mg-Y-Nd-Zr alloys using autoencoder-synthesized data

Abstract

Microhardness plays a critical role in determining the applicability of Mg alloys across various industries. Friction Stir Processing (FSP) refines the microstructure, thereby enhancing the microhardness of the processed material. However, the quantitative relationship between FSP parameters and resulting microhardness remains insufficiently understood. In this study, a machine learning approach was used to model the influence of FSP parameters viz. Tool rotation speed (rpm), traverse speed (mm/min), and shoulder diameter (mm) on the microhardness of Mg-Y-Nd-Zr alloys. The experiment dataset, designed using the Taguchi L₂₇ orthogonal array, consisted of 27 real data points and was augmented with 200 synthetic samples generated using an autoencoder-based data synthesis technique. A hybrid Genetic Algorithm-optimized Artificial Neural Network (GA-ANN) was developed for microhardness prediction. The GA-ANN model trained on the combined dataset achieved an R² score of 0.955 and a mean squared error (MSE) of 0.028, significantly outperforming the model trained on real data, which achieved an R² of 0.75. Additionally, the GA-ANN model outperformed several baseline models, including a standard ANN (R² = 0.85), linear regression (R² = 0.72), and a decision tree regressor (R² = 0.74)