Artificial neural network-based prediction model of residual stress and hardness of nickel-based alloys for UNSM parameters optimization

Abstract

Ultrasonic Nanocrystal Surface Modification (UNSM) is known as one of the surface treatment techniques that utilizes an ultrasonic vibration energy to improve mechanical properties and performance of materials. The dynamic nature of this surface treatment process deforms the top surface and subsurface of materials with a very high strain rate, which makes the direct observation of residual stress and refined layers very difficult. In this study, a novel alternative approach is proposed that is based on the artificial neural network (ANN) concept for predicting residual stress and hardness of various nickel-based alloys that have been subjected to UNSM treatment. Experimental measurement data were used in the ANN training process and validation. The trained model showed the capability of predicting residual stress and hardness accurately with a Pearson correlation value (R) of 0.988 and 0.996, a root mean squared error (RMSE) of 84.231 and 17.028, and a mean absolute error (MAE) of 68.586 and 13.450 for residual stress and hardness models when tested using the test dataset, respectively. It can be concluded that ANN as the alternative approach is a suitable method for accurately performing prediction for practical use in the absence of a mathematical model. Since the experimental result was used in the ANN model training process, the predicted result by the ANN model appears to agree with the experimental results of the UNSM treatment. Because of these demonstration results, the ANN-based prediction model can be used as a tool to optimize the UNSM treatment parameters.