Prediction of aluminum alloy mechanical properties using synthetic data generated by generative adversarial networks

Abstract

Machine learning models are widely used to predict the mechanical properties  of aluminum  alloys.  However,  their  accuracy  is  often hindered   by  the  scarcity  of high-quality   tensile  test  data,  as experimental   data  collection  is  costly  and  time-consuming.  To address this limitation, this study employs Generative Adversarial Networks  (GANs)  to  generate   synthetic   tensile  test  data  for aluminum alloys, improving the accuracy of predictive models. The dataset consists of 200 real samples containing the compositions of nine chemical elements and two mechanical properties-Yield Strength (YS) and Ultimate Tensile Strength (UTS). A trained GAN model   was  used  to  generate   1,000  synthetic  samples,  whose statistical similarity to the original dataset was validated using the Kolmogorov-Smirnov (KS)  test and Pearson  correlation  analysis. The results confirmed  that all synthetic variables  retained  similar distributions and correlation patterns to the original dataset. To evaluate the impact of synthetic data on predictive  accuracy,  three machine learning algorithms-Random Forest Regressor (RF), Gradient Boosting Regressor (GBR), and Ada Boost Regressor (ABR)-were tested under two training schemes: (1) synthetic data for training and real data for testing and (2) real data for both training and testing. The RF model showed the highest improvement in UTS prediction,  with reductions of 38.3% and 46.3% in Mean Absolute Error (MAE) and Root Mean Square Error (RMSE),  respectively. The GBR model exhibited notable enhancements in YS prediction, with MAE and RMSE reductions of22.5% and 28.3%. These results demonstrate that GAN-generated  synthetic data is highly effective in  improving  machine  learning  predictions   of aluminum  alloy properties, particularly when experimental data is limited.