Prediction of stainless steel mechanical properties using synthetic data generated by wasserstein generative adversarial networks with gradient penalty
|
Full Text |
Pdf
|
|
Author |
Maimuzar, Desmarita Leni, Sir Anderson and Yuda Perdana Kusuma
|
|
e-ISSN |
1819-6608 |
|
On Pages
|
558-570
|
|
Volume No. |
21
|
|
Issue No. |
9
|
|
Issue Date |
July 10, 2026
|
|
DOI |
https://doi.org/10.59018/052663
|
|
Keywords |
generative adversarial networks, synthetic data, mechanical properties, stainless steel, machine learning.
|
Abstract
Machine learning has become one of the most widely developed methods for analyzing, predicting, and simulating the mechanical properties of materials. In practice, experimental data on stainless steel remain limited, especially when considering variations in heat treatment and chemical composition. This limitation poses a major challenge in developing robust predictive models. This study aims to address the scarcity of experimental data by applying a Generative Adversarial Network (GAN) model to generate synthetic data that exhibit characteristics similar to real data. Based on an analysis of 16 chemical variables using Pearson correlation, eight variables were identified as the most influential on mechanical properties, namely Cr, Ni, Mo, Si, Mn, P, N, and Cu. The evaluation results indicate that most variables achieve KS values below 0.10, reflecting a high level of similarity between the distributions of synthetic and real data. This study further evaluates the synthetic data using machine learning prediction models to estimate Yield Strength (YS) and Ultimate Tensile Strength (UTS), where 986 synthetic samples were used as the training set and 100 real experimental samples were randomly selected as the test set. The machine learning models employed include Linear Regression, Decision Tree, Random Forest, K-Nearest Neighbors, Gradient Boosting, and Artificial Neural Network. The results show that Gradient Boosting and K-Nearest Neighbors achieved the best performance, with R² values of 0.75 and 0.69, respectively. These findings indicate that the WGAN-GP–based approach can effectively overcome limitations in experimental data availability while accelerating the development and prediction of stainless steel mechanical properties.
Back