Speaker
Description
Limited experimental data remain a major challenge in developing reliable machine learning models for composite material design and optimization. This study investigates the use of Generative Adversarial Network (GAN)-based data augmentation to enhance predictive modeling capabilities under small-data conditions. Conditional Tabular GAN (CT-GAN) and Copula-GAN were employed to generate synthetic datasets from experimentally obtained material and processing parameters. The generated data were evaluated using statistical distribution analysis, correlation assessment, and hypothesis testing to verify their consistency with the original experimental dataset. The augmented datasets were subsequently used to train and evaluate representative machine learning models for predicting nanoparticle distribution, a key indicator influencing composite material performance and quality. Particular attention was given to the effect of synthetic data quality on model accuracy, robustness, and interpretability. Bayesian optimization was employed to improve predictive performance, while SHapley Additive exPlanations (SHAP) were used to identify the most influential processing variables governing nanoparticle distribution. The results demonstrate that GAN-based data augmentation significantly improves predictive capability when experimental datasets are limited, with the highest performance achieved using an optimized gradient-boosting model trained on statistically consistent synthetic data.
| Topic Area | Design Optimization, Additive Manufacturing Technologies & Applications |
|---|