Related Experiment Videos
CTGAN-Based Data Augmentation and XGBoost-LSTM Strength Prediction of CSG
Guanghui Li1,2, Yupeng Zhang1, Qingqing Tian1,3
1School of Water Conservancy, North China University of Water Resources and Electric Power (Longzihu Campus), Zhengzhou 450046, China.
Materials (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces a data augmentation technique using Conditional Tabular Generative Adversarial Networks (CTGAN) to enhance cementitious sand and gravel (CSG) datasets. The improved data boosts the performance of a hybrid XGBoost-LSTM model for predicting CSG mechanical properties.
Area of Science:
- Materials Science and Engineering
- Civil Engineering
- Data Science and Machine Learning
Background:
- Cementitious sand and gravel (CSG) is a vital construction material.
- Traditional CSG mix proportion design is complex and faces experimental limitations like long cycles, high costs, and external factor interference.
- Obtaining high-quality CSG sample data for accurate mix design and performance prediction is challenging due to small sample sizes.
Purpose of the Study:
- To address the challenge of insufficient CSG sample data.
- To develop and validate a data augmentation method for CSG.
- To propose and evaluate a hybrid machine learning model for predicting CSG mechanical properties using augmented data.
Main Methods:
- A foundational dataset of 100 CSG specimens was created through physical experiments, testing compressive and splitting tensile strength.
- Conditional Tabular Generative Adversarial Networks (CTGAN) were employed for data augmentation, generating 100 synthetic samples.
- A hybrid XGBoost-LSTM model was developed, utilizing XGBoost for feature construction and LSTM for sequential learning and prediction.
Main Results:
- The CTGAN method generated synthetic data highly consistent with original CSG data, outperforming comparative methods in quality evaluation (Wasserstein distance).
- The XGBoost-LSTM model achieved improved prediction accuracy after data augmentation, with R2 values of 0.9897 for compressive strength and 0.9801 for splitting tensile strength.
- The hybrid model demonstrated statistically significant superior performance over baseline models (XGBoost, LSTM, RF, SVR) across all evaluation metrics.
Conclusions:
- The CTGAN-based data augmentation effectively expands small CSG datasets, improving the reliability of mechanical property prediction.
- The hybrid XGBoost-LSTM model offers a robust and accurate approach for predicting CSG compressive and splitting tensile strength.
- This study provides a valuable reference for data enhancement and performance prediction strategies for other materials with limited sample data.