Machine learning for ULCF life prediction of structural steels with synthetic data

Abstract

Machine learning (ML) has gradually developed into an effective method for fatigue life prediction. However, training an accurate and robust ML model is challenging when the data points resulting from expensive tests are scarce. This study proposed and validated of using synthetic ultra-low cycle fatigue (ULCF) life data generated by tabular generative adversarial network (GAN) as input for ML models. The ULCF life prediction using synthetic data was conducted for structural steels through artificial neural network (ANN) and multi-fidelity deep neural network (MFDNN). The results demonstrated that the ANN_Syn model trained with original experimental data plus synthetic data possessed higher mean value and lower standard deviation of R² in ULCF life prediction, compared to the ANN model trained with original experimental data only. The pioneeringly constructed MFDNN model with synthetic data by tabular GAN can generally have good predictive performance. The synthetic data size had more significant influence on the predictive ability of MFDNN model than that of ANN_Syn model. The results of this study will promote the application of ML models on ULCF life prediction of structural steels, thereby greatly reducing the cost of parameters calibration of models for ULCF damage evaluation of steel structures.