Abstract：Traditional mechanics models and emerging data-driven models have been widely used to predict the fatigue life of additive materials. However, traditional models represented by Continuum Damage Mechanics (CDM) suffer from low accuracy and limited applicability, while data-driven models represented by Artificial Neural Networks (ANN) cannot be applied to small-sample operating conditions. To address these issues, the knowledge-data dual-driven models that integrate physical knowledge and data information have received extensive attention in recent years. To compare and analyze the predictive capabilities of these models, this study takes the AlSi10Mg alloy in Laser Powder Bed Fusion (LPBF) as the research object. Initially, an automatically calibrated CDM model is constructed and compared with the ANN-based data-driven model under various working conditions. Furthermore, three types of knowledge-data dual-driven models based on CDM and ANN are constructed using feature fusion, parameter fusion, and output fusion methods. The performance of these models is quantitatively analyzed in terms of predictive accuracy and data requirements. The research results indicate that the parameter fusion-based model exhibits a significant effect in correcting the training data and shows minimal sensitivity to the CDM model in terms of predictive accuracy, even when the CDM model fitting results are poor. The feature fusion-based model maximizes the utilization of information from the CDM model and achieves the highest predictive accuracy when data is abundant. The output fusion-based model, which is primarily based on the results of the CDM model and corrected using ANN, demonstrates the best extrapolation performance in non-training domains. These findings provide valuable references for further developing knowledge-data dual-driven models for high-precision fatigue life prediction in additive manufacturing.
2024, Vol. 45 
