Page 61 - IJAMD-2-1

International Journal of AI for
Materials and Design
Fatigue life prediction via contrastive learning

Currently, with the development of advanced intelligent life of 316L and 304 stainless steels, demonstrated strong
algorithms, data-driven methods have been widely generalization capabilities. Chen et al. proposed a
28
developed and applied in the field of material fatigue life multiview neural network model incorporating frequency
prediction. 11-18 From classical shallow machine learning domain analysis. This model integrates convolutional
algorithms to deep learning models based on neural neural networks (CNN), long short-term memory
networks, these methods have demonstrated excellent networks, and FNet with frequency domain analysis in
performance in various fatigue-related problems. 19-24 Jiang a parallel structure, extracting effective features from
et al. proposed a physics-informed multilayer nested the material loading path to predict fatigue life. Through
25
neural network framework, using stacking fault energy, ablation experiments, the extrapolation capability of the
strain amplitude, strain rate, and temperature as input model was verified using specific test datasets. Zhang et al.
29
features. Physical constraints were embedded in the loss used a SHapley Additive exPlanations-informed recursive
function to ensure the model adhered to known physical feature elimination method to identify key features in a
laws. The performance of the model was validated on multiaxial fatigue experiment dataset. Symbolic regression
fatigue data of 316 stainless steel. Zhu et al. developed was employed to extract and encapsulate expressions
26
a deep learning model called Multi-GAT (Multi-Graph predicting fatigue life based on these salient features, which
Attention Network) for predicting the high-cycle fatigue were then integrated into the traditional mean squared
(HCF) life of titanium alloys. This model is integrated error (MSE). This significantly improved the predictive
with an attention mechanism and uses a graph structure accuracy of the model on the existing database.
as the data structure, allowing the full consideration of
relationships between samples. This approach enables In fatigue life prediction, constructing appropriate input
the accurate prediction of HCF life for various titanium features is crucial for improving the model’s predictive
alloys using a limited number of sample data. Liao et al. performance. 30-33 Simply using raw experimental features
27
proposed a path-dependent adaptive physics-informed as inputs may not lead to optimal results. To achieve
neural network to address the non-proportionality caused effective feature engineering, researchers have drawn
by phase differences. The model embedded multiple on various traditional empirical models, such as critical
critical plane models into the loss function and achieved plane models, 34,35 damage mechanics models, 36,37 and
optimization through dynamic weight adjustment. Genetic fracture mechanics models. 38,39 These empirical insights
algorithms and a meta-learning framework were used to help in selecting and designing more relevant features,
optimize the weight hyperparameters. The meta-learning which significantly enhance the model’s ability to predict
framework enabled the weights of different physical terms fatigue life accurately. Figure 1 provides an intuitive
in the loss function to dynamically adapt based on the load comparison between traditional physics-based models and
path information. The explored meta-learning framework, data-driven approaches. In traditional methods, domain
applied through transfer learning to predict the fatigue experts manually extract key features using physics-

Figure 1. A comparison of fatigue life prediction frameworks: physics-based model versus deep learning model

Volume 2 Issue 1 (2025) 55 doi: 10.36922/IJAMD025040004