Page 74 - IJAMD-2-1

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

Table 3. The detailed number of each specimen and defects proposed. This method effectively extracts deep feature
for each type representations from complex multiaxial fatigue stress-
strain responses, which are then utilized in downstream
Feature RMSE of the downstream prediction model
extraction Linear ANN SVM XGBoost fatigue life prediction tasks to enhance prediction accuracy.
method The specific conclusions are as follows:
1D-CNN 0.2889 0.41946 0.76503 1.40253 (i) Compared to network architectures such as ANN,
2D-CNN 0.33526 0.54549 0.8378 1.11583 GRU, and 2D-CNN, the 1D-CNN network achieves
the best contrastive learning performance and is more
GRU 0.45183 0.42774 0.82675 1.30814 stable during training. The deep feature representations
ANN 0.88021 0.85771 1.10393 1.3852 extracted through contrastive learning, when
PLS 0.56382 0.54084 1.17852 1.55079 visualized by t-SNE, show a chaotic distribution in the
PCA 0.51162 0.36833 0.83287 1.55079 reduced-dimensional space, with no obvious clustering
Note: The values in boldface represent the lowest RMSE among the or separation of categories. However, the extracted
four downstream models under the same conditions. features enable the downstream fatigue life prediction
Abbreviations: 1D-CNN: One-dimensional convolutional neural network; model to more easily learn from samples with different
2D-CNN: Two-dimensional convolutional neural network; ANN:
Artificial neural network; GRU: Gated recurrent unit; Linear: Linear loading paths, achieving excellent performance even
regression; PCA: Principal component analysis; PLS: Partial least squares; with a simple linear regression model.
RMSE: Root mean squared error; SVM: Support Vector Machines. (ii) Compared to other unsupervised learning algorithms,
the features extracted using contrastive learning show
Table 4. RMSE results of ablation experiments on 1D‑CNN better similarity and consistency. Through comparative
with contrastive learning experiments, contrastive learning is found to be more
effective in extracting features related to fatigue life
Ablation setting RMSE of the downstream prediction
model prediction from stress-strain hysteresis data, thereby
Linear ANN SVM XGBoost helping downstream models better uncover the
Without data augmentation 0.45243 - - - underlying patterns in the data. Compared to traditional
unsupervised learning algorithms, contrastive learning
Without contrastive learning 1.63522 0.69097 1.12755 1.28143 demonstrates stronger robustness and effectiveness
Without data augmentation & 0.49226 0.78342 0.50171 1.09144 when handling multiaxial fatigue data.
without contrastive learning (iii) The feature representations learned through contrastive
With data augmentation & 0.2889 0.41946 0.76503 1.40253 learning exhibit superior predictive performance
with contrastive learning in downstream tasks. In multiple machine learning
Note: The values in boldface represent the lowest RMSE among the models, contrastive learning consistently achieves
four downstream models under the same conditions. better prediction results. Compared to scenarios
Abbreviations: 1D-CNN: One-dimensional convolutional neural
network; ANN: Artificial neural network; Linear: Linear regression; without contrastive learning, the maximum reduction
RMSE: Root mean squared error; SVM: Support vector machines; in RMSE in models such as SVM, ANN, and others
XGBoost: eXtreme gradient boosting. can reach 86.26%. In addition, the prediction stability
is improved, as evidenced by a reduction in the
Table 3 summarizes the RMSE values of different feature standard deviation of repeated experiments.
extraction methods across various downstream prediction (iv) Contrastive learning has the potential to be further
models. It was observed that the features extracted extended for applications in multiaxial fatigue life
by the contrastive learning framework with 1D-CNN prediction and similar domains. Leveraging the
as the encoder achieved the best performance on the benefits of contrastive learning, it can help achieve
downstream linear regression model. To further validate few-shot or even zero-shot learning for downstream
the contributions of data augmentation and contrastive tasks. This approach can also contribute to addressing
learning, ablation experiments were conducted. Table 4 challenges in fields such as electronic packaging and
presents the results of removing data augmentation, multiscale structural integrity, where data scarcity and
removing contrastive learning, and removing both, the need for robust predictive models are key concerns.
illustrating their impact on model performance. (v) To further enhance the effectiveness and applicability
of the proposed framework, several key directions
5. Conclusion warrant exploration. One important area is integrating

In this study, a multiaxial fatigue hysteresis feature contrastive learning with traditional physics-based
extraction method based on contrastive learning is models to bridge data-driven insights with mechanical

Volume 2 Issue 1 (2025) 68 doi: 10.36922/IJAMD025040004