Page 62 - IJAMD-2-1

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

based, whereas deep learning methods can automatically The essence of feature engineering lies in the
learn high-dimensional abstract representations from extraction and dimensionality reduction of complex,
raw experimental data, capturing complex nonlinear high-dimensional features, which brings samples closer
relationships within the data. This comparison highlights together in high-dimensional space. Traditional models,
the strengths and limitations of both approaches and based on physical laws and phenomenon analysis, focus
underscores the importance of selecting appropriate input on extracting and analyzing key values. However, many
features for accurate fatigue life prediction. Wang et al. of these models are empirical, summarizing patterns
40
proposed two physics-guided machine learning frameworks that may not truly reflect the relationships between data
based on the Paris law and machine learning models. The samples, and they may not have clear physical models to
first framework transforms the original features into new describe the underlying dynamics. From the perspective of
features using the Paris law, which are then combined with data relationships, dimensionality reduction or clustering
the original features and input into the machine learning techniques, such as K-means, variational autoencoder,
45
46
model to predict fatigue life. The second framework and contrastive learning, 47-50 can also be employed with
integrates the Paris law and machine learning models effective results. Among these, contrastive learning
using the Kalman filter, leveraging the advantages of both stands out by maximizing the consistency between
approaches and integrating information from different similar samples and the disparity between different ones,
models. This enables the fusion of physical information enabling the learning of more robust and generalized
and machine learning, allowing the model to account for feature representations. This approach does not require
factors ignored by physics-based models while ensuring prior knowledge of label information, which reduces
consistency with physical results. Dong et al. unified the model usage costs. Furthermore, its adversarial sample
41
influence of different defects through equivalent damage pair learning strategy enhances the model’s generalization
area representation using the M-integral fatigue model. By ability. By optimizing the relative distances between
taking cyclic loading and equivalent damage area as inputs samples, contrastive learning provides a powerful learning
and fatigue life as output, the approach effectively improved mechanism for complex datasets. This is especially
51
the generalization ability and prediction accuracy of beneficial in unsupervised and self-supervised learning
incomplete fatigue data. Fan et al. expanded the original scenarios, where it outperforms traditional algorithms.
42
dataset using the Z-parameter model by restricting key These advantages have led to contrastive learning
parameters such as the size of critical defects, the relative demonstrating outstanding application performance in
depth of critical micro-defects, and stress levels within a variety of fields, such as visual recognition, 52,53 natural
certain ranges. These extended datasets were then used to language processing, 54,55 and sound analysis. 56,57
build machine learning models for ultra-high-cycle fatigue This paper addresses the issue of multiaxial fatigue life
life prediction. With an appropriate increase in the size prediction in materials by applying contrastive learning
of the training set, the model demonstrated significantly to effectively extract sample features across different
improved accuracy and exhibited higher accuracy and multiaxial loading paths. The paper compares the feature
stronger generalization compared to physical models. representation performance of different model frameworks
43
Gan et al. used the Ye-Wang damage theory to derive and tests various downstream task models. Compared to
theoretical estimates of material behavior, constructing other clustering or dimensionality reduction algorithms,
additional features closely related to the desired outputs. contrastive learning consistently achieves good prediction
This approach integrated original data information with results, providing a new feature engineering strategy for
domain knowledge. In addition, output standards were multiaxial fatigue life prediction.
set to provide information for the data-driven process of
model training and prediction, highlighting its potential 2. Data
in addressing small-sample problems. Wang et al.
44
characterized the fatigue life of additive manufacturing 2.1. Experimental data
(AM) parts under different stress amplitudes using In this study, 20 multiaxial fatigue experimental data of
predictions from continuous damage mechanics associated 316L stainless steel were used for training and validation
with AM. These predictions, along with initial features of the contrastive learning model and downstream fatigue
such as experimental conditions, mechanical properties, life prediction model. The experimental details can be
porosity analysis, and surface morphology, were used as found in the referenced published literature. 16,17 As shown
inputs. By learning the dependence on physical principles, in Figure 2, the experimental data included four typical
the model can better map the nonlinear relationships uniaxial and multiaxial loading paths, each with five
between inputs and outputs. different amplitudes, with a stress ratio of R = −1. Table 1

Volume 2 Issue 1 (2025) 56 doi: 10.36922/IJAMD025040004