DeepOFormer: Deep Operator Learning with Domain-informed Features for Fatigue Life Prediction
Chenyang Li, Tanmay Sunil Kapure, Prokash Chandra Roy, Zhengtao Gan, Bo Shen · 2025
Read DeepOFormer: Deep Operator Learning with Domain-informed Features for Fatigue Life Prediction by Chenyang Li, Tanmay Sunil Kapure, Prokash Chandra Roy, Zhengtao Gan, Bo Shen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-22475.txt.
Introduction
Chenyang Li and four co-authors predict how many load cycles a metal survives before cracking, using a neural operator fed with old-fashioned engineering wisdom, in this 2025 paper. Fatigue life is normally read off stress-life curves built from expensive testing. For a new alloy, data are scarce, and ordinary machine learning overfits the little it has.
Their model, DeepOFormer, reframes the task: instead of fitting numbers, it learns the whole curve as an operator, using a transformer encoder and a loss tuned to relative error. Features drawn from classic empirical fatigue formulas, those of Stussi, Weibull, and Pascual and Meeker, are fed in as domain knowledge. On fifty-four curves from aluminum alloys, holding seven alloys back entirely for testing, the model predicted far better than rival networks and gradient boosting. Scarcity is the normal state of materials data, so teaching models the old formulas matters.
Catalog record
| CREATOR | Chenyang Li, Tanmay Sunil Kapure, Prokash Chandra Roy, Zhengtao Gan, Bo Shen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 24 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-22475
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
DeepOFormer: Deep Operator Learning with Domain-informed Features for Fatigue Life Prediction Chenyang Li, Tanmay Sunil Kapure, Prokash Chandra Roy, Zhengtao Gan, Bo Shen Abstract Fatigue life characterizes the duration a material can function before failure under specific environmental conditions, and is traditionally assessed using stress-life (S-N) curves. While machine learning and deep learning offer promising results for fatigue life prediction, they face the overfitting challenge because of the small size of fatigue experimental data in specific materials. To address this challenge, we propose, DeepOFormer, by formulating S-N curve prediction as an operator learning problem. DeepOFormer improves the deep operator learning framework with a Transformer-based encoder and a mean L2 relative error loss function. We also consider Stüssi, Weibull, and Pascual and Meeker (PM) features as domain-informed features. These features are motivated by empirical fatigue models. To evaluate the performance of our DeepOFormer, we compare it with different deep learning models and XGBoost on a dataset with 54 S-N curves of aluminum alloys. With seven different aluminum alloys selected for testing, our DeepOFormer achieves an R 2 R^{2} of 0.9515, a mean absolute error of 0.2080, and a mean relative error of 0.5077, significantly outperforming state-of-the-art deep/machine learning methods including DeepONet, TabTransformer, and XGBoost, etc. The results highlight that our Deep0Former integrating with domain-informed features substantially improves prediction accuracy and generalization capabilities for fatigue life prediction in aluminum alloys. I Introduction Fatigue life characterizes the duration over which a material can function before failure occurs under specific environmental conditions [ 1 , 2 ] . The reliability and security of aluminum alloys are commonly evaluated by measuring their fatigue life across varied production environments [ 3 ] . The stress-life diagram, commonly known as the S-N curve, graphically represents the relationship between applied stress amplitude and the number of cycles until material failure [ 4 ] . This serves as a fundamental method for characterizing fatigue behavior in materials. S-N curves play an essential role in evaluating safety and durability characteristics. Unfortunately, conducting numerous fatigue experiments to generate these curves is both financially burdensome and extraordinarily time-consuming, particularly for high-cycle fatigue testing. Therefore, developing predictive models for S-N curves represents a valuable approach that can efficiently support both the characterization of fatigue properties and the strategic development of alloys engineered for specific fatigue performance requirements. It is very challenging to develop a universal model for predicting S-N curves due to fatigue’s inherent complexity and limited experimental data. Instead, researchers have created various empirical fatigue models that transform experimental data into mathematical equations approximating S-N curve behavior, such as Basquin’s exponential law [ 5 ] and Manson–Coffin relationship [ 6 ] . Additionally, the Stüssi function models a nonlinear S-N curve with fatigue-limited considerations to improve the flexibility of fatigue life predictions [ 7 ] . The Weibull distribution, originally developed as a general statistical model, has been widely adopted to describe fatigue failure probability and model material reliability under varying stress conditions [ 8 ] . The Pascual and Meeker (PM) model estimates fatigue strength using a statistical framework that accounts for random fatigue limits and experimental data variability to enhance fatigue life assessment [ 9 ] . However, these empirical S-N models often fail as predictive tools due to parameter uncertainties and material-specific variations in curve shapes. While they effectively fit data for individual materials, they typically cannot accurately predict fatigue behavior across different materials [ 10 , 11 ] . Machine learning (ML) and deep learning (DL) have emerged as powerful alternatives for material science [ 12 , 13 , 14 ] due to their prediction power. Traditional ML models, including Random Forests, Support Vector Machines, and XGBoost, have been applied to capture complex nonlinear relationships within fatigue data for various materials and conditions [ 15 , 16 , 17 ] . Further, Dai et al. [ 18 ] used artificial neural networks to predict the fatigue life of laser powder bed fusion stainless steel. Wei et al. [ 19 ] employed the long short-term memory (LSTM) network combined with the transfer learning concept to estimate the high cycle S-N curve. Zhu et al. [ 20 ] proposed Multi-Graph Attention Networks (Multi-GAT) to predict the high cycle fatigue life of several types of titanium alloys. Despite the success of ML/DL models, one challenge is that the size of the fatigue experimental dataset is usually small, which raises concerns of ML/DL about poor extrapolation beyond training conditions [ 21 ] . Researchers have proposed to incorporate domain knowledge into ML and DL models. Lian et al. [ 22 ] successfully combined empirical knowledge with ML for aluminum alloy fatigue life prediction, achieving significantly improved accuracy. Gan et al. [ 23 ] incorporated not only real data from experiments but also dummy data generated by empirical fatigue life models to augment training datasets. Chen et al. [ 24 ] implemented the physics-informed neural networks (PINNs) [ 25 ] for the fatigue crack growth simulation and life prediction of structures. However, these models are restricted to traditional ML/simple neural networks with limited domain knowledge. To address these challenges, we introduce DeepOFormer, a deep operator learning with domain-informed features, for fatigue life prediction of aluminum alloys. Specifically, the contributions of our paper can be summarized as follows. • We formulate the S-N curve prediction problem as an operator learning problem and propose a novel model called DeepOFormer. • We explored multiple domain-informed features such as Stüssi, Weibull, and PM features. • We use data from [ 22 ] to demonstrate the effectiveness and efficiency of DeepOFormer by comparing it with state-of-the-art deep learning methods. The rest of this paper is organized as follows. The Dataset used in this paper is introduced in Section II , followed by our proposed DeepOFormer in Section III . Our experiments and results are summarized in Section IV . Finally, conclusions are discussed in Section V . II Dataset Description The S-N curve can illustrate the relationship between stress amplitude ( \sigma_{a} ) applied during cyclic loading and the number of cycles ( N ) to failure. It is typically plotted with stress amplitude on the vertical axis and the number of cycles on the horizontal axis. It provides a graphical representation essential for predicting fatigue performance and establishing safe operating limits. The fatigue S-N curve dataset used in this work is from [ 22 ] , which has 54 S-N curves from seven types of aluminum alloys comprising 257 high-cycle (10 4 –10 10 cycles) tensile fatigue test data. All tests are performed at room temperature conditions. In the dataset, mechanical properties such as ultimate tensile stress (UTS), tensile yield strength (TYS), and Fatigue Strength are taken into consideration because of their high positive correlation with fatigue life. Temper [ 26 ] is one important categorical feature that contains information on treatments in aluminum alloy production and correlates to the microscopic structure. Stress ratio (R) that equals \sigma_{max}/\sigma_{min} is also included in the feature set, where \sigma_{min} is minimum cyclic stress, \sigma_{max} is maximum cyclic stress. We also consider stress amplitude \sigma_{a} and \sigma_{a}^{3} in the feature set for modeling. The logarithmic of fatigue life, i.e., \log(N) , is used as the target variable in the dataset