Comparative Study of Physics-Informed Neural Network Architectures for Steel Fatigue Strength Prediction under Data Scarcity

Yancai Sun; Wenzhong Deng; Yan He

doi:10.6919/

Authors

Yancai Sun College of Electromechanical Engineering, Qingdao University of Science and Technology, Qingdao 266061, China;Guangxi Key Laboratory of Special Engineering Equipment and Control, Guilin University of Aerospace Technology, Guilin 541004, China;University Engineering Research Center of Non-standard Intelligent Equipment and Process Control Technology, Guangxi, Guilin 541004, China;Guangxi Region Precious Metal Materials Advanced Process Research Center, Guilin 541004, China
Wenzhong Deng Guangxi Key Laboratory of Special Engineering Equipment and Control, Guilin University of Aerospace Technology, Guilin 541004, China;University Engineering Research Center of Non-standard Intelligent Equipment and Process Control Technology, Guangxi, Guilin 541004, China;Guangxi Region Precious Metal Materials Advanced Process Research Center, Guilin 541004, China
Yan He College of Electromechanical Engineering, Qingdao University of Science and Technology, Qingdao 266061, China

DOI:

https://doi.org/10.6919/

Keywords:

Physics-informed neural networks, fatigue strength prediction, small-sample learning, residual physics network, multi-fidelity learning.

Abstract

The effect of architecture choice on physics-informed learning under data scarcity remains unclear for steel fatigue prediction. This study presents a comparative evaluation of four physics-informed neural network (PINN) architectures on the NIMS steel fatigue dataset with 437 samples: a loss-level baseline MLP-PINN, Residual Physics Network (ResPhys-Net), Multi-Fidelity PINN (MF-PINN), and Physics-Aware Attention PINN (PA-PINN). The models are assessed by 5-fold cross-validation and by repeated small-sample experiments with training fractions from 10% to 100%. In the full-data setting, PA-PINN achieves the highest mean R² of 0.985, followed closely by ResPhys-Net (0.985) and MF-PINN (0.984). Under the 10% training-data setting, ResPhys-Net and MF-PINN retain mean R² values of 0.884 and 0.879, respectively, compared with 0.828 for XGBoost and 0.815 for Random Forest. A complementary accuracy–stability analysis further indicates that residual and multi-fidelity architectures are less sensitive to repeated subsampling than attention-based fusion. These results suggest that architecture preference depends on the data regime: attention-based fusion is slightly advantageous with richer data, whereas residual and multi-fidelity designs show more stable behavior when the available fatigue data are very limited.

Downloads

Download data is not yet available.

References

[1] Murakami, Y. (2002). Metal fatigue: Effects of small defects and nonmetallic inclusions. Elsevier.

[2] Agrawal, A., Deshpande, P. D., Cecen, A., Basavarsu, G. P., Choudhary, A. N., & Kalidindi, S. R. (2014). Exploration of data science techniques to predict fatigue strength of steel from composition and processing parameters. Integrating Materials and Manufacturing Innovation, 3, 8.

[3] Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 686–707.

[4] Zhao, Z., Ding, X., & Prakash, B. A. (2024). PINNsFormer: A transformer-based framework for physics-informed neural networks. In Proceedings of ICLR.

[5] Wang, S., Li, B., Chen, Y., & Perdikaris, P. (2024). PirateNets: Physics-informed deep learning with residual adaptive networks. Journal of Machine Learning Research, 25, 1–48.

[6] Meng, X., & Karniadakis, G. E. (2020). A composite neural network that learns from multi-fidelity data: Application to function approximation and inverse PDE problems. Journal of Computational Physics, 401, 109020.

[7] Maynier, P., Jungmann, B., & Dollet, J. (1978). Creusot-Loire system for the prediction of the mechanical properties of low alloy steel products. In D. V. Doane & J. S. Kirkaldy (Eds.), Hardenability concepts with applications to steel (pp. 518–545). AIME.

[8] Li, H., Sun, G., Tian, Z., Huang, K., & Zhao, Z. (2024). A physics-informed neural network framework based on fatigue indicator parameters for very high cycle fatigue life prediction of an additively manufactured titanium alloy. Fatigue & Fracture of Engineering Materials & Structures, 47(9), 3171–3188.

[9] Zhou, S., Henrich, M., Wei, Z., Feng, F., Yang, B., & Münstermann, S. (2025). A general physics-informed neural network framework for fatigue life prediction of metallic materials. Engineering Fracture Mechanics, 322, 111136.

[10] Akiba, T., Sano, S., Yanase, T., Ohta, T., & Koyama, M. (2019). Optuna: A next-generation hyperparameter optimization framework. In Proceedings of ACM SIGKDD (pp. 2623–2631).

[11] Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., & Yang, L. (2021). Physics-informed machine learning. Nature Reviews Physics, 3, 422–440.

[12] Guo, J., Zhu, H., Yang, Y., & Guo, C. (2025). Advances in physics-informed neural networks for solving complex partial differential equations and their engineering applications: A systematic review. Engineering Applications of Artificial Intelligence, 161, 112044.

[13] Jo, J., Jeong, Y., Kim, J., & Yoo, J. (2024). Thermal conductivity estimation using physics-informed neural networks with limited data. Engineering Applications of Artificial Intelligence, 137, 109079.

[14] Wang, S., Yu, X., & Perdikaris, P. (2022). When and why PINNs fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449, 110768.

[15] Krishnapriyan, A. S., Gholami, A., Zhe, S., Kirby, R. M., & Mahoney, M. W. (2021). Characterizing possible failure modes in physics-informed neural networks. In Advances in Neural Information Processing Systems (Vol. 34, pp. 26548–26560).