A Data-Driven Approach to Predicting Thermal Stress and Fatigue Life in Mechanical Components Using Machine Learning Techniques

https://doi.org/10.5281/zenodo.18871054

Abstract

Background: The accurate forecasting for thermal stress and fatigue life of mechanical components are crucial to reliability and safety in engineering applications. Classical approaches such as FEA are frequently not sufficient because they are too complicated and computationally extensive. Data-driven methods, especially machine learning (ML), offer a promise of higher predictability with the lower computational cost.

Objective: The objective of the present research is to investigate the application of machine learning, namely Physics-Informed Neural Networks (PINN), to predict thermal stress and fatigue life for recipients in different conditions.

Method: SimDA uses a composite dataset of 5,000 samples spanning various materials (steel, aluminum, titanium) over a diverse set of thermal and mechanical stress parameters to train and retrospectively test multiple machine learning models including Gradient Boosting Machine (GBM), Long Short-Term Memory (LSTM), PINN, and traditional FEA. The performance measures, namely R² score, RMSE, and MAPE were considered.

Results: Among these models, PINN model had better prediction performance with R² equal to 0.948 than LSTM (0.935) and GBM = (0.912). The RMSE and MAPE was minimum for PINN model as to the predictive accuracy among the models. Feature importance analysis showed that mean stress and plastic strain energy were the two most important parameters that influenced the fatigue life.

Conclusion: The machine learning algorithms, especially PINN, provide a new perspective for prediction of fatigue life under thermal in this study. Such models offer precise and economic techniques that could be used to be part of predictive maintenance systems increasing component life-times and safety.