Machine learning surrogate models for Hertzian contact stress prediction in gear design: A comparative study of multiple approaches

• Machine Learning for faster gear design • Study compares multiple regression and neural network models • Focus on surface fatigue and pitting over root stress Accurately predicting contact stress in gears is essential for ensuring durability and optimizing performance during the design stage. This study investigates machine learning surrogate models for predicting Hertzian contact stress in involute gear pairs, aiming to accelerate the gear design process. Unlike approaches based on finite element simulations or experimental data, the proposed method relies solely on stress values from ISO 6336 formulations. A comprehensive dataset covering various macro geometries and loading conditions is used to train and evaluate several regression models, including Elastic Net, Support Vector Regressor, ensemble methods, and Neural Networks. To improve computational efficiency and address the high dimensionality of the input space, Principal Component Analysis is explored as a dimensionality reduction technique. The study provides a detailed comparison of surrogate modeling approaches based on predictive accuracy and generalization. Results show that surrogate models can accurately reproduce ISO based predictions, offering a faster alternative to traditional methods. Focusing on Hertzian stress, rather than root stress, offers deeper insight into surface fatigue and pitting resistance for gear life improvement. This work establishes the foundation for future surrogate models that optimize gear pairs by balancing multiple design objectives and constraints, aiding engineers in selecting optimal configurations.