Study on a Process Parameter-Driven Deep Learning Prediction Model for Multi-Physical Fields in Flange Shaft Welding.

Area of Science:

• Materials Science and Engineering
• Computational Mechanics
• Artificial Intelligence

Background:

• Large flange shafts are critical components in high-end equipment, with welding quality directly impacting performance and safety.
• Traditional experimental and finite element methods for predicting welding multi-physical fields are time-consuming and inefficient.

Purpose of the Study:

• To develop fast and accurate prediction models for welding temperature, deformation, and residual stress in large flange shafts.
• To combine thermal-mechanical coupled finite element simulation with machine learning for enhanced prediction capabilities.

Main Methods:

• A dataset of 100 process parameter groups was generated using finite element simulations.
• Machine learning models including MLP, RF, RBF-SVR, TabNet, XGBoost, and FT-Transformer were constructed and compared.
• Strategies like early stopping and dropout were employed to mitigate overfitting, with 10-fold cross-validation used for verification.

Main Results:

• The MLP model demonstrated superior performance, accurately predicting peak temperature (error < 5%), deformation, and residual stress (error < 10%).
• The MLP model showed good agreement with simulation values, with average peak residual stress error around 6 MPa.
• The Random Forest (RF) model ranked second, offering good interpretability and engineering applicability.

Conclusions:

• The developed MLP model effectively reproduces welding multi-physical fields, enabling rapid prediction and optimization for large flange shaft welding.
• The study highlights the potential of machine learning to overcome limitations of traditional methods in predicting welding quality.
• Future work can enhance model generalization by expanding the dataset and incorporating experimental validation.