Abstract
The application of the double-ellipsoid heat source model proposed by Goldak is prevalent in numerical simulations of arc welding, where the selection of model parameters significantly impacts computational accuracy and efficiency. However, the absence of quantitative methods for parameter description often necessitates reliance on trial-and-error calculations or previous experience. In this work, BP neural network, GA-BP neural network, and second-order polynomial regression models are employed to establish the relationship between the heat source model parameters and arc welding process parameters. The modeling process utilizes 49 datasets sourced from publicly available studies on heat source parameters, with dimensionless processing applied to the data. The results show that the three models have significant performance differences in parameter estimation. Notably, the GA-BP neural network model demonstrates excellent predictive ability, as evidenced by an RMSE value of 0.09. The approach presented in this work is capable of precisely and effectively determining Goldak heat source parameters based on any given welding process parameters, with significant potential to improve the accuracy and computational efficiency of numerical simulations in arc welding.
Introduction
Numerical simulation is a cornerstone of modern welding engineering, allowing researchers to predict temperature fields and residual stresses without the need for extensive physical experimentation. However, the fidelity of these simulations depends heavily on the accuracy of the heat source model. For decades, the "Goldak double-ellipsoid" has been the standard mathematical representation for arc welding heat sources, yet its implementation is hindered by the difficulty of determining its characteristic shape parameters.
The Challenge of Parameter Selection
The Goldak model utilizes several geometric parameters—$a, b, c_f$, and $c_r$—to describe the spatial distribution of heat in the weld pool. Because these parameters are not directly measurable from standard welding equipment settings (such as voltage or current), engineers often resort to tedious trial-and-error cycles or expensive destructive testing to calibrate their simulations. This lack of a robust quantitative framework makes the selection process inefficient and prone to error.
Methodology and Data Processing
To bridge the gap between process settings and model parameters, this study evaluates three distinct mathematical approaches to map the relationship between inputs (Voltage $U$, Current $I$, and Speed $V$) and outputs ($a, b, c_f$):
- Second-Order Polynomial Regression: A classical statistical method used to fit a curved surface to the experimental data.
- BP Neural Network: A standard three-layer artificial neural network. While powerful, standard BP networks can be inconsistent due to random weight initialization.
- GA-BP Neural Network: The core contribution of this research. A Genetic Algorithm (GA) is used to optimize the initial weights and thresholds of the neural network. By "evolving" the starting state, the GA ensures the network avoids local minima and achieves higher precision.
The researchers utilized 49 datasets extracted from existing literature. To ensure that variables with different units did not skew the learning process, all data underwent dimensionless normalization (scaling values between 0 and 1) before training.
Results and Performance Comparison
The GA-BP model significantly outperformed the other methods across all evaluation metrics. Key findings include:
- Predictive Accuracy: The GA-BP model achieved a Root Mean Square Error (RMSE) of 0.09, indicating excellent predictive capability.
- Parameter Precision: For parameter $a$, the model reached a Mean Squared Error (MSE) of just 0.0052.
- Optimization Necessity: Standard BP networks struggled particularly with the $c_f$ parameter (MSE of 0.1076), proving that Genetic Algorithm optimization is essential for a reliable predictive tool.
While the polynomial regression model proved robust for simpler estimates like weld depth ($b$), the GA-BP model is necessary for high-fidelity research where all shape parameters must be precisely determined.
Conclusion
This research demonstrates that machine learning, specifically GA-optimized neural networks, can effectively automate the calibration of arc welding simulations. By leveraging existing experimental data, this approach eliminates the need for manual trial-and-error. Future integration of this GA-BP framework into commercial Finite Element Analysis (FEA) software could significantly streamline industrial welding workflows and improve simulation accuracy across various materials and processes.
Summary updated