<p>Laser–arc hybrid additive manufacturing (LHAM) technology combines two heat sources, laser and arc, whose coupling effect exerts a significant nonlinear influence on the formation dimensions of the deposition layers, posing certain challenges for high-precision prediction. The accurate prediction of deposition layers dimensions is critical for controlling the structural accuracy of LHAM-formed parts, playing a vital role in improving the forming precision of metal components and achieving process controllability. This study innovatively proposes a genetic algorithm-optimized backpropagation neural network (GA-BPNN) prediction model that integrates multi-strategy optimization, using the leading mode, the distance between laser beam and arc tip (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(D_\text {LA}\)</EquationSource> </InlineEquation>) and energy ratio (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(RQ_\text {LA}\)</EquationSource> </InlineEquation>) as key process parameters. The model optimizes the initial weights and thresholds of the BPNN through a GA, incorporates the bootstrap method to propagate measurement uncertainty to the prediction results, and integrates L2 regularization, early stopping, and hyperparameter scanning to establish a comprehensive overfitting control system. Ultimately, it achieves high-precision prediction of the width, height, and depth of LHAM deposition layers. The optimized GA-BPNN model achieves a maximum prediction error of only 1.673% and elevates the minimum coefficient of determination (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(R^{2}\)</EquationSource> </InlineEquation>) to 0.972. Compared to the BPNN model, the GA-BPNN model reduces the maximum prediction error by 48.3% and improves the minimum <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(R^{2}\)</EquationSource> </InlineEquation> value by 8.0%, demonstrating significantly enhanced prediction accuracy and model fitting capability. This model not only addresses the gap in existing hybrid intelligent models for small-sample manufacturing scenarios in LHAM, but also endows the predictions with physical interpretability by correlating statistical trends with physical mechanisms such as laser–arc energy synergy and molten pool dynamics. It thereby provides a quantitative tool with both practical utility and theoretical foundation for the precise intelligent control of LHAM processes.</p>

Integration of physical mechanism and data-driven prediction of deposition layers dimensions in laser–arc hybrid additive manufacturing: a genetic algorithm-optimized neural network model

期刊论文
发表日期：

2026年3月30日

查看全文

Junhua Wang,
Dongbo Lu,
Luhaotian Feng,
Junhang Wang,
Mingxin Liu,
Yuanming Mao,
Liaoyuan Chen,
Kun Li,
Tancheng Xie,
Ruijie Gu

摘要

Laser–arc hybrid additive manufacturing (LHAM) technology combines two heat sources, laser and arc, whose coupling effect exerts a significant nonlinear influence on the formation dimensions of the deposition layers, posing certain challenges for high-precision prediction. The accurate prediction of deposition layers dimensions is critical for controlling the structural accuracy of LHAM-formed parts, playing a vital role in improving the forming precision of metal components and achieving process controllability. This study innovatively proposes a genetic algorithm-optimized backpropagation neural network (GA-BPNN) prediction model that integrates multi-strategy optimization, using the leading mode, the distance between laser beam and arc tip ( \(D_\text {LA}\) ) and energy ratio ( \(RQ_\text {LA}\) ) as key process parameters. The model optimizes the initial weights and thresholds of the BPNN through a GA, incorporates the bootstrap method to propagate measurement uncertainty to the prediction results, and integrates L2 regularization, early stopping, and hyperparameter scanning to establish a comprehensive overfitting control system. Ultimately, it achieves high-precision prediction of the width, height, and depth of LHAM deposition layers. The optimized GA-BPNN model achieves a maximum prediction error of only 1.673% and elevates the minimum coefficient of determination ( \(R^{2}\) ) to 0.972. Compared to the BPNN model, the GA-BPNN model reduces the maximum prediction error by 48.3% and improves the minimum \(R^{2}\) value by 8.0%, demonstrating significantly enhanced prediction accuracy and model fitting capability. This model not only addresses the gap in existing hybrid intelligent models for small-sample manufacturing scenarios in LHAM, but also endows the predictions with physical interpretability by correlating statistical trends with physical mechanisms such as laser–arc energy synergy and molten pool dynamics. It thereby provides a quantitative tool with both practical utility and theoretical foundation for the precise intelligent control of LHAM processes.