Objective In laser directed energy deposition (LDED) for thin-walled parts, laser power, scanning speed, and powder feeding rate exhibit strong coupling. Laser power dictates molten pool temperature: excessive power may cause over-melting or burnout, while insufficient power leads to incomplete fusion. Scanning speed modulates laser-material interaction time—faster speeds risk inadequate melting and weak bonding, slower ones may induce overheating and deformation. Powder feeding rate correlates directly with cladding quantity; mismatched rates disrupt the balance between powder supply and molten pool capacity, causing uneven deposition. Critical is that minor parameter fluctuations amplify during layer-by-layer deposition, severely hindering precise dimension control and triggering issues like inconsistent wall thickness, warpage, or collapse. Therefore, it is essential to study the relationship of LDED parameters and thin-walled part dimensions. It improves forming accuracy, ensures structural integrity, optimizes parameter selection, and advances LDED’s application in high-precision thin-walled component manufacturing.

Methods To address this technical issue, this study develops an orthogonal experimental scheme via Taguchi design and collects thin-walled part dimensional data precisely using a high-definition electron microscope. A prediction model based on the Multi-Verse Optimizer-Generalized Regression Neural Network (MVO-GRNN) is developed, with the height and width of thin-walled parts as output variables and laser power, scanning speed, and powder feeding rate as input variables. The orthogonal test ensures efficient exploration of parameter influences with limited experiments, while high-definition electron microscopy provides high-precision dimensional data to enhance model reliability. The MVO-GRNN model combines the nonlinear mapping capabilities of a generalized regression neural network with the optimization capabilities of a multi-verse optimizer. Its goal is to effectively predict dimensional control in laser directed energy deposition by capturing the intricate nonlinear relationship between process parameters and thin-walled part dimensions.

Results and Discussions The optimal parameters obtained by the MVO algorithm were respectively substituted into the GRNN models for height and width, thereby establishing the MVO-GRNN prediction model for the dimensions of thin-walled parts fabricated by LDED. The comparison diagrams of the prediction results are shown in Fig.9 and Fig.10. For the height prediction model, the coefficient of determination R² reaches 0.95954, and the root mean square error is 0.313 95. For the width prediction model, the coefficient of determination R² is 0.96959, with a root mean square error of 0.114 95. Moreover, as shown in Fig.11 and Fig.12, in the prediction of the height and width of thin-walled parts, compared with the standard GRNN model, the root mean square error of the MVO-GRNN model is reduced by 13.5% and 26%.5% respectively, and the mean absolute error is decreased by 11.5% and 19.3% respectively. These results indicate that the prediction results of the MVO-GRNN model are closer to the actual values than those of the standard GRNN model. In terms of various indicators, the goodness of fit of the MVO-GRNN model is higher than that of the standard GRNN model, which suggests that the MVO-GRNN prediction model can better explain the nonlinear relationship between the LDED process parameters and the dimensions of thin-walled parts. The results show that the MVO-GRNN model exhibits excellent prediction performance in the dimension prediction of thin-walled parts, and it also indicates that the parameter optimization of the GRNN model plays an important role in improving the prediction accuracy of the dimensions of thin-walled parts manufactured by LDED.

Conclusions To explore the intricate mapping relationship between the process parameters of LDED and the dimensions of thin-walled components. This study proposes a novel prediction method based on the MVO-GRNN. Key process parameters in LDED, such as laser power, scanning speed, and powder feeding rate, exhibit strong coupling characteristics. Meanwhile, thin-walled part dimensions are highly sensitive to parameter fluctuations. Accurately capturing their inherent correlation has long been a technical challenge in the field. The proposed MVO-GRNN method innovatively combines the global optimization capability of the MVO algorithm with the excellent nonlinear fitting performance of the GRNN. Specifically, it employs the MVO algorithm to search for the optimal parameters of the GRNN, thereby overcoming the limitation that the traditional GRNN is prone to falling into local optima due to improper parameter settings. On this basis, an MVO-GRNN prediction model tailored to the dimensions of thin-walled parts is constructed. This model effectively quantifies the complex nonlinear relationship between LDED process parameters and thin-walled part dimensions. It also provides a new technical approach for achieving high-precision prediction of the dimensions of LDED thin-walled parts. Moreover, this model is expected to lay a theoretical foundation for the parameter optimization and quality control in the manufacturing of LDED thin-walled parts.