<p>In this study, a hybrid lightweight bidirectional deep generative framework is proposed to model the nonlinear interactions between process parameters and melt-pool geometry in the laser powder bed fusion (LPBF) process. The framework is built upon a modified Conditional Variational Autoencoder (CVAE), in which conventional input concatenation is replaced by an interaction-based architecture that explicitly embeds physics-inspired parameter relationships into the latent space. This design reduces the effective dimensionality of the learning problem and directly addresses a key challenge in applying machine learning to manufacturing, namely the small sample sizes. By leveraging the generative nature of the CVAE, which captures joint rather than purely conditional distributions among variables, the framework enables both forward and inverse digital-twin capabilities. In the forward mode, the model predicts melt-pool width and depth and reconstructs physically consistent melt-pool cross-sections through an image-based generative pipeline. In the inverse mode, the framework performs reverse engineering by inferring feasible LPBF process-parameter ranges capable of reproducing a prescribed melt-pool geometry, thereby providing a data-driven tool for experimental design and parameter selection. The proposed numerical CVAE demonstrates strong predictive performance, achieving an overall coefficient of determination of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$R^{2} = 0.955$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mi>R</mi> <mn>2</mn> </msup> <mo>=</mo> <mn>0.955</mn> </mrow> </math></EquationSource> </InlineEquation> for the full-dataset training and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$R^{2} = 0.87$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mi>R</mi> <mn>2</mn> </msup> <mo>=</mo> <mn>0.87</mn> </mrow> </math></EquationSource> </InlineEquation> for 80/20 train-test split. Finally, this study proposes a hybrid two-stage model for image-based digital twin. Compared with a single-stage Convolutional Neural Network (CNN) combined CVAE, which achieves an image reconstruction accuracy of <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$R^{2} = 0.7691$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mi>R</mi> <mn>2</mn> </msup> <mo>=</mo> <mn>0.7691</mn> </mrow> </math></EquationSource> </InlineEquation>, the hybrid two-stage model further improves the reconstruction performance to <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$R^{2} = 0.7790$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mi>R</mi> <mn>2</mn> </msup> <mo>=</mo> <mn>0.7790</mn> </mrow> </math></EquationSource> </InlineEquation>. In parallel with the forward digital twin, the inverse model infers feasible LPBF process parameters with prediction errors that remain within the natural variability of the corresponding melt-pool regimes. Overall, the proposed framework serves as a robust and practical digital twin tool for LPBF manufacturers operating under limited data availability.</p>

A Bidirectional Physics-Inspired Digital Twin for Melt-Pool-Process Parameter Interplay in Laser Powder-Bed Fusion

期刊论文
发表日期：

2026年7月23日

查看全文

Sabbir Alom Shuvo,
Xiyuan Liu,
M Shafiqur Rahman

摘要

In this study, a hybrid lightweight bidirectional deep generative framework is proposed to model the nonlinear interactions between process parameters and melt-pool geometry in the laser powder bed fusion (LPBF) process. The framework is built upon a modified Conditional Variational Autoencoder (CVAE), in which conventional input concatenation is replaced by an interaction-based architecture that explicitly embeds physics-inspired parameter relationships into the latent space. This design reduces the effective dimensionality of the learning problem and directly addresses a key challenge in applying machine learning to manufacturing, namely the small sample sizes. By leveraging the generative nature of the CVAE, which captures joint rather than purely conditional distributions among variables, the framework enables both forward and inverse digital-twin capabilities. In the forward mode, the model predicts melt-pool width and depth and reconstructs physically consistent melt-pool cross-sections through an image-based generative pipeline. In the inverse mode, the framework performs reverse engineering by inferring feasible LPBF process-parameter ranges capable of reproducing a prescribed melt-pool geometry, thereby providing a data-driven tool for experimental design and parameter selection. The proposed numerical CVAE demonstrates strong predictive performance, achieving an overall coefficient of determination of $R^{2} = 0.955$ $R^{2} = 0.955$ for the full-dataset training and $R^{2} = 0.87$ $R^{2} = 0.87$ for 80/20 train-test split. Finally, this study proposes a hybrid two-stage model for image-based digital twin. Compared with a single-stage Convolutional Neural Network (CNN) combined CVAE, which achieves an image reconstruction accuracy of $R^{2} = 0.7691$ $R^{2} = 0.7691$ , the hybrid two-stage model further improves the reconstruction performance to $R^{2} = 0.7790$ $R^{2} = 0.7790$ . In parallel with the forward digital twin, the inverse model infers feasible LPBF process parameters with prediction errors that remain within the natural variability of the corresponding melt-pool regimes. Overall, the proposed framework serves as a robust and practical digital twin tool for LPBF manufacturers operating under limited data availability.