<p>Used in industries such as automotives and aeronautics, metal laser welding remains a challenging process to master. It is characterised by its high penetration-to-width ratio, which is useful when the weld is realised through a material. Due to a physical transformation involving non-linear laser-matter interaction, thermodynamic and fluid mechanics, it possesses numerous interactions between its parameters, making it a poor candidate for simulation or design of experiment modelling. The main method to get a production-ready set of parameters remains the time-consuming and labour-intensive trial and error. In this study, an artificial intelligence model is investigated on a large dataset of diverse materials to predict weld penetration from sets of process parameters. The dataset was pre-processed by adding a physical description of the material as input to the model. An optimised multi layer perceptron (MLP), a type of shallow neural network, achieved good performance, both on the validation sets and in process-parameter-map generation. This is done because we do not only consider the R<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="170_2025_16437_Article_IEq1.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="8" /> </InlineMediaObject> <EquationSource Format="TEX">$^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mn>2</mn> </mmultiscripts> </math></EquationSource> </InlineEquation> but also the continuity of the predicted process parameter maps during this study. When testing is performed on materials unknown to the model, the performance of the MLP model are found to depend on the extracted material’s similarity to those left in the training set. However, we found that a high R<InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="170_2025_16437_Article_IEq1.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="8" /> </InlineMediaObject> <EquationSource Format="TEX">$^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mn>2</mn> </mmultiscripts> </math></EquationSource> </InlineEquation> does not to ensure the prediction of a continuous and representative process parameter prediction map. Such a model could be used to identify the set of parameters able to achieve a given laser weld penetration in new metals.</p>

Machine learning for prediction of laser welds penetration on a multi-material dataset

期刊论文
发表日期：

2025年9月3日

查看全文

Victor Hayot,
Rabih Amhaz,
Sylvain Lecler,
Andre Alves Ferreira,
Grégoire Chabrol

摘要

Used in industries such as automotives and aeronautics, metal laser welding remains a challenging process to master. It is characterised by its high penetration-to-width ratio, which is useful when the weld is realised through a material. Due to a physical transformation involving non-linear laser-matter interaction, thermodynamic and fluid mechanics, it possesses numerous interactions between its parameters, making it a poor candidate for simulation or design of experiment modelling. The main method to get a production-ready set of parameters remains the time-consuming and labour-intensive trial and error. In this study, an artificial intelligence model is investigated on a large dataset of diverse materials to predict weld penetration from sets of process parameters. The dataset was pre-processed by adding a physical description of the material as input to the model. An optimised multi layer perceptron (MLP), a type of shallow neural network, achieved good performance, both on the validation sets and in process-parameter-map generation. This is done because we do not only consider the R $^{2}$ $^{2}$ but also the continuity of the predicted process parameter maps during this study. When testing is performed on materials unknown to the model, the performance of the MLP model are found to depend on the extracted material’s similarity to those left in the training set. However, we found that a high R $^{2}$ $^{2}$ does not to ensure the prediction of a continuous and representative process parameter prediction map. Such a model could be used to identify the set of parameters able to achieve a given laser weld penetration in new metals.