Machine learning techniques for the optimisation and simulation of metal additive layer manufacturing process chains

Le Roux, Leopold 2024. Machine learning techniques for the optimisation and simulation of metal additive layer manufacturing process chains. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Download (56MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (957kB)

Abstract

Metal Additive Manufacturing is starting to revolutionise parts fabrication with its capability to create shapes by adding material layer by layer and its ability to create parts that would have been complex to create with traditional methods. However, despite its great potential, optimising quality control and predicting the final part properties remains an ongoing challenge. In this context, a European Union project (MANUELA) funded by the Horizon 2020 research and innovation programme was conducted to develop new and more efficient metal additive manufacturing (AM) based pilot lines covering the full product development cycle. As part of that project, this thesis aimed at developing novel deep learning techniques for the enhancement of the AM process. Many Deep learning techniques are emerging showing impressive modelling capabilities in various areas. However in the field of manufacturing, their full potential is not yet fully achieved due to their rapid development combined with constantly improving computing and monitoring devices. Thus, the first achievement of this thesis was a new monitoring system developed to assess, during machining, the quality of the Electron Beam Melting (EBM) printing process using images of each printed layer. This study demonstrated that using deep learning, the information contained in these layer images can be extracted and serve as reliable indicators of the printing quality. The developed predictive model achieved high accuracy and appeared to performbetter than the human experts who created the training data. A detailed analysis of misclassified images revealed a significant margin of human error not replicated by the produced deep learning models, thereby substantiating the superiority of the machine-learning-based approach. This work was further improved by segmenting the large manually labelled monitoring images into smaller regions for the training of CNN models, allowing the detection of defects with higher resolution. The work showed that this new approach can be used to avoid the time-consuming (or even unachievable) manual labelling of small powder bed areas that would be required for the construction of "higher resolution" training data sets, thus reducing the cost and time required to generate accurate predictive models. The effectiveness of this refined technique was confirmed through varying degrees of accuracy, dependent on sub-image dimensions and on the types of defects considered. This type of deep learning model is opening the door for realtime decision-making during the EBM process. The two other main achievements resulted from exploring the application of deep learning in estimating part deformations resulting fromtheMetalAMprocess. The main idea is to use deep learning models to emulate relatively slow Physics-based Metal AM simulations, thus enabling rapid deformation prediction, which would facilitate swifter design iterations. Thus, firstly a new CNN modeling approach based on MeshCNN was developed. The models ii created can predict the final geometry deformations of parts printed without supports in less than half a second, under fixed printing conditions. For training, a large dataset of deformed parts was generated using an inherent strain based AM simulation to deform parts. To make this dataset compatible with the Deep Learning method used, a pre-processing method was developed to ensure parts validity with a manifold mesh, to create a fixed number of edges and to normalise the dataset. During the training of the DL model, a new custom losswas used,guiding the DL modelswithAMknowledge to try to make them converge faster to a solution. After training, another metric was introduced, to shift the error evaluation from the difference between predicted and simulated deformation vectors to the difference between the two final volumes. In the end, this DL model proved to be capable of having low errors with the custom metric in 75% of its evaluation and showed that the addition of domain-specific knowledge in the loss function did not improve the performance of the DL model. Secondly, the limitations of this modelling approach were tackled in the last main achievement of this thesis by developing a new CNN modelling approach based on Graph NN that can mimic Physics-based simulation steps for a wide array of printing parameters and considering support structures. These modelling approaches provide a mechanism for quick iteration across various geometries and printing conditions before resorting to more time-consuming Physicsbased simulations. Unlike in the previous section, which used only the final simulation results, in this approach all simulation steps were used to create datasets of parts being created and with different printing parameters and geometries. The generated Graph NN were capable of distinguishing the printing parts from their supports and their different thermal and physical evolution behaviours. To train this Graph NN, another custom loss function was created, to add simulation knowledge to its training. Overall, the contributions of this thesis set the stage for the intelligent automation of additive manufacturing, heralding a new era of efficiency and reliability in production techniques.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Engineering
Uncontrolled Keywords:	1) Machine Learning 2) Deep Learning 3) Additive Manufacturing 4) Graph Neural Network 5) Defaults detection 6) Surrogate model
Date of First Compliant Deposit:	6 January 2025
Last Modified:	06 Jan 2025 15:06
URI:	https://orca.cardiff.ac.uk/id/eprint/175030

Actions (repository staff only)

Edit Item