Figure - available from: Journal of Intelligent Manufacturing
This content is subject to copyright. Terms and conditions apply.
Scanning speed dependence of the D/W and H/W ratios for different laser power levels (200, 350, and 500 W)
Source publication
In this paper, the effects of two key process parameters of the selective laser melting process, namely laser power and scanning speed, on the single-track morphologies and the bead characteristics, especially the depth-to-width D/W and height-to-width H/W ratios, were investigated using both experimental and Machine Learning (ML) approaches. A tot...
Citations
... Many models have been established to describe the input-output mapping relationships in metal AM via conventional regression equations [11,12]. In recent years, machine learning (ML) algorithms based on data-driven, such as artificial neural network (ANN), random forest (RF), Naïve Bayes (NB), and support vector regression (SVR) [13][14][15][16][17][18], have been widely used to solve the classification and regression tasks of multiple objectives by automatically establishing data dependence models. These algorithms possess the merits of excellent self-learning capacity and high adaptability. ...
... They used eight different ML models to estimate the layer geometry from process parameters and material properties, verifying the regression and classification performances of these ML models. Hong et al. [16] used laser power and scanning velocity as inputs of RF and ANN models to predict the layer geometry in LPBF, indicating that the overall R-squared values of these models exceeded 90% on the validation dataset. Zhu et al. [17] established an SVR model to predict the layer height and width by the laser power, travel speed, and powder feed rate in powder-laser DED. ...
The quantitative prediction of more process parameter variables for fewer layer geometry variables is challenging in wire-laser DED. This study’s novelty is combining machine learning models with a non-dominated sorting genetic algorithm-II (NSGA-II) to predict process parameters for desired layer geometries accurately. Thirty single-layer deposition experiments are conducted to obtain response data of layer geometries to process parameters. Two support vector regression (SVR) models are trained by these data to predict the layer height and width, respectively, and the mean absolute percentage errors (MAPEs) of these models are 4.16% and 1.76%. A reverse system, consisting of both SVR models and the NSGA-II algorithm, is designed to search the optimal process parameters for the desired layer geometries. The maximum MAPE between the actual layer geometry deposited by the predicted process parameters and the desired layer geometry is less than 5.5%, providing solid confirmation of this methodology’s reliability.
... Moreover, ML methods do not rely on restrictive assumptions about the analyzed processes, allowing them to adapt to a broad spectrum of process variability. ML models are also renewable and generalizable [16][17][18], enabling the retention and update of pre-trained models with new data and facilitating their application to other AM processes that can be analyzed using the updated ML models or transfer learning. Furthermore, ML can address various auxiliary challenges in AM, including cost estimation [19], manufacturability assessment [20], and closed-loop quality control [21], which are traditionally difficult to tackle using conventional physics-based methods. ...
The advancement of additive manufacturing (AM) technologies has facilitated the design and fabrication of innovative and complicated structures or parts that cannot be fabricated with traditional subtractive manufacturing processes. To achieve the desired functional performance of a specific part, quality and process should be well monitored, controlled, and optimized with advanced modeling techniques. Despite the effectiveness of existing physics-based and data-driven methods, they have limitations in providing generalizability, interpretability, and accuracy for complex metal AM process optimization and prediction solutions. This work emphasizes Physics-Informed Machine Learning (PIML) as a significant recent development, embedding physics knowledge (e.g., thermomechanical laws and constraints) into Machine Learning (ML) models to ensure their reliability and interpretability, as well as enhancing model predictive accuracy and efficiency while addressing the limitations of traditional approaches. The paper further classifies PIML into three categories, emphasizing physics integration in terms of Physics-Informed Domain Knowledge, Simulation-Based Input Data, and Physics-Guided Model Training. In this context, the Physics-Informed Neural Network (PINN) serves as a notable example of Physics-Guided Model Training. PINN is particularly noteworthy for its ability to yield more explainable and reliable results in forward problem solving, even with noisy training data. In addition, the paper further discusses the limitations and potential solutions of PINN.
... Such empirical models, known as "black box" models, can mathematically describe the relationships between welding parameters and its joint performance, but cannot explain the underlying bonding mechanisms [3,16]. Thus, the model performance and training convergence are strongly dependent on the selection of the model hyperparameters and unexpected errors could appear as extrapolating beyond the observation range [16][17][18][19]. ...
In recent years, ultrasonic metal welding (USMW) has gained considerable research attention due to the increasing application of lithium-ion batteries in electric vehicles. Although USMW is widely used in battery manufacturing, the fundamental understanding and monitoring of the welding process needs to be investigated to ensure the process robustness. Therefore, this work aims to establish in-situ observation of the bond formation process by using Laser-Doppler Vibrometry (LDV). The acquired velocity data was analyzed in the time and time- frequency domains. Based on the relative motion, a five-stage model was developed to describe the welding process. The changes in system dynamics led to the formation of sidebands in the frequency domain. Addi- tionally, the power consumption was recorded by the welding system, which had a good correlation with the system dynamics. The results provided the scientific understanding behind the bond formation process and suggested relevant signals for in-process monitoring.
... Le- Hong et al. (2021) investigated the effects of two key process parameters of SLM (i.e., laser power and scanning speed) on the single-track morphologies and the bead characteristics, especially the depth-to-width D/W and height-to-width H/W ratios using ML approaches (i.e., random forest and ANN). Both models could predict reasonably well the two aspect ratios, D/W and H/W, with an overall R 2 value reaching about 90%, respectively. ...
Laser beam powder bed fusion (LB-PBF) is a widely-used metal additive manufacturing process due to its high potential for fabrication flexibility and quality. Its process and performance optimization are key to improving product quality and promote further adoption of LB-PBF. In this article, the state-of-the-art machine learning (ML) applications for process and performance optimization in LB-PBF are reviewed. In these applications, ML is used to model the process-structure–property relationships in a data-driven way and optimize process parameters for high-quality fabrication. We review these applications in terms of their modeled relationships by ML (e.g., process—structure, process—property, or structure—property) and categorize the ML algorithms into interpretable ML, conventional ML, and deep ML according to interpretability and accuracy. This way may be particularly useful for practitioners as a comprehensive reference for selecting the ML algorithms according to the particular needs. It is observed that of the three types of ML above, conventional ML has been applied in process and performance optimization the most due to its balanced performance in terms of model accuracy and interpretability. To explore the power of ML in discovering new knowledge and insights, interpretation with additional steps is often needed for complex models arising from conventional ML and deep ML, such as model-agnostic methods or sensitivity analysis. In the future, enhancing the interpretability of ML, standardizing a systemic procedure for ML, and developing a collaborative platform to share data and findings will be critical to promote the integration of ML in LB-PBF applications on a large scale.
... Despite their promising results, ANNs are commonly classified as "black box" models (Oliveira et al., 2015) since, for example in the context of joint strength prediction, they do not explain the underlying bonding mechanisms that give rise to increased joint performance. Furthermore, the model's performance and its training convergence have a strong dependency on the selection of the model hyperparameters (Le-Hong et al., 2021). Therefore, identifying an appropriate set of hyperparameters is essential to obtain acceptable results. ...
The use of composite materials is increasing in industry sectors such as renewable energy generation and storage, transport (including automotive, aerospace and agri-machinery) and construction. This is a result of the various advantages of composite materials over their monolithic counterparts, such as high strength-to-weight ratio, corrosion resistance, and superior fatigue performance. However, there is a lack of detailed knowledge in relation to fusion joining techniques for composite materials. In this work, ultrasonic welding is carried out on a carbon fibre/PEKK composite material bonded to carbon fibre/epoxy composite to investigate the influence of weld process parameters on the joint’s lap shear strength (LSS), the process repeatability, and the process induced defects. A 3 ³ parametric study is carried out and a robust machine learning model is developed using a hybrid genetic algorithm–artificial neural network (GA–ANN) trained on the experimental data. Bayesian optimisation is employed to determine the most suitable GA–ANN hyperparameters and the resulting GA–ANN surrogate model is exploited to optimise the welding process, where the process performance metrics are LSS, repeatability and joint visual quality. The prediction for the optimal LSS was subsequently validated through a further set of experiments, which resulted in a prediction error of just 3%.
The printed quality of laser directed energy deposition (L-DED) technology is significantly influenced by the characteristics of the molten pool. Therefore, it is crucial to perform in-situ monitoring of the molten pool and accurately predict track geometric features in order to effectively control the processing quality. In this study, an innovative feature fusion deep learning (FF-DPL) network is proposed, which combines both dynamic image features and static processing parameters to significantly improve the prediction accuracy by utilizing only a smaller dataset and requiring fewer data features. Width average prediction accuracy reaches an impressive 98.45%. It is worth noting that on the extrapolation test dataset, the FF-DPL model achieves an average prediction accuracy in depth and height that is over 10% higher than the other six popular deep learning models. The extracted features of the images under different convolutional layers are visualized, confirming the significance of spatter and steam plume features for track geometry prediction. Finally, a correlation analysis is performed to investigate the relationship between input parameters and the track geometry, revealing the influence trends and magnitudes of different processing parameters on the variation of the track geometry, which provides important guidance for controlling the molten pool size in practical engineering production.
Purpose
316 L stainless steel alloy is potentially the most used material in the selective laser melting (SLM) process because of its versatility and broad fields of applications (e.g. medical devices, tooling, automotive, etc.). That is why producing fully functional parts through optimal printing configuration is still a key issue to be addressed. This paper aims to present an entirely new framework for simultaneously reducing surface roughness (SR) while increasing the material processing rate in the SLM process of 316L stainless steel, keeping fundamental mechanical properties within their allowable range.
Design/methodology/approach
Considering the nonlinear relationship between the printing parameters and features analyzed in the entire experimental space, machine learning and statistical modeling methods were defined to describe the behavior of the selected variables in the as-built conditions. First, the Box–Behnken design was adopted and corresponding experimental planning was conducted to measure the required variables. Second, the relationship between the laser power, scanning speed, hatch distance, layer thickness and selected responses was modeled using empirical methods. Subsequently, three heuristic algorithms (nonsorting genetic algorithm, multi-objective particle swarm optimization and cross-entropy method) were used and compared to search for the Pareto solutions of the formulated multi-objective problem.
Findings
A minimum SR value of approximately 12.83 μm and a maximum material processing rate of 2.35 mm ³ /s were achieved. Finally, some verification experiments recommended by the decision-making system implemented strongly confirmed the reliability of the proposed optimization methodology by providing the ultimate part qualities and their mechanical properties nearly identical to those defined in the literature, with only approximately 10% of error at the maximum.
Originality/value
To the best of the authors’ knowledge, this is the first study dealing with an entirely different and more comprehensive approach for optimizing the 316 L SLM process, embedding it in a unique framework of mechanical and surface properties and material processing rate.
Optimum process parameter window for recently developed metal powders used for laser powder bed fusion (LPBF) is strongly correlated with characteristics of each single track and single layer. In the present research, the influence of LPBF process parameters on geometrical and microstructural characteristics of A20X aluminum single tracks was studied theoretically and experimentally. Increasing the laser scan speed led to formation of non-homogenous, irregular single tracks. Moreover, enlarging the powder layer thickness to 120 µm induced balling phenomenon owing to a poor wetting of substrate by melt pool. Experimental measurements indicated that width and depth of melt pools decreased up to 53 % and 68 %, respectively by increasing the scan speed from 500 to 1700 mm/s and layer thickness from 40 to 120 µm. These findings were in agreement with predictions of an exponentially decaying heat input model used in this study. The model took into account thermo-physical properties of the alloy at different states. Bead height was observed to increase with the powder layer thickness, but remain unchanged with the scan speed. The process parameter window (energy density range) resulting in the conduction melting mode for the A20X alloy was found to be more than two times wider compared to conventional LPBF aluminum alloys. For all process parameters, a fine equiaxed grain structure was observed in the vertical section of the single tracks. Enhancing the scan speed and reducing the powder layer thickness resulted in a substantial grain refinement. This was well explained by the simulated cooling rates. In contrast to cast Al-Cu alloys, shape, size, and volume fraction of the second phase precipitates were determined to be independence of cooling rate over the investigated cooling rate range.
