Microstructural analysis and mechanical properties of AlSi10Mg alloy components produced via SLM

. Selective Laser Melting (SLM) is a manufacturing process that involves layer-by-layer fusion of metal powder material using a laser beam. AlSi10Mg alloy, a hypoeutectic aluminum-silicon alloy, is known for its favorable mechanical properties, low density, and corrosion resistance, making it widely used in various industries. While traditionally used for casting, the unique properties of AlSi10Mg alloy have opened up opportunities for additive manufacturing, including SLM technology. This paper explores the tracks quality and preferable parameter interval for production via SLM. SLM products exhibit unique microstructures with distinct advantages and limitations. This study underscores the importance of careful parameter selection during SLM and the need for effective quality control in achieving desired mechanical properties


Introduction
Selective Laser Melting (SLM) is a transformative additive manufacturing technique that has gained significant attention in recent years.It involves the precise layer-by-layer fusion of metal powder material using a laser beam, allowing for the creation of intricate and complex geometries.Among the materials used in SLM, the AlSi10Mg alloy has emerged as a notable candidate.This alloy belongs to the category of hypoeutectic aluminum alloys known as silumins, characterized by their high silicon content.
Silumins, including AlSi10Mg, have been traditionally favored in industries such as automotive, aerospace, and military applications due to their excellent mechanical properties, low density, corrosion resistance, and castability.However, the application of AlSi10Mg alloy in additive manufacturing, particularly through SLM, has opened new avenues for the production of components with unique properties.In this context, it becomes essential to thoroughly investigate the microstructure and mechanical properties of products manufactured using AlSi10Mg alloy via.Understanding the distinctive characteristics of SLM-produced components is crucial for determining the feasibility of utilizing this technology for specific applications.[1][2] The microstructural differences between SLM products and those produced by other methods, such as powder metallurgy, are significant.Notably, SLM products exhibit a distinct grain structure, with silicon particles distributed along grain boundaries.These differences have a direct impact on mechanical properties, including yield strength.While SLM products excel in certain mechanical properties, such as yield strength, they fall short in terms of ductility when compared to traditionally produced components with finer and more uniform grain structures.Furthermore, the relationship between microstructure and mechanical properties in AlSi10Mg alloy remains complex and warrants further investigation.For instance, while castings may exhibit superior strength characteristics after appropriate heat treatment, they do not display a straightforward correlation between microstructure and mechanical properties.[2][3][4] One of the key challenges in SLM-produced components is the presence of porosity and single-track quality, which significantly affects tensile strength, yield strength, and hardness.Achieving high mechanical properties in SLM products requires meticulous parameter selection during the experimental technological process and the development of effective quality control measures.This paper aims to delve deeper into tracks quality and appropriate window parameters for further investigations.[5] 2 Methods Addsol D50 (figure 1) was used to fabricate single tracks for continuity inspection and further cross-section analysis.AlSi10Mg powder particles with the size below 50  were used and the layer thickness for the single track was chosen 50  (equals to size of the laser beam).To narrow down preferable parameters interval continuity inspection was conducted.Convolutional Neural Networks (CNNs) can learn complex thresholding functions from data for image segmentation tasks.
Finding optimal learning schedule (figure 2) is critical in case of limited dataset and available GPU resources.The logic behind "planning" this so-called schedule is to if you're starting with a large learning rate and then reduce it once training stops making fast progress, good solution can be reached faster than with the optimal constant learning rate.There are many different strategies to reduce the learning rate during training.It can also be beneficial to start with a low learning rate, increase it, then drop it again [7].Comparing CNN and "traditional" analytical algorithm, the CNN are of course powerful but computationally intensive and require substantial data for training, making them wellsuited for complex computer vision tasks.Otsu's Method, on the other hand, is a simpler and more computationally efficient technique primarily used for basic image thresholding and segmentation, making it suitable for real-time applications and situations where labeled data is scarce.The choice between them depends on the specific requirements and complexity of the task at hand.Due to the limited number of images available and desire to maximize IoU (intersection over union) authors used following data preparation and training process: 1. Using a Pretrained Model (U-Net): In this study, the authors opted to employ a pretrained model, specifically the U-Net architecture.Pretrained models are neural network models that have been previously trained on a large dataset for a related task.The selection of U-Net is noteworthy as it is a convolutional neural network architecture known for its exceptional performance in image segmentation tasks, particularly in the medical imaging domain.U-Net's architecture is characterized by a contracting path followed by an expansive path, making it highly suitable for semantic segmentation tasks where pixel-level accuracy is essential.By leveraging a pretrained U-Net model, the authors could capitalize on the model's learned features, which are advantageous for their specific task, likely related to image segmentation.[8][9] 2. Augmentation of the Dataset: Data augmentation is a critical step in the data preparation process for training deep learning models, especially when dealing with a limited number of images.Augmentation techniques involve creating new training examples by applying various transformations to the existing dataset, such as rotation, scaling, cropping, flipping, and introducing noise.By augmenting the dataset, the authors effectively increased the diversity of the training samples, which can help prevent overfitting and improve the model's generalization capability.This augmentation process likely allowed the model to learn a broader range of features and variations present in the data, ultimately enhancing its ability to make accurate predictions.
3. Adaptive Learning Schedule: An adaptive learning schedule is a strategy for adjusting the learning rate during the training of a deep neural network.The learning rate is a hyperparameter that determines the step size at which the model's weights are updated during gradient descent optimization.An adaptive learning schedule typically involves dynamically changing the learning rate based on the model's performance or the progress of training.[10] This can be done using techniques such as learning rate decay, learning rate warm-up, or using learning rate schedules like cyclical learning rates.By employing an adaptive learning schedule, the authors aimed to fine-tune the training process, ensuring that the model converges to an optimal solution effectively.This strategy can be particularly beneficial for achieving a high Intersection over Union (IoU), which is a metric commonly used to evaluate the accuracy of image segmentation models by measuring the overlap between predicted and ground truth regions.An adaptive learning schedule can help the model converge more quickly and effectively to maximize IoU.In our research we aimed to use CNN for image segmentation (figure 3) and ended up using traditional Otsu's method, more about this in the Results and Discussion section.In total after data augmentation, we have around ~600 images and corresponding masks.[11][12] Besides track's continuity, which is the simple boolean parameter we also used the so-called extent value -  , which is the ratio between total area of the binary mask (in other words track's area) to binary mask's bounding box (figure 4).Since we now the mask's area -  (initially in  2 ) there is no difficulty to obtain its bounding box's area -  : Based on this data quality map for extent value was plotted.We set the infimum cap for 0.3 (since extent parameter is always between   ∈ (0; 1)) for filtering out the tracks we consider continuous, but not enough sintered.[13] 2. Cross section analysis.
For the cross-section analysis only part of the original set was selected (about 30%) this is due to the extent and continuity filtering process.Contact angle and dimensions of the etched track were calculated.These parameters gave us understanding and opportunity to link the track width and height with its depth.[14] 3 Results and discussion 1.
CNN model training.Due to the limit with available GPU memory and considerably small dataset we trained U-Net models between 30 and 300 epochs.Here is the main scores and statistics: There is noticeable linear trend (figure 5).Firstly, due to the relatively small data set IoU is not exceeding 0.6.Secondly, we need to run model with more epochs to see convergence.This will be done in the future works with the extended dataset.2. Continuity and extent parameter.After obtaining binary masks extent parameter was calculated on the continuous tracks.Extent parameter vary between 0.4 ÷ 0.7 (figure 6).Based on the calculated extent parameter subset of the tracks were selected, forementioned three power columns.In the corresponding pictures with extent parameter in the  interval -[0.4; 0.7] contact angle varies between.To summarize, we can say that at an angle of less than 160 degrees, the cross sections of the tracks signal the instability of certain laser sintering processes.[15][16] Starting from angles of 160 degrees, the effect of a keyhole or a so-called mercury drop is observed (figure 8).
Considering forementioned filtering power-velocity interval is proposed.Authors suggest to print 3D sample, i.e., cubes or cylinders for porosity and strength evaluation in the interval framed in the figure 9.

Conclusion
1.The study utilized Addsol D50 to create single tracks for continuity inspection and cross-section analysis, employing AlSi10Mg powder particles of a specific size and chosen layer thickness .Various thresholding techniques, including traditional methods like Otsu's Method and deep learning-based approaches, were explored for continuity inspection.The choice of method depends on the complexity of the task and available data.
2. To address limited data availability, the authors employed strategies such as using a pretrained U-Net model, augmenting the dataset with diverse transformations, and implementing an adaptive learning schedule.These techniques likely contributed to improving the model's generalization and accuracy.
3. Extent Value and Continuity: In addition to assessing continuity, the study introduced the extent value (  ) as a parameter.Tracks with   values below 0.3 were considered insufficiently sintered.This approach provided a more nuanced understanding of track quality.
4. Cross-Section Analysis: A subset of the dataset was selected for cross-section analysis, focusing on contact angles and dimensions of etched tracks.These parameters allowed for the evaluation of track width, height, and depth, revealing insights into laser sintering processes.5. CNN Model Training: Due to GPU memory limitations and a small dataset, U-Net models were trained for a limited number of epochs.The Intersection over Union (IoU) scores showed a linear trend, with room for improvement through extended training with a larger dataset in future work.
6. Contact Angle: Contact angle analysis revealed that angles less than 160 ∘ indicated instability in laser sintering processes, while angles greater than 160 ∘ resembled a keyhole or "mercury drop" effect.
In summary, the study demonstrated a comprehensive approach to assessing the quality of laser sintered tracks, combining various methods for continuity inspection, data preparation, cross-section analysis, and contact angle evaluation.The findings suggest opportunities for future research to enhance the accuracy and robustness of the assessment process.

Fig. 1 .
Fig. 1.SLM machine Addsol D50. 1. Continuity inspection.The final goal of this inspection is conclusion about track continuity (simply speaking does it have unsintered sites).For this simple Python script was written for obtaining binary masks of the tracks.There is a different method for obtaining masks of the gray-scaled image (Global, Adaptive, Edge-Based, Color, Machine Learning-Based, Deep Learning-Based Thresholding) [6].Authors used: 1.1.Global Thresholding : 1.1.1.Otsu's Thresholding maximizes the inter-class variance to find an optimal threshold.1.1.2.Kapur's Entropy Thresholding maximizes the sum of entropies between classes and within classes.1.1.3.Huang's Thresholding minimizes the entropy within classes.