Experimental study and modeling the tensile strength of 3D‐printed aluminium polylactic acid (PLA) parts using artificial neural networks

Introduction

Additive manufacturing processes such as 3D printing are widely used today in many industries. 3D printing has become more widely used in industries, increasing the number of consumers in need of 3D printed products of high quality.¹ As a result of increased competition, 3D parts with more precise tolerances, better product finish, and overall durability have become increasingly popular. Despite the fact that 3D printing costs are on the decline, part quality issues such as integrity, strength and aesthetics are still a concern.² Because of this, industries have been driven to continuously improve quality control of their products. These quality measures are very important to mechanical parts because it affects the cooperation of parts together, working accuracy, physical properties and chemical properties altogether. 3D printing industries rely on these qualities to produce better-looking and more functional parts while also saving money by reducing the amount of time needed to manufacture the parts.³ The parameters that are taken into considerations are layer thickness, infill density and number of shells.⁴ The product's tensile strength is a critical factor to take into account. In this study, a central composite design experimental plan was developed with three influencing input parameters, namely infill density, layer thickness and number of shells. Aluminium polylactic acid (PLA) materials were used as a work material. Tensile test was conducted on the specimen and recorded. Using the data, an artificial neural network (ANN) model was developed to predict the tensile strength of new combination of parameters, and a validating test carried out on the model.

Methods

This study was conducted in November 2020. A design of experiment (DOE)/Central Composite Design (CCD) experiment plan was designed with three input parameters (layer thickness, infill density and number of shells) shown in Figure 1. There are eight corners, six face centers, and one center, for a total of 15 specimen points.^4,5 Table 1 shows the three levels of parameters. Table 2 displays the experiment plan as well as the test results.

Figure 1. Central composite design.

Tensile test specimens were printed in the Makerbot 3D printer. PLA filament mixed with 15% aluminium powder (Al-PLA- Fabbxible, 1.75 mm diameter) was used as work material. One tensile test was carried out using the Instron 3360 series dual column tabletop equipment. The results were captured using Instron Bluehill- Universal 3 software. American society for testing of materials (ASTM). D-638 standard size was used to prepare the specimen.

Artificial neural networks (ANN) are artificial computing systems that mimic biological neural networks, which are similar to the human brain. The system has self- learning capabilities to perform tasks or produce better results.⁶ Using the data, an ANN model was developed to predict the tensile strength of newer combinations of input parameters, and a validating test carried out on the model. Matlab R2021b neural network software (Scilab is an open-source alternative) was used to create a neural network with the printer process parameter and tensile strength data (Table 2).

In order to develop the ANN model, the experimental data were imported into Matlab software. With the deep learning toolbox feature, a framework for creating and designing neural networks was made. Whilst giving the option to import, create and export networks and data, it can also generate a neural network system which can be designed, simulated and analysed.

Results and discussion

There are four specimens in Figure 2 that stand out as having the highest values of tensile strength. Other specimen includes 1, 3, 5 and 7. They have the same layer of thickness (0.1 mm). This indicate that the lower the layer thickness the higher the tensile strength. As an example, if we look at specimens 13, 14, and 15, they all have the same number of layers (0.3 mm) and three shells, but different infill density levels. Specimen number 13 that has the lowest infill density of 20% appears to have the lowest tensile strength of 0.71 kN. As a result, the breaking load decreases as infill density decreases. There are no differences in layer thickness or infill density between specimens 1 and 3, but there is a discrepancy in the shell count between specimens 1 and 3. In spite of the fact that specimen 3 has a much higher number of shells than specimen 1, it has a slightly higher breaking load. The same is true for specimens 2 and 4, which have similar values of layer thickness (0.3 mm) and infill density (20%). Specimen 4 have a very slight increase of breaking load compared specimen 1, even though specimen 4 has a much higher number of shells compared to specimen 2. Specimens 5 and 7, which have similar values of layer thickness (0.1 mm) and infill density of 60%. Specimen 7 also has a very slight increase of breaking load compared specimen 5, even though specimen 7 has a much higher number of shells compared to specimen 5. This means that the values of layer thickness and infill density have a significant impact on the value of tensile strength. However, the number of shells does not the effect the variation of breaking load in test specimens. The tensile strength increases with decreasing layer thickness. The specimen's tensile strength increases as the infill density increases.⁵

Figure 2. Tensile strength of all specimens.

An ANN model was created as shown in Figure 3, by entering input such as layer thickness, number of shells and infill density and tensile strength data as a target data.

Figure 3. ANN model.

Errors in validation and training can be used to gauge a trained network's performance as a whole. Performing a regression plot between the network response and the targeted values, as shown in Figure 4, is one way to achieve this goal. We can tell if the neural network is desirable or not by generating the regression plot. The regression plot depicts the relationship between a network response variable (output) and target variables. If we had a perfect fit, where outputs exactly equal to targets, the slope would be of value “1” and the y-intercept would be at point (0, 0). This condition is considered to be desirable, which means a good correlation of targeted values to output values. Vice versa, if the slope is not close to value “1”, then it can be considered as not desirable; such a condition is shown in Figure 4a. If it is not desirable, we can retrain the network by closing the window and clicking on train network again until we get a regression plot’s value for all that is close to the value “1” as shown in Figure 4b. As a result of oscillating between input and output values, the neural network learns and repeats the process. This will eventually give a better regression model close to the value “1”. It is a good idea to train several networks to ensure that a network with good generalization is found. Figure 4b shows a good regression plot. According to Figure 4, the open circles represent the values of network output plotted against the goals. The best fitted slope is indicated by the dashed line and the solid line indicates the fitted slope generated by the open circles.

Figure 4. (a) Undesirable regression plot, (b) good regression plot.

According to Figure 4a, Matlab's first trained network isn't ideal because the regression slope value for all is 0.94767 and the y-intercept isn't close to 0. By retraining the network, we are able to get results such that in Figure 4b, which is our desirable regression. Next simulating by going to the simulate tab, selecting the inputs and clicking simulate network. Using the developed model tensile strength was predicted the new set of input parameters and tabulated in Table 3.

To validate the developed ANN model, 5 new test specimens were fabricated as shown in Table 3. The five test specimens were then subjected to a tensile test, and the results were tabulated. Based on the table, the specimen with the lowest layer thickness has the highest tensile strength with 1.05kN. This can be validating with what we have discussed previously where the lower the layer thickness, the higher the tensile load. The percentage difference of experimental values and theoretical values for all the specimens are less than 5%.

According to Table 3, the percentage difference between experimental data and ANN model predicted data for all specimens is less than 5%, which proves the model's validity and reliability.

Conclusions

An ANN model was successfully developed and validated using Matlab to predict the 3D printed aluminium part data. The three variables of layer thickness, number of shells, and infill density have been shown to affect the tensile strength of the 3D printed specimens. The specimen with the lowest layer thickness of 0.1 mm, four shells, and a 20 percent infill density had the highest tensile strength of 1.12 kN, based on the results of the tests. In addition, specimen 13 has the lowest tensile strength of 0.71 kN when compared to specimens 14 and 15, which all have the same layer thickness (0.3 mm) and number of shells (3) but different infill density. For this reason, the thicker layers have lower tensile strengths. However, as the number of shells and infill density increases, the tensile strength increases. In summary, the 3D printed part with the lowest layer thickness, highest infill density, and most shells has the highest maximum tensile strength, according to the research. A comparison of predicted and experimental results allowed us to verify the accuracy of the ANN model.

Data availability

All data underlying the results are available as part of the article and no additional source data are required.