Analysis of the influence of FDM parameters in the tensile strength response using machine learning

This research presents a comprehensive experimental study on the effect of temperature, material and process parameters related to tensile strength response in 3D printing manufacturing process with ABS material. A Hyper Latin Square design was chosen for the experimental points distribution. Thirteen parameters with multiple levels are considered LAYERHEIGHT, WALLTHICKNESS, TOPBOTTOMTHICKNESS, TOPBOTTOM-LINEDIRECTION1, TOPBOTTOMLINEDIRECTION2, INFILLDENSITY, INFILLLINEDI-RECTION1, INFILLLINEDIRECTION2, PRINTSPEED, EXTRUSIONTEMP, BEDTEMP, WORKSPACETEMP and POSITION. Type IV tensile specimens are fabricated and tested with an universal testing machine. Maximum stress is measured, evaluated and analyzed in three different building positions.Machine learning algorithm with Orange Data mining software are used to study underlying relations between factors and response. Experimental results indicate that INFILLDENSITY, TOPBOTTOMTHICKNESS and INFILLLINEDIRECTION1 has a strong positive correlation with tensile strength. Meanwhile, TOPBOTTOMLINEDI-RECTION1, WORKSPACETEMPERATURE and PRINTSPEED has a negative correlation with tensile strength. Position 1 with line depositions parallel to Y axis produce the higher tensile strength response. Findings imply that machine algorithms can be used to study multiple parameters at time.


Introduction
Additive manufacturing (AM) is a component manufacturing process by modeling in computeraided design (CAD) software and then mechanically bonding materials, which are layered one on top of the other.There are different types of MA, one of them is fused deposition modeling (FDM), patented in 1989 [1].
The most popular method of FDM is 3D printing, in which various process parameters must be controlled, such as process parameters (deposition angle, layer height, fabrication direction, deposition width, print speed, density of fill, air gaps, fill pattern, extrusion temperature), 1299 (2024) 012003 IOP Publishing doi:10.1088/1757-899X/1299/1/012003 2 workspace parameters (bed temperature, chamber temperature, humidity, nozzle diameter) and material parameters (material type, color, diameter of the filament) [2].Regarding 3D printing, the parameters and their influence on the properties and behavior of the manufactured part have been investigated.Much of this research is oriented toward mechanical properties such as resistance to tension, compression, bending, torsion and impact energy absorption.[1] [2] [3].

3D print Material
The material used for the study was acrylonitrile butadiene styrene (ABS).It is the most commonly used petroleum-derived plastic in 3D printing.It is characterized by its high mechanical and thermal resistance and its great flexibility.The ABS is from the eSUN brand, red color, with the properties and printing parameters recommended in Table 1.

3D printing machine
The printer used for the fabrication of the test specimens is a INTAMSYS FUNMAT HT (Figure 2) with a build volume of 260 x 260 x 260 mm with 50-micron high-resolution industrial quality 3D printing.The thermal system design of this 3D printer includes a 90 °C constant temperature chamber, 160 °C heated build plate and 450 °C high-temperature extruder with an all-metal hot end.

Printing process parameters
Ten process parameters are included for the research:

Factor parameter levels
A Latin Hypercube design was chosen to assess the impact of the process factors on the output.This experimental design offers a more uniform spreading of the random sample experimental points [22].Level ranges were established base on previous researches and are show in Applying the LHS for the distribution of the experimental points, a set of data is obtained as shown in Table 3 2.5.Printing process Models from SolidWorks were saved as .stltype and preprocessed in Ultimaker Cura slicer software.Tests specimens were printed in the three positions with the same parameters to compare the effect of the same experimental point in the three building positions, as shown in Figure 3. Testing specimen Specimen for the tension test is a ASTM D638 Type IV [23].The specimen was modeled in SolidWorks software with dimensions as in Figure 4.
The Universal testing machine is a 50 kN Armsfield with a ±0.5% precision.Test's velocity was set in 1 mm sec .After the test, stress-strain curves and results were registered in the machine software.Yield, ultimate and break stress, and maximum load were measured from the graphics.Young's modulus was calculated from the slope of the stress-strain curves.
IOP Publishing doi:10.1088/1757-899X/1299/1/0120036 2.7.Machine learning processing For machine learning analysis, a visual data mining tool was applied.Orange Data Mining [24] is a computer program for data mining and predictive analytics developed at the Faculty of Informatics at the University of Ljubljana.It consists of a series of components developed in C++ that implement data mining algorithms, as well as data preprocessing and graphical representation operations as in Figure 5.The principal advantage is that to explore data with Orange requires no programming or in-depth mathematical knowledge.4 show calculated results for the tension tests.

Outliers
Outliers were calculated with local outlier based on the local density from the k-nearest neighbors factor method [25].Three outliers were found, run 6, 7 and 8, as shown in Figure 6.
Hereinafter, next results are analyzed with the inliers points, ignoring outliers.

Correlation results
For the correlation's analysis, Pearson method was computed [26].Results in Table 5 shows that positive and negative correlation exist between factors and the maximum stress.INFILLDENSITY has the biggest correlation coefficient among the positive correlations, followed by TOPBOTTOMTHICKNESS and INFILLLINEDIRECTION1. TOPPBOTTOMLINEDIRECTION1 has the biggest negative correlation factor, with WORKSPACETEMPERATURE and PRINTSPEED as the only three negative.
INFILLDENSITY, TOPBOTTOMTHICKNESS and INFILLLINEDIRECTION1 shows a strong positive correlation (¿0.5), meanwhile WALLTHICKNESS, EXTRUSSIONTEMPERA-TURE and INFILLDIRECTION2 swows weak positive correlation.LAYERHEIGHT, TOP-BOTTOMLINEDORECTION2 and BEDTEMPERATURE has a correlation closes to zero (weak correlation).Printing Position 1 has the biggest mean and the higher value in respect to tensile stress response.A value of 20.56 MPa against 17.30MPa and 16.48MPa for position 2 and 3.

Classification tree algorithm
Classification tree algorithm works well with numerical and categorical variables.In this case, we can use decision tree to analyze numerical variables with the printing position.

Conclusions and Future Works
Fourteen parameters were analyzed using machine learning techniques for a 3D printing precess.Significant correlation were identified, with strong and weak relations between parameters and the tensile strength response.
It was found that INFILLDENSITY, TOPBOTTOMTHICKNESS and INFILLLINEDIREC-TION1 increase the tensile strength response.
To a lesser extent, the factors WALLTHICKNESS, EXTRUSIONTEMP and INFILLLINEDI-RECTION2 produce an increase in the response.
Printing position 1 produce the best tensile strength response compared with the other two positions.
For future works, research on the deposited line directions in the infill and top-bottom with diffeferent infill patterns shall be considered.

Figure 3 .
Figure 3.The three positions of the testing specimens

Figure 4 .
Figure 4. ASTM D638 Type IV Tension test specimen (dimensions in mm)

3. 4 .
Visual Explorative data analysis Three visualization techniques for explorative data analysis, Linear Projection, RadViz y FreeViz were used to analyze relations between factors and the tension stress response, and compare with correlation results.Visual tools are show in Figures 7, 8 and 9.The trend of the positive and negative factors is the same as in Table 5, with INFILLDENSITY, TOPBOTTOMTHICKNESS and INFILLLINEDIRECTION1 in direct strong correlation with the response.

Figure 11 .
Figure 11.Box plot for POS Vs. maximum stress.

Table 2 Table 2 .
Factors an range for the experimental points sample design

Table 3 .
Experimental points generate with LHS

Table 4 .
Tension test results