Investigation of the impact of power consumption, surface roughness, and part complexity in stereolithography and fused filament fabrication

Today, additive manufacturing (AM) is being utilized in a plethora of fields and applications. This study investigates two of the most commonly used AM technologies, fused filament fabrication (FFF) and stereolithography (SLA), under different part complexities and their effects to finished part quality. The process parameters investigated are infill density (ID), infill pattern (IP), and layer height (LH). The output variables observed are power consumption (PC), surface roughness (SR), and mass. Statistical analysis of experimental data indicates that LH is the most influential parameter for all output variables except mass. A validation study is later performed using three specimens with more diversely complex geometries. The model complexity is then evaluated and analyzed using three different ratios. Normalizing PC with respect to volume indicates a lower PC per unit volume for FFF. Experimental data also indicates SLA has a better surface finish and uses less power. This paper reports the state-of-the-art FFF and SLA process knowledge blocks created through this research study.


Introduction
Many industries are employing additive manufacturing (AM) to perform tasks otherwise performed by traditional means [1,2]. In addition, AM can produce ready-to-use parts and tools. The process of AM involves layer-by-layer forming of an object [3,4]. Within AM, there are seven processes [5] which include fused filament fabrication (FFF), stereolithography (SLA), powder bed fusion, directed energy deposition, binder jetting, material jetting, and sheet lamination. Material extrusion and fused deposition modeling are other names for FFF [6]; vat photopolymerization is another name for SLA [7]. This study focuses on FFF and SLA AM processes.
Arguably, the most used AM process is FFF, due to its ease of operation, relatively low cost, and high accessibility [8,9]. FFF machines employ polymers in 1.75-mm or 3.00mm diameter varieties rolled into a spool; these materials are typically referred to as filament. The filament is pushed through by a stepper-motor-driven gear. A heating element is then used to melt the filament and allow the flow through a nozzle. The material is deposited onto a platform, usually referred to as "print bed," layer by layer as shown in Fig. 1. The print bed is usually heated to help with adhesion and warping [3,10]. A FFF produced part is directly derived from a computer-aided design (CAD) file [11]. The CAD file is pre-processed accordingly to meet the FFF machines' parameters.
The CAD file is exported as a standard tessellation language (STL) file prior to pre-processing for fabrication. An STL file contains data that describes the shape of the solid object through representation of triangle coordinates usually referred to as facets [12]. Facets are used to approximate the surface characteristics of the model. An STL file can be viewed as a mesh for the surface of the model.
The pre-processing of a STL file to produce through FFF is referred to as slicing [14]. The slicing process sets all the process parameters for fabrication. These parameters include infill patterns, infill density, extrusion width, bed and nozzle temperatures, speed, and many others. This study will focus on changing and finding effects of three process parameters: infill density (ID), infill pattern (IP), and layer height (LH). The IP parameter controls the internal structure of a model, while the ID controls how often the IP is repeated within a structure. LH is best described as the vertical resolution of a model. These parameters affect the physical attributes of the model, which directly affects the mechanical and physical properties of the fabricated part.
The other contender for the most used AM process is SLA, also known as Vat photopolymerization [15]. SLA has recently become highly accessible while being low cost and easy to use. Instead of polymers in spool form, SLA machines employ liquid photosensitive resins as their material. The liquid resins are solidified using a light source, i.e., laser, in a controlled manner. SLA machines have build-platforms that get submerged in the liquid resin before retracting, allowing excess resin to drip back in the reservoir. The light source is then used to solidify the resin according to the model. Like FFF, SLA also requires pre-processing of a STL file.
Pre-processing of a STL file for SLA is also called slicing. The process differs immensely from FFF pre-processing. The SLA slicing process does not allow for setting infill parameters, as SLA parts are usually fully solid [16,17]. The only mainstream overlapping process parameter between SLA and FFF is layer height. Unlike FFF, the parts produced using SLA almost always require post-processing [18]. This is mainly due to the supports that are required for SLA parts. Other post-processing tasks include washing the parts to eliminate any undesired uncured resin. Afterwards, parts are cured using a designated curing station or sunlight. Without curing, the mechanical properties of SLA parts are typically inferior [19].
These two most commonly used AM technologies, FFF and SLA, are being heavily utilized in various industries in recent times. They have their limitations along with advantages and disadvantages. As an example, FFF-fabricated parts will always display visible layer lines that are produced due to the nature of the process, while SLA is superior in producing a perfectly smooth surface [20]. On the other hand, FFF is known to be capable of lowering fabrication times and larger components with entry-level FFF machines, while SLA machines capable of such properties are of highlevel and higher cost.
In addition, research studies have claimed that "complexity is free" when it comes to AM [21], with evidence supporting that claim. Part complexity in traditional manufacturing (TM) often leads to higher production cost and time [22]. The increase in cost, in subtractive manufacturing for example, is typically due to the need to change set ups, toolheads, and several other factors, which is usually the case for moderately complex structures [23]. In AM on the other hand, the nature of the process allows efficient production of diversely complex structures without interference or the need to change tools.
Kim et al. [24] aimed to test different AM technologies, including FFF and SLA, by investigating their mechanical properties, geometrical accuracy, surface roughness, speed, and cost of materials. From a dimensional accuracy and surface roughness standpoint, SLA yielded parts with the highest dimensional accuracy and lowest surface deviation. Jayaram et al. [25] investigated several AM technologies including SLA and FFF. A test artifact is designed to be of varied complexities to later be tested. Results indicated  [13] that SLA was the most esthetic AM process compared to all the others, considering SLA had all the intricate features fabricated with the lowest surface roughness, while stating that FFF parts required the least post-processing steps. Choudhari et al. [26] explored the benchmarking of a few different AM processes and discovered that SLA-produced parts displayed the least shrinkage and the best surface finish compared to FFF-produced parts. In addition, the authors stated that FFF is ideal for low-cost production since the material used is significantly cheaper than materials used in other AM technologies. Terry et al. [27] gave a look from a few different aspects into the advantages of using smart manufacturing, including AM. Their review was focused on the advantages of energy saving and production efficiency. Hinshaw et al. [28] experimented with different FFF process parameters (LH, ID, side shells) to investigate energy consumption. Their results indicate the highest layer height consumed the least energy, as expected.
Brajlih et al. [29] introduced four methods to identify geometric complexities of 3D models based on the corresponding STL files. The four methods define complexity based on the number of facets and part volume divided by the number of facets, bounding box volume, and surface area of the model. Merkt et al. [21] investigated the three complexity measures Brajlih et al. proposed to be added to an integrative technology evaluation model (ITEM). The model is supposed to identify whether selective laser melting (SLM) is the most efficient process based on the three complexity factors. They concluded that their model is not complex for SLM process but is complex for machining and casting processes.
In this study, an experiment is conducted to understand how process parameters affect power consumption, mass of the final parts, and surface roughness for SLA and FFF technologies. Thereafter, another experiment is conducted to validate the results obtained. This experiment employs complex designs as opposed to simple models. The methods used for geometric complexity identification by previous literature are used to investigate how power consumption responds to the different complexity factors for both SLA and FFF. As far as the authors know, no literature explicitly compares complexity factors to power consumption.

Initial parametric study
The main objective of this study is to explore how some process parameters affect the desired output variables for FFF and SLA technologies. The data resulting from this study is later used for further validation. The process parameters are layer height (LH), infill pattern (IP), and infill density (ID). For each process parameter, three levels are investigated. The response variables in question are power consumption (PC), mass, and surface roughness (SR). A 3 × 3 × 3 factorial design is employed to conduct this experiment, shown in Table 1.

Test specimens
For the parametric study, half-inch cubic specimens are used, shown in Fig. 2. The cubic specimens are designed and edited in ANSYS SpaceClaim. To reiterate, IP and ID are two of the process parameters investigated. However, SLA fabrication does not usually employ IP or ID parameters. For this reason, SpaceClaim is used to set the infill parameters considering it edits the STL file itself and not as a g-code. The STL file can then be imported to both slicing software. In addition to setting infill parameters, one of the cubes' faces is removed to ensure no resin is trapped in the SLA process. The infill patterns chosen are simple yet highly utilized infill patterns, thus increasing relatability.

Setup and equipment
The machine used for FFF fabrication is a Raise3D Pro2 [30] with the Raise3D-developed slicer, ideaMaker. The material used is polylactic acid (PLA) by Walleye. The machine used for SLA fabrication is a Formlabs Form 2 [31] paired with Preform slicing software. The material used for SLA fabrication is V4 Clear resin developed by Formlabs. The  [32] measuring device is used. To log the data captured by the measuring device, Logger Pro software is used. The power logger device is plugged into an outlet, while the FFF or SLA machine is plugged into the power logger device. Another connection is required between the logging software and the power logger, which is achieved via USB cable to a laptop computer with the software installed. It should be noted that for SLA, curing time and energy is not considered in this study, because the curing station is not a vital component of the SLA process, and the parts can be cured in the sun without consumption of energy.
Mass measurements are simply taken by placing the cubic specimens on a measuring scale. The scale used is a Mettler Toledo PL-602S [33], which measures with resolution up to a 100th of a gram considering the specimens are similar in mass and higher resolution is needed. The specimens, specifically SLA specimens, are stripped of supports and fully cured prior to scaling.
For SR measurements, a Mitutoyo SJ-210 [34] profilometer is employed. The profilometer uses a needle-size probe, which is positioned to be dragged across the layer lines. Measuring across layer lines yields meaningful data, as it is measuring the roughest area of an additively manufactured part. Probing along layer lines may result in influencing the probe to stay between two-layer lines which yield inconclusive data.

FFF vs. SLA
While simply looking at results can yield conclusions about the two technologies and how they compare, further statistical analysis is required. The analysis indicates that SLA used, on average, about 23% of the energy used by FFF to fabricate the same parts. Furthermore, SR measurements, on average, indicate that SLA-fabricated cubes exhibit 72% less surface deviation than FFF-fabricated cubes. Because densities of the clear resin and PLA differ, mass comparison between the two is redundant and not in the scope of this study. Comparison is done to see how process parameters affect the mass within each technology. Build time, however, is also recorded to draw conclusions on which technology achieved lower fabrication times. On average, SLA-fabricated cubes took 141 min to be produced, while FFF-fabricated cubes took 121 min. This accounts to only a 16% increase in fabrication time for SLA over FFF.
The build volumes of the two machines used in this study vastly vary in size. The build volume of the Raise3D Pro2 is 30.4 × 30.4 × 30.0 cm (12 × 12 × 8 inches), while Form 2 is 14.5 × 14.5 × 17.5 cm (5.71 × 5.71 × 6.89 inches). This further suggests the machines are built and equipped to handle different volumes of parts. For this reason, it is beneficial to normalize the PC results with respect to each machine's volume for another aspect of comparison. For the Pro2, the PC per unit volume yields 0.017 Whr/cm 3 , while the Form 2 results yield 0.03 Whr/cm 3 . This indicates that the PC per unit volume for FFF is almost half of that of SLA's. It should also be noted that the main influencer in the high PC for FFF is maintain build platform temperature, which means that the PC for FFF can be significantly reduced if using a material that does not require a heated bed for adhesion.

Process parameter effects within FFF and SLA
To understand how each process parameter affects response variables, multivariate analysis of variance (MANOVA) is employed using Statistical Analysis Software (SAS). For this analysis, a priori of 0.01 is selected to be later compared to resulting p-values. The data is imported into SAS, and a code is written to apply MANOVA to the data. MANOVA results indicate whether a parameter has an overall influence on the response variables, unlike analysis of variance (ANOVA), which can indicate whether a process parameter has influence on a specific response variable. The reason for choosing MANOVA over ANOVA is due to a correlation study that is done prior. The response variables are highly correlated which leads to using MANOVA instead of ANOVA.
Analyzing MANOVA results is performed by considering p-values and comparing them to the priori. In addition, F-values are also considered to further support a conclusion. A p-value higher than the priori indicates the process parameter has no main effect on the response variables. A p-value lower than the priori indicates the process parameter does have an overall main effect on the response variables. Looking at FFF results in Table 2, all three parameters have overall main effects on the response variables, with LH being the most influential considering the higher F-value. For SLA results, it can also be seen that all three process parameters have overall main effects on the response variables, with LH again being the most influential process parameter. In addition, simple statistics like the average and standard deviation can be found in Table 3 for response variables.

Process parameters vs. response variables
Looking at the plots in Figs. 3, 4, and 5, it can be observed that all response variables fluctuate with the three different layer heights, except mass data for SLA. The first plot displays the relationship between PC and LH, which looks like an inversely proportional relationship; the higher the layer height, the lower the PC and vice versa. This is an expected  Fig. 4, displays the relationship between SR and LH. Logically, it is a proportional relationship. Since higher layer heights produce visible layer lines which can trigger higher SR values during measurement. It should be noted that for SLA, there is no visible difference in SR between LHs of 0.025 mm and 0.05 mm. In addition, no statistically significant differences exist between the levels of LH for SR in FFF, based on Table 4 (discussed later), despite the visible difference observed in the plots.
The machine is capable of a LH as low as 0.01 mm according to the manufacturer. That does not indicate the machine deposits the correct amount of material at lower LHs. On the contrary, LH for SLA-fabricated parts does not exhibit a change in mass with the different LHs. This further supports the fact that SLA is capable of higher detail than FFF is capable of. This conclusion and the visible differences in Fig. 5 are again validated by Table 4, considering the 0.1 LH for FFF statistically differs from its counterparts.
The way ID affected response variables is no surprise, with PC and mass having proportional relationships with ID and no effect from ID on SR. For IP, it is observed that the process parameter has no effect on PC or SR. This outcome is expected because for PC to be affected significantly, the parts must be larger and contain more infill structures. In addition, IP only alters the internal structure of the model, which should not affect the outer surface and its roughness. A conclusion, however, can be made about how IP affects mass for both FFF and SLA. In Fig. 6, it is seen that out of the three IPs, Honeycomb, or Hex, consistently has the lowest mass compared to the other two infill patterns.  This observation is to justify using this IP for the validation study.
To statistically validate the conclusions drawn based on the previous plots, a pairwise contrast analysis is performed for all possible combinations of process parameter levels for each response variable. The Tukey method is used for the pairwise analysis using SAS platform. The priori for this analysis is 0.05. In other words, the p-values shown in Table 4 will be compared with 0.05, thus allowing conclusions made previously to be statistically validated on how different process parameter levels affect each response variable, for both SLA and FFF. The results of the Tukey test mostly verify and clarify the conclusions drawn based on the plots.

Validation study
The purpose of the validation study is to achieve the same or similar results for PC using models of different backgrounds and different complexities. The models are first downloaded from the STL online library website, Thingiverse (www. thing iverse. com). Three models, shown in Fig. 7, are downloaded and edited in ANSYS SpaceClaim to have the process parameters desired. The process parameters used in this study are based on the ones used in the initial parametric study. A 0.1-mm LH is chosen considering that it is the LH the FFF machine is most capable with. For ID, 20% is chosen simply because it is a commonly used ID and considered a middle ground between the other two levels. As previously mentioned, Honeycomb is chosen for the IP since it produced the lightest parts for both technologies. The models chosen are a knee joint, a steering knuckle, and a threaded bolt. These three models vary significantly in features and complexity.
The outcome is expected to be in favor of SLA-fabricated parts when it comes to lower PC. Experimental results verify expected outcome, with FFF using about 8 times more energy to fabricate the sane parts. Even with normalizing the results with respect to build volume, it is still in favor of the SLA process. Normalizing with respect to build volume yields an average Whr/cm 3 of 0.049 and 0.048 for FFF and SLA, respectively. Looking at Fig. 8, however, it seems that the knuckle model yields the highest consumption. Typically, models that are tall consume more energy than shorter models, which means the knuckle should have been the most economic since all three models have the same volume. Taller models have more frequent discontinuities during fabrication for layer changes in the z-axis. This only leaves the reason to be the supports needed for the knuckle model significantly increasing the fabrication time which directly increases PC. Taller models do consume more energy using SLA according to Fig. 8, mainly due to the pause requirement between layers to allow resin to drip back into the reservoir. Considering the findings, SLA proves to be much more economical than FFF from a PC standpoint, which is proven using simple half-inch cubic specimens along with various complex structures. The reason for these results is because SLA, like FFF, is typically less efficient in layer changes in the z-axis. The low efficiency for z-axis movement in SLA is solely due to the nature of the process and the material used.

Complexity evaluation
Reiterating, several methods are proposed to quantify model complexity by Brajlih et al. [29] with the number of facets, or by diving the part volume with number of facets, box volume, and surface area. In addition, Pradel et al. [35] proposed a similar method which is simply the surface area of the model divided by the volume of the envelope space between the functional faces, which is essentially the volume of the models chosen in this study. Considering Pradel's ratio is just the inverse of one of Brajlih's methods, only Brajlih's methods are considered. The three ratios by Brajlih et al. are defined as A low C vf ratio would indicate there is a high number of facets for a given unit volume. This observation yields to the conclusion that a low C vf ratio suggests a more complex part, considering there is a need for more facets per unit volume due to changes in relative planar geometry. The C vbv can suggest conclusions about whether parts are bulky and simple or more free form and complex. The C vsa ratio can be used to quantify the relationship between a part and its curved or free form surfaces. In addition, the ratio can be an indicator for differences between thin-walled and bulky parts.
After scaling all models to have the same part volume for comparative reasons, the properties shown in Table 5 are acquired from slicing software by simply clicking on each model individually. The surface area, however, is acquired from Autodesk Fusion 360 considering the slicing software does not convey that information. The ratios are simply calculated using their corresponding equations.

Discussion
The data in Table 5 indicates that the knee model has the highest number of facets, then comes the bolt and followed by the knuckle. Logically, the knuckle model has many observable flat surfaces which directly correlate to the low facet count. On the contrary, the knee model exhibits purely organic shapes throughout the model apart from the base. With organic shapes, flat surfaces are seldom found which also directly correlates to the high number of facets. From a manufacturing standpoint, organic shapes are random and typically considered complex geometry for traditional manufacturing processes. For this reason, the knee is considered the most complex out of the three, based on the nature of the model and on the number of facets. The C vf ratio is based on the number of facets which, as established, leads to the conclusion that the knee model is the most complex. The trendlines featured in the top of Fig. 9 indicate that PC is not necessarily higher for a significantly large number of facets. In fact, FFF results show an inversely proportional relationship with the number of facets in a model. As an observation, the knee model has justifiable reasoning for the high number of facets due to the random organic structure as opposed to a high-resolution flat-surfaced part. Results for SLA indicate that the lower number of facets has the least PC. However, it is previously mentioned that the low PC for this specific model is most likely due to the lower z height compared to the knee and bolt models. The C vf ratio can be described as the inverse of a normalized # of facets number with respect to part volume. This yields to C vf having a linear relationship with PC for FFF; the lower the C vf ratio, the lower the PC. SLA results show no specific trend or absolute relationship with the number of facets or the C vf ratio. According to C vbv, the bolt model is the least complex. This result is expected as the bolt only has one variation of volume with respect to its bounding box. This one variation is the shank area of the bolt since it is much smaller than the bolt head; the bolt head defines the bounding box dimensions. This ratio, however, would not account for the smaller details in the bolt, like the actual threads or "knurls" on the bolt head, which may lead to false conclusions that the bolt is the least complex. This leaves the knuckle model to be of highest complexity according to the C vbv ratio and the knee to be considered of average complexity. Looking at Fig. 10, however, no conclusions can be made about the relationship between the C vbv ratio and PC. This may be due to the proximity of the part volumes and the bounding boxes.
Looking at C vsa results, the knee model is suggested to be the most complex; this is mainly due to the curves in the model itself. In Brajlih's work, a threaded plug very similar to the bolt had a significantly higher C vsa ratio than the rest of their models. The plug was significantly larger than some of the other models in their study, which may have tremendously affected the comparative results. The trendlines in Fig. 11 indicate an inversely proportional linear relationship exists between the C vsa ratio and PC for FFF. For SLA, however, no solid conclusions can be made considering the non-linear trendline. Based on the three ratios, it can be concluded that the knee model is considered the most complex out of the three. These results agree with intuition since TM processes would not be able to fabricate the model without significant effort. Associating the PC results for both technologies, along with complexity evaluations, is further proof that complexity is indeed free with AM technologies.

Conclusion
This study explored a number of process parameters using the two most used AM technologies, FFF and SLA. The process parameters are LH, ID, and IP. The response variables investigated are PC, SR, and mass. Statistical analysis, A validation study is performed to further investigate the results using more complex models: a knee joint, a steering knuckle, and a threaded bolt. The models are fabricated using a LH of 0.1 mm, an ID of 20%, and the Honeycomb IP. Results indicate that SLA uses 8 times less energy compared to FFF to fabricate the three models. Normalizing the PC data with respect to build volume yields 0.049 Whr/cm 3 and 0.048 Whr/cm 3 for FFF and SLA, respectively. The results of both studies indicate that SLA is far superior when it comes to PC and SR compared to FFF.
Complexity evaluation is performed using four methods, three of which depend on the part volume. Three out of four methods indicate the knee model is the most complex. Looking at PC data, the knee model has the lowest PC for FFF, and not the highest in SLA. This study is further evidence that complexity is indeed free for AM processes.