3D printed object topography analysis and the viability for forensic examination

Abstract

With the recent surge of commercially available 3D printing technologies, untraceable firearm components are being created with ease. As these firearms become more popular, there is a need for viable methods of source identification for these 3D printed objects. Both the printer hot end nozzles and print bed surfaces leave distinct characteristics that are observable upon surfaces of the 3D printed objects. Few studies focus on toolmark analysis of 3D printed evidence, and fewer utilize algorithms to determine the similarity of these toolmarks. The goal of this research was to determine if toolmarks found on 3D printed objects are distinct and reliable enough to be used for source conclusions. To test this, 150 printed objects were created with 10 different print beds and 10 different nozzles, with 15 prints for each bed‐nozzle pair. 3D scans of the top and bottom surfaces of each object were made using the Cadre Forensics TopMatch‐GS system. Cadre's pattern‐matching algorithms were then applied to the 3D scans, giving each comparison a score between 0.0 and 1.0 depending on the similarity of the surfaces. All print bed surface and nozzle surface scans were intercompared, resulting in 11,175 comparisons for each top and bottom surface. Data were analyzed using receiver operating characteristic (ROC) curves and area under curve (AUC) scores. AUC scores demonstrated that the algorithm was able to consistently differentiate between same and different‐source printed objects. This study demonstrates that source identification of 3D printed objects using toolmarks may be viable for forensic examination.

Highlights.

• Objects printed with the Prusa MK4 3D printer were analyzed for toolmark evidence.

• Surfaces of 3D printed objects were scanned and intercompared with pattern‐matching algorithms.

• Algorithm and visual comparisons indicated that surfaces of the 3D printer do not wear significantly over time.

• Algorithm and visual comparisons indicated that toolmarks on 3D printed objects were distinct.

• Toolmarks found on 3D printed objects are viable for forensic examination.

1. INTRODUCTION

The increasing accessibility and sophistication of 3D printing technology has led to a rise in the creation of untraceable firearms and firearm components, commonly referred to as “ghost guns.” These 3D printed objects present a significant challenge to forensic investigations due to the difficulty in linking them to their original source. Traditionally, when a firearm is found on the scene of a crime, an investigator can utilize the serial number embedded in the receiver of the firearm to assist in determining previous owners of the firearm. The Gun Control Act of 1968 mandated that all non‐antique firearms must have a unique identifier on the receiver of the firearm [1]. With this, criminals have been obliterating serial numbers on firearms for decades. However, the recent surge in 3D printing technology has enabled these criminals to simply create firearm components or even entire firearms without any unique identifiers.

3D printing technologies function by depositing thin layers of material onto itself to create a complex three‐dimensional object. One of the most popular and consumer‐friendly 3D printing techniques is known as fused deposition modeling (FDM). In FDM, spooled filaments in the form of thermoplastic polymers are extruded through a heated nozzle. As the filament extrudes through the heated nozzle, it reaches a semi‐liquid state. The plastic is then extruded onto the print bed surface, where the first layer is deposited. From this point, the thermoplasticity of the filament allows it to bind itself to previous layers of material and harden at room temperature [2]. This extrusion assembly is mounted to a machine that moves along the x‐, y‐, and z‐axes and is controlled by a complex computerized system. 3D models made for printing can be downloaded from websites for free or can be custom‐made using 3D modeling programs. Once downloaded, the 3D model is run through a program called a “slicer.” This program “slices” the 3D model into several small layers. The 3D printer translates these layers into machine code pathing that the printer follows during the printing process. Much like standard firearm manufacturing, the theory of characteristics created on a firearm during the manufacturing process extends to the 3D printing of firearms.

Few studies have discussed the presence of toolmarks on 3D printed objects and their viability for forensic examination. One study observed toolmarks present on the bottom layer of the 3D printed object [3]. It was theorized that the plasticity of the 3D printed polymers during the printing of the base layer of an object would inherit individual characteristics present on the print bed surface [3]. With this hypothesis in mind, analysis of the toolmarks was attempted to determine if characteristics present on the bottom layer of the printed object would be a reliable area for forensic examination [3]. Using the characteristics on the bottom of the printed object, source conclusions were made between two printed objects from the same printer as well as a printed object compared with a silicone casting of the print bed surface [3]. These print bed characteristics are promising areas for comparison but have limitations. The main difficulty of these characteristics is that they heavily depend on where the object was printed on the print bed surface [3]. Since the object only inherits markings on the base of the printed object, the area on the print bed affects the markings that would be present. This means that two objects printed in different areas of the same print bed surface would have different characteristics.

In another study, the theory of unique marking during the manufacturing process was applied to the hot end nozzle of an FDM 3D printer [4]. The hot end nozzle of an FDM 3D printer is the last mechanism that the polymer filaments interact with before layers are deposited [4]. Several brass nozzles were examined and microscopic markings incidental to manufacture were observed at the tips of the nozzles [4]. It was hypothesized that these markings contribute to striae and other markings on the surfaces of a 3D printed object [4]. First, areas of interest on the 3D printed object that were most likely to inherit the characteristics of the hot end nozzles were observed. These toolmarks were found on the top layers of the 3D printed object and observed as striated in nature [4]. To determine if the striated characteristics were sufficient in quantity and agreement, several different 3D printed objects were manufactured with the same hot end nozzle. After several visual comparisons between different printed objects, it was deemed that the areas on the top layers of a printed object had sufficient striated characteristics to make source conclusions [4]. However, the researchers also found that the striated characteristics left by the nozzle were dependent on the direction of the nozzle's travel [4]. This meant that a nozzle traveling from left to right would deposit different characteristics than the same nozzle traveling right to left. When two of the same printed objects were compared, this limitation did not pose an issue.

These exploratory studies yielded intriguing insights into the application of traditional firearm and toolmark analysis techniques to 3D printed objects, but they also discussed the limitations of the research as well as the need for further experimentation. One common limitation is the subjective nature of the visual comparisons made during both studies. Similar to traditional toolmark analysis, conclusions made regarding 3D printed objects are subjective in nature and are based on an examiner's personal training and experience. However, emerging technologies are creating ways to quantify the number of discernible characteristics on firearm evidence using pattern‐matching algorithms.

Virtual comparison microscopy (VCM) is an evolving technology that aims to assist in the reliability of conclusions made regarding firearms evidence. Traditionally, firearm examination is conducted using light comparison microscopy (LCM), where microscopic characteristics present on firearms evidence are compared using light sources and microscopes. VCM technology is improving the comparison process by creating both high‐resolution 2D images of firearm evidence and 3D topographical scans of firearm evidence. One such instrument used for collecting 3D scans is the Cadre Forensics TopMatch‐GS 3D [5]. Additionally, Cadre Forensics has developed algorithms that are able to quantify the amount of similarity between pieces of evidence that are both impressed and striated in nature [5]. Several studies have been conducted using the TopMatch‐GS 3D instrument and algorithms [6, 7, 8]. Many aimed to validate the use of VCM and algorithm‐assisted comparisons in making source conclusions regarding firearms evidence [6, 7, 8]. The results of these studies boasted low error rates and an improved evidence handling workflow [6, 7, 8]. Overall, these novel systems and technologies are revolutionizing the way that examiners compare firearms evidence, along with increasing the objectivity of the examiner's conclusions.

Illicit applications of 3D printing have become an increasing threat to law enforcement as the popularity of these technologies increases. Along with this surge in popularity, research in the field of 3D printer forensic identification is also in great demand. Although there has been little research on the topic of 3D printed feature comparison forensics, the goal of this study was to determine the viability of toolmark comparison techniques applied to 3D printed objects. This study utilized 3D VCM technology and employed algorithm feature comparison software to quantitatively determine the similarity of topographical signatures present on the surfaces of 3D printed objects.

2. MATERIALS AND METHODS

2.1. 3D printer

The Prusa MK4 printer was chosen as the instrument to print all the 3D printed objects for this study because of its reliability, ease of use, open‐source nature, and high‐quality prints (Figure 1A). This printer utilized the FDM technique, where thin layers of polymer are deposited onto a print surface and subsequent layers of polymer are added onto previous layers to create a 3D object. The 3D printer was acquired solely for this study, meaning all printing and printing conditions were within a controlled environment.

FIGURE 1.

The Prusa MK4 printer used in this study (A) along with a brass nozzle (B) and textured print bed surface (C). Scale tick marks are in millimeters.

2.2. Filament

The prints for this study were printed using Hatchbox PLA (Polylactic Acid) filament. PLA was chosen due to it being one of the most common and sturdy 3D printer filaments for creating firearm components. The specific branded Hatchbox PLA was chosen for its high quality and performance. The filament was stored in a plastic container along with desiccant to ensure controlled moisture levels when not in use.

2.3. 3D model

The printed object that was used for analysis was a custom file made in Autodesk Fusion 360 specifically for this study [9]. The object resembled a double‐sided 0.45 Auto caliber cartridge case (Figure 2) so that it would easily fit into the cartridge case holders on the TopMatch instrument for 3D scanning. The length is a bit shorter than a 0.45 Auto to reduce the amount of filament needed and the printing time. The object was designed to have a flat surface at the bottom of the print and a simulated primer surface at the top of the print. The flat surface on the bottom was chosen to inherit the characteristics of the print bed surface with as much surface area as possible. The simulated primer at the top of the object was chosen to inherit characteristics of the hot end nozzle surface in focus of the 3D scan. The custom .stl file was then sliced into gcode for printing using PrusaSlicer software [10]. Based on recommendations and test prints, the print bed surface for all prints was 60°C, and the nozzle temperature for the first layer was 230°C with all remaining layers being printed at a nozzle temperature of 210°C [11]. The first layer was purposefully printed at an elevated temperature to ensure appropriate bed adhesion. Hatchbox recommends applying blue painter's tape to the bed surface to promote adhesion, but that would have prevented transfer of the textured surface [11]. Good adhesion was observed with the heated bed and increased first layer nozzle temperature, so those are the settings used across all prints in this study.

FIGURE 2.

Object designed and printed for analysis. Side view (A) along with the top showing nozzle marks (B) and bottom showing print bed marks (C). Scale tick marks are in millimeters.

2.4. 3D printed object scanning and comparison

The instrument chosen to visualize these objects was the Cadre Forensics TopMatch‐GS 3D scanner. The TopMatch‐GS was designed to scan the headstamp and primer areas of cartridge cases. Thus, the design of the object used for the study resembled a 0.45 Auto cartridge case to be compatible with the scanner. The TopMatch‐GS was chosen because of its ability to create high‐quality 3D images of cartridge case surface topography, as well as its ability to use powerful pattern‐matching algorithms to compare cartridge case topographies. The TopMatch‐GS utilized photometric stereo with a painted elastomeric gel sensor to create 3D images of cartridge cases without the influence of surface reflectivity. This was important because 3D printed objects can be printed with many different types and colors of filament, which tend to reflect large amounts of light. This makes traditional comparison microscopy analysis of 3D printed evidence difficult.

The TopMatch system's algorithms are intended to assist examiners in making conclusions by determining the level of similarity between two surfaces. Two of the TopMatch system's algorithms were used for objective comparison in this study: the breech face algorithm and the aperture shear algorithm. The breech face algorithm is utilized for the impressed characteristics on the primer area of firearms evidence. Therefore, the breech face algorithm was used on the surface of the object that contacts the print bed surface. This is because the characteristics on the bottom surface of the print bed resembled impressed characteristics found on the breech face area of firearms evidence. The aperture shear algorithm is utilized to compare striated markings found on the aperture shear section of the primer. Because the aperture shear algorithm is used for the comparison of striated markings, it was used on the top layer of the print bed that contacts the nozzle. This is because the characteristics on the top layer of the printed object that contact the nozzle were striated markings.

Once scans were uploaded to the TopMatch's incident analysis software, they were intercompared with each other using the TopMatch breech face algorithm version 1.2 and aperture shear algorithm version 1.3. The algorithms independently scored each comparison on a scale from 0.0 to 1.0 in terms of topographical similarity, where larger numbers indicated greater similarity. These comparison scores were then used for statistical analysis.

2.5. Print bed surface

The first area of the printer this study focused on is the print bed surface (Figure 1C). The print bed surface is where the initial layer of the 3D printed object contacts the print bed. Because both the print bed surface and the PLA are heated, it is hypothesized that the thermoplastic properties of the PLA will allow it to inherit the characteristics of the print bed surface. The area of the printed object that contacts the print bed was scanned with the TopMatch‐GS to produce a high‐quality 3D image of the object's surface topography (Figure 3). Figure 3 highlights the granular characteristics that the 3D printed object inherits from the textured print surface on which it was printed. The lines present within Figure 3 are the 3D print path lines that the nozzle follows when creating the object layer by layer and are present within all FDM 3D printed objects. Once the bottom surface of the object was scanned by the TopMatch system, the program had the user mask areas of the image that are suitable for algorithm comparison by highlighting the areas in a red color. This allowed the user to control what features the algorithm used to compare to other scans. When masking the scans of the print bed surface, the print path lines were not highlighted for algorithm analysis because they were present on the bottom layer of all FDM 3D prints. While not subclass characteristics in the traditional sense because they are not from a manufacturing process, these print path lines are analogous to subclass information as these could be repeated on another printer by coincidence leading to false identifications. Furthermore, slicing software that determines the nozzle path could generate a different pathway, even on the same printer and object, leading to a false elimination. Figure 3 depicts how the bottom layer of the 3D printed objects were masked, with the areas that the algorithm will compare highlighted in a red color. The algorithm used for analysis of the print bed surface was Cadre's Breech Face 1.2 algorithm. Scores that the algorithm generated were based solely on the characteristics within the red highlighting. The image area in Figure 4 is less than that shown in Figure 3. This is because the TopMatch is designed for fired cartridge cases, and therefore it crops to the center area expecting to focus on only the primer area for algorithm comparison purposes. Since the print beds are double‐sided and the surface area of the printed object is small, the print beds were divided into sides (side A or B) and subdivided into 20 sections (50 mm × 50 mm) lettered A through T on each side for consistent printing. Characteristics of the object inherited by the print bed surface depend on where the object was printed. This means each different section (A through T) of the same print bed surface should impart different characteristics on each object. The print bed surfaces used in this study were Prusa Textured PEI Powder‐Coated Spring Steel Sheet. The textured print bed surface was chosen due to its ease of printed object removal, as well as its durability. Although the textured print bed surfaces did not come with the Prusa MK4 printer, many other similar FDM printers come standard with textured print bed surfaces.

FIGURE 3.

A TopMatch‐GS scan of the bottom of the printed cartridge case that contacts the print bed. Multiple characteristics of the print bed that transferred to the bottom of the printed object can be easily observed.

FIGURE 4.

A focused TopMatch‐GS scan of the bottom of the printed cartridge case that is appropriately masked for algorithm scoring analysis.

2.6. Hot end nozzle

The second area of the printer this study focused on is the printer's hot end nozzle (Figure 1B). The hot end nozzle is where heated filament is extruded from. The filament maintains constant contact with the nozzle as the object is being printed. During the manufacturing process, nozzles inherit characteristics incidental to their manufacturing process. Figure 5 highlights the microscopic differences in the tips of unused hot end nozzles. It is hypothesized that the characteristics of the nozzle are transferred to the printed object as the nozzle moves to create the printed object. These markings are most commonly observed as striated marks on the top layer of the 3D printed object. The top layer of this study's printed object was designed to highlight these striated marks present in the 3D scanned image (Figure 6). Once the top of the object was scanned by the TopMatch system, the program had the user mask areas of the image that are suitable for algorithm comparison by highlighting the areas in a green color. Once again, this allowed the user to control what features the algorithm uses to compare to other scans. The algorithm used for analysis of the nozzle surface was Cadre's Aperture Shear 1.3 algorithm. Figure 6 depicts only a small, masked area of the printed object. This section of the object was chosen to be masked because it is a portion of the object that shows a nonoverlapping area of nozzle striations. These sections of the print that depict nonoverlapping areas of striated marks are hypothesized to be the best areas to use for comparison [3]. This is because areas that do not overlap inherit the characteristics of the entire nozzle surface and are not destroyed by the normal print path. Since all these objects were printed from the same 3D modeled file, all prints would have a similar area of nozzle striation suitable for comparison purposes. All hot end nozzles used in this study were brass 0.4 mm Prusa Nextruder nozzles. All hot end nozzles were new and bought directly from Prusa.

FIGURE 5.

An image of the microscopic differences found between two unused hot end nozzles.

FIGURE 6.

A TopMatch scan depicting the striations left by the nozzle of the 3D printer on the top layer of the object. The area in green is a selected “mask” of an area for the algorithm to use for comparison purposes.

2.7. Nozzle and print bed wear

One of the most fundamental questions to consider when evaluating how viable a piece of evidence is for forensic toolmark examination is how its source changes over time with consistent use of the tool. In terms of 3D printed evidence, the tools and surfaces from which the object inherits its characteristics must be evaluated to see if their surfaces change significantly over time. To test this, three unused Prusa Nextruder nozzles were subjected to printing the same object using the same gcode 100 times in a row to determine if the nozzle surfaces changed significantly over time. These prints were also printed on three different print surfaces to determine if the print bed surface changed significantly over time as well. This resulted in 300 prints: 100 prints from each nozzle, with a different print surface corresponding to each nozzle. Every 10th print, from the 1st to the 100th had the nozzle side and the print bed side scanned on the TopMatch system to visualize the surfaces. This resulted in 11 scans for each of the three nozzle surfaces and 11 scans for each of the three print surfaces.

Because the goal of this project was to evaluate marks on printed objects, to assess changes over time, the printed objects were examined instead of directly looking at the nozzles and print beds. This informed the degree to which transferred marks may be altered due to changes occurring to the nozzle and print bed surfaces. Once the 3D scans of each of the print surfaces were complete, the scans were masked for algorithm analysis based on the surface type (nozzle or print bed). Once the scans were masked, the algorithm intercompared the scans for all the nozzles and print beds coming from the same source. This provided a comparison score from 0.0 to 1.0 quantifying the surface topographies. This was done to assess if there was a declining trend in comparison scores of prints coming from the same source over time, which would indicate the surface used to manufacture the object was wearing significantly.

2.8. Print bed surface and nozzle variability

This part of the study aimed to determine if the toolmarks present on the bottom and top surfaces of a 3D printed object are distinct enough to be used in forensic analysis. Because the printed objects themselves are more commonly submitted to the forensic laboratory, rather than the printer itself, variability assessment was approached by examining the printed surfaces. In this part of the study, 10 different hot end nozzles and 10 different print bed surfaces were used. Because the print bed surfaces were double‐sided, five textured print sheets were used, and objects were printed on section M of each side of each sheet. Fifteen prints were conducted per nozzle and per print bed surface. The first 10 prints for each surface were printed consecutively with no adjustment, but for the last five prints, the nozzle and print bed surfaces were removed and then reapplied to the 3D printer. This was done to simulate normal working conditions when printing objects. All prints for this part of the study used the same gcode to print the object.

Since the nature of the model was double‐sided, each top and bottom surface of the 3D printed object was scanned. This created 150 3D scans for the bottom surface that contacted the print bed (15 prints on each of 10 different print bed areas) and 150 scans for the top surface that contacted the hot end nozzle (15 prints from each of 10 different nozzles). This resulted in a total of 300 3D scans with a known origin. Each scan was manually masked and intercompared with all other scans of the same surface (nozzle or print bed) using the TopMatch's 3D comparison algorithms. As mentioned previously, the breech face algorithm was utilized for the scans that contacted the print bed surface, and the aperture shear algorithm was used for the scans that contacted the hot end nozzle. Once all scans were uploaded and the algorithm had intercompared all scans, the data were exported to conduct data analysis.

2.9. Open‐set unknown comparisons

A blind study was then set up to test the ability to identify and eliminate printed objects with an unknown origin. Twenty‐five additional objects were printed by another individual, with the origins of the objects being known only to that individual. These 25 prints had each side scanned, for a total of 50 scanned surfaces: 25 of the print bed surface and 25 of the nozzle surface. These scans were compared with scans of known origin from the nozzle and print bed variability portion of this research in an open‐set design, meaning not all the objects of unknown origin were guaranteed to be identified to a known source. These scans were uploaded to the TopMatch's incident analysis program for visual comparison and algorithm analysis. Once all scans were uploaded and the algorithm had cross‐compared all scans, the algorithm's scores were used as a triage tool to quickly identify potential same source objects. The scans with the highest comparison scores were visually compared, and conclusions were made for each of the 50 scans. Comparisons between the known objects and the unknown objects had their conclusions plotted into a confusion matrix. A confusion matrix is a performance measurement table that visualizes the number of true‐positives, true‐negatives, false‐positives, and false‐negatives [12].

3. RESULTS AND DISCUSSION

3.1. Print bed and nozzle wear

For each of the three nozzles and print beds used to evaluate the level of wear present, scores from printed objects were plotted to determine if they trended downward over time. Because all print settings were kept constant throughout this study, a downward trend in algorithm scores could indicate that the surfaces were wearing significantly over time. To visualize this, the first print for each surface was compared with every 10th print, up to the 100th print. Figure S1 depicts the graphs of the breech face algorithm scores of the three print bed surfaces used in this portion of the study (bed B, side A, sections L, M, and N). Although all objects were printed on the same print bed and same side of the print bed, printing in different sections resulted in different characteristics. As seen in Figure S1, the first print compared with all prints retained a score of around 0.9999 for all three print surfaces. This trend was seen throughout all three of the tested print bed surfaces.

The nozzle surfaces were evaluated using the same methodologies as the print bed surfaces; however, the algorithm used for the nozzle surfaces did not provide the same level of consistent scoring as the algorithm used for the print bed surfaces. It is hypothesized that since the algorithm is traditionally used for very fine and deep striated marks, the shallower striated toolmarks found on the nozzle surface of the prints were not as ideal for the TopMatch aperture shear algorithm comparison. Algorithm scores were still used for objective comparison of all nozzle surfaces in this study, but visual comparisons were also used to support the conclusions of this study. Figure S2 depicts the graphs of the aperture shear algorithm scores for all three nozzles used in this portion of the study. Although the nozzle algorithm scores are generally much lower than the print bed algorithm scores, there is still the ability to observe any trends present. The scores for nozzle 4 seen in Figure S2 do not display evidence of significant changes occurring. Figure 7 also depicts images of visual comparisons from the first print of nozzle 4 compared with the 100th print of nozzle 4. Figure 7 focuses on three different areas of comparison where sufficient agreement can be seen between the two comparisons: the left, top, and bottom sides of the nozzle surface.

FIGURE 7.

Images of a visual comparison between the 1st and 100th prints made by nozzle 4. Three different areas containing sufficient characteristics were viewed: the left (left), top (middle), and right (right) sides of the surface.

Similar trends were seen between the other two nozzles used for the wear portion of the study. Similar to nozzle 4, visual comparisons were conducted with nozzle 5 and nozzle 6. Sufficient agreement was easily found between the 1st and 100th prints. Visual examination of the algorithm score graphs suggests no significant changes to characteristics on printed objects across 100 prints, and visual comparisons were conducted to corroborate this conclusion.

No direct inferences can be made here about tool wear because the surfaces themselves were not examined. However, no significant differences were observed on the printed objects across 100 prints from each surface (print beds and nozzles). The framework presented here would be useful for additional research in this area where the print bed and nozzle surfaces themselves are examined at various print intervals.

3.2. Print bed surface and nozzle variability

The similarity scores that the algorithm provides are a quantifiable measure of the similarity of two surface topographies. This feature is essential for a quantitative analysis of similarity, as opposed to traditional subjective analysis. The print bed surface of the printed objects had 14 same‐source and 135 different‐source comparisons each (recall, 15 objects were printed across 10 different print beds). The print nozzle surfaces of the printed objects also had 14 same‐source and 135 different‐source comparisons each. When intercompared with each other, this resulted in a total of 11,175 comparison scores for print bed surfaces and 11,175 comparison scores for nozzle surfaces. Using data from these comparisons, receiver operating characteristics (ROC) curves were used to evaluate the ability of the TopMatch's algorithms to differentiate between same‐source and different‐source 3D printed surfaces [13]. ROC curves were generated for both print surface comparison data as well as nozzle comparison data. The area under the curve (AUC) metric was used to identify the probability of a randomly selected same‐source comparison having a higher algorithm score than a randomly selected different‐source comparison. Figure 8 depicts the ROC curve and the AUC score for all 11,175 print bed comparisons. The print bed comparisons received an AUC score of 1.00, meaning that the algorithm was correctly able to distinguish between same‐source and different‐source comparisons 100% of the time, using only algorithm scores and no visual comparisons. These data objectively suggest that toolmarks present on the textured print bed surface of a printed object can be used to identify and exclude objects coming from a certain source.

FIGURE 8.

ROC curve and AUC scores generated for all print bed and nozzle comparisons.

A ROC curve and AUC score were also generated for the nozzle comparisons. Figure 8 depicts the ROC curve and AUC score for all the 11,175 nozzle comparisons. The AUC score for all of the nozzle comparisons was 0.86, meaning that the algorithm was correctly able to distinguish between same‐source and different‐source comparisons 86% of the time. It is important to note that this statistic is calculated using only the algorithm's scores of the surface topographies and no visual examination was used.

To further understand if certain print beds and nozzles performed differently from others, ROC curves and AUC scores were calculated for each individual nozzle and print bed tested in this study. Figure S3 depicts all 10 ROC curves and AUC scores calculated for the print bed surface combined onto one plot (beds C through G, sides A and B of each). Since the AUC score of all print surfaces combined was 1.00, each individual print bed surface also has an AUC score of 1.00. This suggests that all print bed surfaces consistently reproduced enough characteristics for the algorithm to discern between same‐source and different‐source comparisons. Figure S4 depicts the ROC curves and AUC scores calculated for all 10 of the nozzle surfaces combined onto one plot. As seen in Figure S4, the AUC scores from different nozzles ranged from a high of 0.95 to a low of 0.74, with only two nozzles (N10 & N12) falling below an AUC of 0.84. Although a wide range of AUC scores is present within the nozzle comparisons, these data suggest that nozzles that may not mark as well still leave characteristics that are distinct enough that the algorithm was able to correctly discern between same‐source and different‐source comparisons at least 74% of the time. Similar to traditional toolmark examination, these data provide insight into how different surfaces the object contacts during printing may have areas that leave characteristics more consistently than others.

To further understand the variability of the AUC scores for the print bed surfaces and nozzle surfaces, algorithm scores between the objects of a known same source were plotted using box plots. This analysis allows for the visualization of variabilities present within 3D printed objects of the same source. Figure S5 depicts the box plot of the same‐source comparison scores for the print bed surfaces. As shown in Figure S5, there is not much variability in the same‐source comparison algorithm scores for the print bed surfaces, with the exception of a few outliers. Although print bed E–B had the most variability in same‐source comparison scores, the variability did not affect the algorithm's ability to distinguish between same‐source and different‐source comparisons. These data depict that the variability between same‐source print bed comparisons is not significant enough to have an effect on the ability to make sound conclusions on toolmarks found on the bottom layer of 3D printed objects.

Figure S6 depicts the box plot of the same‐source comparison scores for the nozzle surfaces. As seen in Figure S6, the variability in algorithm scores for same‐source nozzle comparisons is much greater than that of the print bed comparisons. This was expected, as the nozzle comparisons in this study did not score as well as the print bed comparisons. This box plot also depicts two nozzles that had the highest variability in same‐source comparison scores, nozzle N10 and nozzle N12. Recall that both nozzle N10 and nozzle N12 also had the lowest AUC scores for their individual ROC curves, indicating that the markings on these nozzles may be harder to identify than others. Contrary to the previous two nozzles, nozzle N07 and nozzle N15 had the lowest same source score variability with the two highest AUC scores. This may indicate that these nozzles leave more distinct and identifiable markings on 3D printed evidence.

Overall, the data collected indicate that toolmarks present on textured print bed surfaces and nozzle surfaces are distinct enough to be used for source conclusions. The data collected also indicate that there is minimal variability in markings present on the bottom layer of a printed object that contacted the Prusa textured print bed, where there is more variability in striated markings present on the top layer of a printed object that contacts the Prusa Nextruder nozzle. Although the variability in nozzle surface comparison scores impacted the algorithm's ability to distinguish between same and different sources, the data from this study still suggest that nozzle surface comparison is viable for toolmark comparison analysis.

3.3. Open‐set unknown comparisons

After assessing print bed and nozzle variability, an unknown set of 25 print bed objects and 25 nozzle objects were printed and evaluated for conclusion accuracy in a scenario that would resemble traditional casework. Comparisons between the known objects and the unknown objects had their conclusions plotted into a confusion matrix. The confusion matrix allows for the visualization of the number of true‐positive, true‐negative, false‐negative, and false‐positive examiner conclusions. Figure 9 depicts the overall confusion matrix scores from all 50 unknown comparisons. As seen in Figure 9, there were a total of 19 true‐positive conclusions, 23 true‐negative conclusions, 8 false‐negative conclusions, and 0 false‐positive conclusions. This totals an overall correct conclusion rate of 84%. To gain a deeper understanding of the conclusions made, confusion matrices were created for both print bed surfaces and nozzle surfaces individually.

FIGURE 9.

A confusion matrix plotting all conclusions made in the open‐set study.

Figure S7 depicts the confusion matrix scores for only the print bed surface conclusions. As seen in Figure S7, there were a total of 10 true‐positive conclusions, 15 true‐negative conclusions, 0 false‐negative conclusions, and 0 false‐positive conclusions. This totals an overall correct conclusion rate of 100%. This was expected, as the print bed surfaces mark more consistently and with less variability than that of the nozzle surfaces. Figure S8 depicts the confusion matrix scores for only the nozzle surface conclusions. As seen in Figure S8, there were a total of 9 true‐positive conclusions, 8 true‐negative conclusions, 8 false‐negative conclusions, and 0 false‐positive conclusions. This totals an overall correct conclusion rate of 68%. The visual comparisons of the nozzle surfaces were much more difficult than the print bed surface comparisons, which was expected based on the ROC curve analysis. It is also important to note that unknown prints coming from the same unknown source were able to be correctly correlated 100% of the time, even if their comparison to the known source was incorrect. Figure 10 depicts a visual comparison with an unknown nozzle surface (left) compared with a known nozzle surface (right) of nozzle N10. The unknown nozzle surface in Figure 10 was eliminated from being made by nozzle N10, but the ground truth of the comparison indicated that they came from the same source. Markings on the eight falsely eliminated unknown nozzle surfaces were significantly different from the surface of known origin. There may be several reasons for this significant difference in characteristics between the known and unknown prints.

FIGURE 10.

Images of a visual comparison between an unknown nozzle surface (left side) and known nozzle surface of N10 (right side). Three different areas containing significantly different characteristics were viewed: the left (left), top (middle), and right (right) sides of the surface.

The first explanation of these significantly different characteristics was that some of the printed objects were printed with different gcode. This different gcode allowed for printing multiple objects at the same time to improve the efficiency of the printing process. This change of the printed file may have slightly changed the print path of the nozzle, leaving different characteristics on the printed object. To investigate this, two different objects were printed using both different gcode files to determine if the change in gcode had an effect on the striated characteristics present on the nozzle surfaces. Figure 11 depicts two different printed objects printed with the same nozzle, but with slightly different gcode. As seen in Figure 11, the striated markings present on the nozzle surface of the two objects printed with different gcode meet the threshold for an identification because the microscopic marks were in sufficient agreement. This indicates that the minor change in gcode did not have a significant effect on the toolmarks present on the nozzle surfaces of the objects. The change in gcode also did not affect the print bed surface of the objects because characteristics inherited by the print bed surface are influenced by where the object is printed on the print bed surface as opposed to the nozzle path.

FIGURE 11.

Images of a visual comparison between two nozzle surfaces printed with slightly different gcode. Three different areas containing similar characteristics were viewed: the left (left), top (middle), and right (right) sides of the surface.

The second explanation relates to the mechanical nature of the printer itself. There was a significant amount of time between the printing of the known prints and the unknown prints. During this period, several other prints unrelated to this study were conducted. It is possible that wear and degradation of the mechanical parts of the 3D printer occurred between the time of printing the known and unknown prints. However, this does not explain how some unknown nozzle surfaces were able to be easily identified and some were not able to be identified. The cause of the significant differences in markings on some of the unknown print surfaces is currently unknown and requires further research.

Overall, the conclusions made in this open‐set blind comparison study suggest that examination of toolmarks present on 3D printed objects is still viable for forensic analysis. However, different files used to print the objects may have an effect on toolmarks present on the nozzle surface of the printed objects, and normal wear of the mechanical parts of the 3D printer may also influence markings found on the 3D printed object. Further expansion of this study should include a design where the same file is used to print all 3D printed objects for comparison. Further research is also needed on how much printer mechanical part degradation affects markings found on 3D printed objects.

3.4. Limitations

The data from this study suggest that toolmarks imparted onto 3D printed objects during the 3D printing process can be used to draw source conclusions. However, only one 3D printer, a Prusa MK4, was used to print all 3D printed objects. Even though different print bed surfaces and nozzles were used to represent several printers, many other 3D printer manufacturers may use different techniques to manufacture their print surfaces or hot end nozzles. Differences in manufacturing methods of 3D printer parts may lead to different or even no characteristics at all. Another limitation is that only one type of filament was used. Although PLA is one of the most common 3D printer filaments, different plastics may have different properties. This could change the number of characteristics that are transferred from the printer to the object, thus changing the effectiveness of identifying the source of the print.

Further limitations of this research relate to the algorithms used during this study. Cadre's Breech Face 1.2 algorithm is intended to be used in the analysis of microscopic, impressed characteristics on the primer surface of fired cartridge cases. Cadre's Aperture Shear 1.3 algorithm is intended to be used in the analysis of microscopic, striated markings on the primer surface of fired cartridge cases. Although the algorithm scoring performed fairly well in differentiating between same and different‐source conclusions for both print bed and nozzle surfaces, it is important to note that these algorithms were not designed for the analysis of 3D printed evidence.

Other limitations of this study are based on the design of the study. First, the sample size included only 15 prints per nozzle and print bed surface. The average 3D printer nozzle can typically handle hundreds of prints before it needs to be replaced. Secondly, only 100 small prints were conducted for the wear portion of the study. Printers often handle much larger and more frequent printed objects. The results of the surface wear portion of the study may not be as generalizable to printer surfaces that have withstood creating much larger and much more frequent 3D printing objects.

Another limitation of this study is that only textured print surfaces were used. Although textured print surfaces are popular, smooth print surfaces are also used to manufacture 3D printed objects. Figure 12 shows a 3D scan of the bottom layer of an object printed on a smooth print surface. No significant markings were observed on the objects printed on the smooth print surfaces. The diagonal lines seen in Figure 12 have even spacing and may be observable on all smooth print bed surfaces manufactured by Prusa. However, previous research suggests that markings found on smooth print surfaces after their manufacture are inherited by the printed object [4]. Overall, more research must be conducted before determining if characteristics found on the bottom layer of 3D printed objects printed on smooth print surfaces are distinct enough to be used for source conclusions.

FIGURE 12.

A TopMatch‐GS scan of the bottom of an object that was printed on a smooth print surface.

4. CONCLUSION

Data from this study not only suggest that the tools and surfaces that a 3D printed object inherits its characteristics from do not degrade significantly over time, at least to an extent that would change the characteristics on printed objects, but also suggest toolmarks present on 3D printed objects can be used to identify or exclude if an object came from a certain source. Impressed characteristics found on the bottom layer of the printed object that contacts the print bed surface and striated markings found on the top layer of the printed object that contacts the nozzle surface produced the best characteristics for comparison purposes. Utilizing surface topography similarity algorithms allowed for statistical and objective analysis of 3D printed surface topographies to aid in determining that they were suitable for forensic analysis.

The results of this study are significant for the field of toolmark examination because 3D printing technology is increasingly being used to manufacture firearms and illegal firearm components. Although further research in the field of 3D printed object toolmark analysis is needed, the data from this study strongly suggest the possibility of source identification using toolmarks present on 3D printed objects.

Recall that other 3D printer manufacturers may use different techniques to manufacture their print surfaces or hot end nozzles, which may leave different or no characteristics at all. Continued work in the field of 3D printed object toolmark analysis could analyze objects printed using different models of 3D printers to determine if those objects can also be used for forensic analysis. Future work may also include black box studies of 3D printed object toolmark analysis, designed to evaluate examiner conclusions. This would aid in the implementation of these methods into the field of firearm and toolmark analysis.

CONFLICT OF INTEREST STATEMENT

The authors do not have any conflicts of interest to disclose.

Supporting information

FIGURE S1. A plot depicting the algorithm scores for each of the three print surfaces used to determine surface wear over time. The first print on each print bed was compared with every 10th print, up to the 100th print.

FIGURE S2. A plot depicting the algorithm scores for each of the three hot end nozzles used to determine surface wear over time. The first print using each hot end nozzle was compared with every 10th print, up to the 100th print.

FIGURE S3. A combined plot depicting all ROC curves and AUC scores for each individual print bed surface.

FIGURE S4. A combined plot depicting all ROC curves and AUC scores for each individual nozzle surface.

FIGURE S5. A box plot depicting the variability of algorithm scores of same‐source comparisons within each individual print bed surface.

FIGURE S6. A box plot depicting the variability of algorithm scores of same‐source comparisons within each individual nozzle surface.

FIGURE S7. A confusion matrix plotting the conclusions made for the print bed comparisons.

FIGURE S8. A confusion matrix plotting the conclusions made for the nozzle surface comparisons.

Blair CJ, Law EF, Garrison K. 3D printed object topography analysis and the viability for forensic examination. J Forensic Sci. 2026;71:2028–2039. 10.1111/1556-4029.70364

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.