Trace Evidence, Physics and Pattern Interpretation, and Fire Investigation | Session 1 from the 2026 NIJ Forensic Science Research and Development Webinar Series

Gregory Dutton: Good afternoon, and welcome to the 2026 NIJ Forensic Science R&D Webinar Series. We're excited to have you with us for our opening session today. My name is Greg Dutton, program manager at the National Institute of Justice, and I'll be moderating today's session. NIJ is the research, development, and technology agency of the U.S. Department of Justice. We're very excited to share with you some of the research that our programs support.

Today's webinar, Trace Evidence, Physics and Pattern Interpretation, and Fire Investigation, will showcase NIJ-supported research in this wide range of disciplines. Presentations will highlight innovative analytical techniques, statistical approaches to pattern evidence, and emerging research that enhances the scientific foundations of these disciplines. Before we begin, next slide, please. Before we begin, let's review some important information.  

To adjust your audio settings, click on the down arrow button next to the grayed out microphone icon to adjust as needed. There'll be a few minutes after each presentation for questions. Please submit any questions you have during the talk into the Q&A box as you think of them. I will read the questions to the speaker at the end of the talk. To submit your questions, click the Q&A icon at the top of the screen.  

And please note that we may not be able to address all questions. Closed captioning is available. To access, click the more button with the three dots, select language and speech, and then select show live captions. Today's presenters are first, Tatiana Trejos, associate professor of forensic and investigative sciences at West Virginia University. Second is Shuna Ni, assistant professor of fire protection engineering at the University of Maryland, College Park. Third is Meredith Coon, certified latent print examiner. And finally, Ranjan Maitra, professor of statistics at Iowa State University.

Please note that our speakers may not have time to address all your questions. We appreciate your understanding, and we will do our best to get to as many as possible. And with that, let's hear from our first speaker, Professor Tatiana Trejos. Tatiana, the floor is yours. Tatiana Trejo: Thank you, Greg. Good afternoon. It is a pleasure to be here today to share our experience in translating research into practice with a focus on our research on glass, tape, and gunshot residue. I would like to thank our co-PIs, Dr. Luis Arroyo from West Virginia University and Troy Ernst from Michigan State Police - Grand Rapids Forensic Laboratory. Next slide.

This research was funded by NIJ. Our awards are listed below to West Virginia University, but I have to disclaim that the opinions, findings, and conclusions are those of the authors and do not necessarily reflect those of the Department of Justice. Next.

Today, I'm going to discuss our experiences with research/practitioner/industry collaborations, the motivations for the studies, how we establish interlaboratory studies to assist with the adoption of methodologies, and finally, the impact and future needs. Next. The focus of my presentation is to illustrate how multidisciplinary collaboration identifies relevant needs in the field and practical solutions that create paths to implementation.

Next slide, please. At the national level, we have been very lucky to establish collaborations among industry, academia, and practitioners in more than 20 states. Next.

Next, and then at the national level, at the international level, we have expanded that collaboration, illustrating how the discipline shares common needs. Next slide. A key success in the research and practitioner partnerships is the vision of upper management in forensic laboratories. Advancements in the field really start with leaders who are open to change and to modernize protocols and workflows in their laboratories. Identifying shared goals to advance knowledge and practice starts with effective communication, listening and working together to identify needs, and investing time and effort to assess the impact of those changes into the criminal justice system, having a proactive instead of a reactive mindset. Next.

The motivation behind the studies I'm presenting today is to illustrate trace evidence to its full potential. As we all know, trace evidence can provide information not only at the source level, but most importantly, at the activity level. It can answer when, how, what, and where the crime was committed, which can ultimately help us to better utilize trace evidence in the early investigative stages. So the main goal of these grants was to use trace evidence more effectively. Next.

The first study that I'm discussing today evaluated the use of modern micro XRF systems coupled with silicon drift detectors. And the motivation of this study started with a survey that we did within the OSAC Glass Test Group that identified not only the majority of the laboratories in the United States at that time were using micro XRF technology for glass examinations, but most surprisingly, 94% of the laboratories have updated the detectors to modern silicon drip systems. There was therefore a need to evaluate how this advanced technology could impact our practice, as laboratories very rarely adopt technology so quickly. This research gap requires us to update consensus test methods. Next.

Therefore, we assembled a group of collaborators from Michigan State Police, NIST, Battelle, and West Virginia University to assess the reliability of this modern micro XRF technology and to expand its impact in trace examinations. Next. The reason for that is that prior research in our group had shown that these modern micro XRF systems offer improved sensitivity and precision. But that also meant that we needed to evaluate whether these improved capabilities require updates on how we make sampling decisions, conduct examination, analyze data, and interpret the results.

Next. The main hypothesis was that the SDDs improve detection limits and will lead to improved precision; which requires modified comparison criteria; superior discrimination power, which needs to be considered in data interpretation; faster analysis time, which can improve workflows in the laboratories; and improve capabilities for small samples that can expand service outcomes. Next.

The main goal was to increase capacity and modernize trace evidence examinations by assessing the micro XRF performance for glass and tape; providing recommendations for sampling, analysis, and interpretation criteria; evaluating expanding capability to modernize products such as phones, eco glass, electrical tapes; and whether we can do that for smaller samples and faster. And finally, enhancing objectivity in spectral comparisons. The main products led to increased shared knowledge and strengthened foundations through a physical database of thousands of glass and tape fragments that were duplicated at MSP and WVU and available to be shared with other forensic laboratories, a digital repository of over 7,000 spectral files, and publications and presentations that led to the resource optimization and updates to consensus-based methods.

Next. One of our goals was to create a representative database of modern glass and tape that can be used for improved evidence interpretation and exploring mechanisms for sharing that data across laboratories. Next.

The database included contemporary windshield, side vehicle, and architectural glass, as well as PEDs for phone glass and electrical tapes. Each set of samples included samples of known origin from both the same sources and different sources to assess error rates and recommend new sample and comparison criteria. Next.

These recommendations were fully assessed through blind interlaboratory exercises from 11 participants. Next. Which led to the presentation at 22 meetings, 2 workshops, 12 publications, and updates to the ASTM 2926. Next.

So from these studies, I want to highlight the main findings on some of them. So if you can move to the next slide. Our publication in the Journal of Forensic Science presents quantitative similarity metrics for XRF spectral comparison that we named SCAR, which first complements the traditional spectral overlay comparison with more objective statistical methods to support the examiner's opinions.

Second, it offers a practical workflow that normalizes the data, making it now feasible to compare measurements across instruments and addressing a common challenge in sharing XRF spectra among users. The performance of this similarity metric was tested in the inter-laboratory studies. Next.

The study published at Spectrochimica Acta Part B resulted in concrete evidence-based recommendations for forensic practitioners using these SDD-equipped instruments, which include the need for modified comparison criteria for elemental ratios to adapt to the improved precision, use of a larger number of fragments, for adequate characterization of the known comparison samples, and implementation of new strategies for handling modern glass coatings to minimize false exclusions. The study shows the elemental composition of a variety of modern soda lime glasses that includes vehicle and architectural glass with temper, E-coding, and team treatment, providing updated information for practitioners. Next.

Two studies published at the Journal of Forensic Science and Forensic Chemistry demonstrate for the first time the capability of modern micro XRF systems to analyze samples as thin as 10 microns, expanding the suitability of services to cases where only minute fragments remain, otherwise not possible with older systems. Next. Two other studies offer the central foundation and specific analytical and interpretative protocols for the forensic analysis of aluminosilicate glass from PEDs. The proprietary formulations of these phones create opportunities to expand the capacity to investigative leads in cases where broken phone screen has been transferred during the criminal event. Next.

Finally, the studies show opportunities to modernize and improve workflow for incorporating these systems into electrical tape analysis and drastically reduce response times while enhancing the information. The experimental design examines samples handling and the interpretation of chemical data from tapes. when latent fingerprint development is also required. Next.

The ultimate impact on this grant was through the interlaboratory studies, through a full assessment of standardized examination and comparison criteria that led to consensus methods. Next, please. So our research in glass and tape formulations support the development and reliability of trace evidence examination. It provided the most extensive existing repository of micro XRF data to enhance interpretation and implement cost-effective strategies as well as to adapt to contemporary formulations.

The consensus-based best standards help the field standardize examination and interpretation techniques, promoting wider adoption in operational forensic laboratories, and re-evaluated more economical approaches to use trace evidence in early investigative processes. Next. Our second study aims to modernize gunshot residue examinations and interpretations by first recognizing that GSR is more than primer residue. There is a large amount of information coming from the organic propellant that is not being utilized but can increase confidence in the examination. And second, providing streamlined processes and new information on the position, transfer, and persistence of both pGSR and organic gunshot residue. Next.

This research includes various phases, including: the development of fast and accurate screening methods compatible with universal collection tools, demonstrating their versatility for laboratory-based and on-site portable applications across multiple substrates and applications, making it more cost-effective; the development of confirmatory methods for use of organic gunshot residue in workflows to make sequential analysis of pGSR and OGSR practical; the development of novel standard materials for pGSR and OGSR; novel methods to visualize the GSR deposition, transfer, and persistence; and finally, incorporating population studies for more comprehensive interpretation of the evidence. Due to time constraints, I will focus on points two and six in this presentation. Next.

Again, the development of large population databases was essential to test the performance of the methods, including error rates. Our GSR database, including over 4,000 samples that represented collections of hands from shooters collected after firing a gun at different time intervals and after various pulse shooting activities, as well as non-shooters with low and high risk of containing GSR-like particles and mock cases. Next.

For screening, we propose the use of LIBS and electrochemistry due to the advantages. They are rapid, and when I say rapid is seconds to minutes. Selective, sensitive, require minimal to no sample operation. They are compatible with confirmatory methods and are portable. The combination of LIBS and electrochemistry respond to the main challenges in shooting reconstructions, including speed of analysis for rapid investigative needs, on-site detection, versatile solution, and cost-effectiveness to aid in workflows and backlogs.

The main findings show that the use of screening tools is possible in GSR investigations, something that is conventionally used in many disciplines, such as drugs and explosive, but not in firearm-related investigations. These screening tools can modernize the way that we do these investigations and provide accuracies over 90% even in stress-tested populations. The research also reveals that these screening tools are compatible with confirmatory testing on the same sample by SEM EDS, and LC-MS. And the analytical tools are portable as well, so they can provide real-time information in minutes, both at the scene and at the laboratory to assist with triage.

Next slide. The versatility of LIBS expands from lab-based to mobile units, from bullet hole identification to shooting distance estimations, and finally to small particles recovered from hands. Next. One of the breakthroughs in our investigation was the ability to incorporate cameras into the system in analyzing micron-sized single particle analysis. Next. Another focus of our research was the incorporation of organic gunshot residue to complement pGSR analysis. To do that, the main questions were what the best methods are for collecting and detecting organic gunshot residue, whether those methods were compatible with existing pGSR SEM analysis, what the occurrence of GSR was in the regular populations, and how we could use all that information to interpret organic gunshot residue.

The main research focus on evaluating the performance of collection and analytical methods, the integrity of dual testing of GSR on the same sample, the stability of GSR, background occurrence studies, and the deposition, transfer, and persistence of these residues. Next. And perhaps one of the most important discoveries was made using a multi-sensor approach that evaluated the position, transfer, and persistence using laser scattering, high-speed videography, particle counting sensors, and analytical measurements using SEM-EDS, and LC-MS.

Next. The studies were conducted indoors, outdoors, and in vehicles. Next. With laser sheet scattering, now we can make visible the dynamic and chaotic distribution of GSR, as you can see here in these videos. And the particle sensors can monitor how long they remain in the air near the shooting point. As we can see here, GSR particles took up to three hours to settle in indoor environments while they disperse in less than one minute outdoors, even without noticeable wind conditions, something that is crucial to consider in investigations.

Next. For that reason, the risk of indirect exposure by a bystander or a person entering the room after a firearm has been discharged was assessed under various conditions. So if you don't mind moving there, thank you. Here we can see that in enclosed conditions, the GSR cloud reaches the person that was about 3 to 5 feet away from the shooter in a few minutes.

Next. And in this graph, I'm presenting the SEM-EDS results for trials from top to bottom in this image repeated indoors, outdoors, and in vehicles, and from left to right, samples collected from the shooter's hands, bystanders, and passersby. The results indicate that although the number of characteristic pGSR is higher in the shooter, particles are also likely to be detected in bystander and passerby.

Interestingly, when we look at the image on your right-hand side, when analyzing organic gunshot residue by LC-MS, OGSR was consistently detected in shooters, but not in any of the bystanders and passersby hands. This presents a valuable opportunity to use OGSR and pGSR together to interpret GSR evidence, particularly when indirect exposure is questioned in the courtroom. Next. Several studies have confirmed these findings in different populations and mock cases scenarios. Next. So ultimately, these awards have expanded the knowledge base and provided an opportunity to modernize the use of GSR in firearm-related investigations using cost-effective tools. At this time, validated collection methods are now compatible for dual screening of pGSR and OGSR on the same sample.

The analytical performance of these methods has been established, and databases and interlaboratory studies have provided the foundations for background occurrence, transfer, and persistence that can be used for more comprehensive interpretation. Next. And we can move to the next slide.

Thank you. So in conclusion, the most valuable aspect of these studies are not necessarily the end product, but the process that leads to that improved technology decision making and helping the community to transition from research to adoption. The added value of these studies is five fold, it helps with harmonization, collective improvement, development of systematic approaches, organizational culture.

There is no doubt that those inter-laboratory discussions in these exercises evolve the discipline and help with cultural change. And finally, collaboration and communication. And last but not least, if you can move to the next slide, I could not fit all the acknowledgements in one single slide, so I would like to thank all the collaborators who have made the advancement of these fields possible through these partnerships.

Here's the list of our major collaborators in GSR. Next. My co-PI in the GSR project, Luis Arroyo, and all former and current students that have made this possible. And finally, we move to the next slide.

All the collaborators, students, and partners that made possible the research on glass and tape. And with that, I would like to thank NIJ for funding these studies and for the opportunity to share our experiences with you. Thank you.

Gregory Dutton: All right, thank you, Tatiana. We will move into the Q&A session. So a reminder to everyone to submit your questions in the Q&A box. And we will try to queue them up. I have an initial question from someone asking, so essentially, is it easier to collect gunshot residue from a shot fired indoors compared to outdoors? And How does rain affect residue outdoors?  

Tatiana Trejo: Thank you. That's a great question. So interestingly, we can detect GSR indoors or outdoors. In both situations, the success rates have been very high. The main differences is if we are considering indirect the position, like for example, whether the person fired the gun or was in the vicinity as a passerby, someone that entered the scene or someone that was close by, the indoor versus outdoor scenarios produce a very different indirect transfer of GSR as opposed to the direct transfer to the shooter, especially when we consider organic gunshot residue in addition to pGSR because how differently they deposit in those circumstances. So the indoor and outdoor raise a flag of the potential indirect transfer when we have indoor environments with very low flow, there is a potential risk that we need to consider depending on the hypothesis that are evaluated or the questions that are being raised during the investigation.

But in terms of detection from the shooter, both indoors and outdoor provide a high success rate on those both pGSR and OGSR. But of course, the amounts of deposition and transfer will depend on those conditions. We haven't done studies on the rain and how that could affect. We did a very small portion so that we cannot produce a lot of generalizations of running, how running and other activities can help with the decrease and losing particles, both above GSR and inorganic gunshot residue. In a few of the trials that we were doing, running, we had a little bit of rain. And the GSR still persisted, but we didn't have a very heavy rain. So we obtained some results by accident because we couldn't control the rain environment during those trials that happened, even though the weather said that it was going to be sunny.

And we still were able to detect both type of residues. But we haven't done a systematic study to evaluate how rain can affect those circumstances. Where the studies, however, have been very important is to raise the evaluation of factors that might be important to consider during the interpretation of evidence, especially when the scientist is presented with different hypotheses and different aspects of how the GSR got there needs to be accounted.

Gregory Dutton: Thanks, Tatiana. Someone is asking a point of clarification. What is pGSR?  

Tatiana Trejo: Yes, absolutely. So pGSR comes from primer GSR. So when we have an ammunition, there are different sources of residues that are released during the discharge of the firearm. When we deal with residues that come from the primer, that's the primary explosive that is going to be activated or detonated when the pin of the firearm hits the cartridge.

That primary explosive produces residues that are very characteristic of GSR that contains usually lead, barium, and antimony that are materials that are used in the primer. In general terms, usually the discipline referred to that as inorganic gunshot residue, but inorganic gunshot residue is a little bit broader. It includes those primer GSR in addition to other inorganic components that can come from the cartridge or the bullet as well.  

Gregory Dutton: All right, excellent. Another question, are the laser study videos available online, or can they be shared with labs to use for training staff?  

Tatiana Trejo: Yes, absolutely. So we published – one of the publications has the videos in the supplemental section. And we, as part of the NIJ funding, we were able to publish that open source. So finding that, you can e-mail me, that you can access that publication. And in the supplemental, we have all the videos that are free to access.  

Gregory Dutton: All right, excellent. Another question, have you had any troubleshooting issues with using the mobile or fieldable LIBS instrument? Could you expand on that? If so, are there any limitations you experienced? Did you use multiple instruments?

Tatiana Trejo:  Yeah. Yes, excellent question. No, we did use only one brand or manufacturer for both our bench top and our mobile LIBS. We collaborated with one manufacturer that fine tunes and customizes both the ablation chambers as well as the camera so that we can do single particle analysis.

And the instrument is very robust. It is not horrible in the sense that it can, it doesn't fit in a case like other commercial products. Ours is more like a mobile LIBS. So it fits into a really small cart, similar to the ones that are used in emergency rooms that can be moved to one place to another.

So it's meant to be fit in a van, for example, and be brought to the crime scene. We have done a lot of more cases, because we have a crime scene complex here, where we have to transport the instrument and use it. And the instrument is very, very robust. We didn't really have much troubleshooting, and we have been using it for about five years now. And so we have had very good success with the instrument, very little maintenance and almost no troubleshooting so far, knock on wood. But we only have worked with one particular manufacturer.

Gregory Dutton:  All right, excellent. Another question, what confirmatory technique was considered as the primary source for the results?  

Tatiana Trejo: So for confirmation of the inorganic or pGSR, we use the standard protocol, SEM-EDS, and that is the consensus-based method used in the forensic field. And for organic gunshot residue, I didn't have the time to talk a little bit about these, but we primarily use LC-MS, which is the method of choice. But we also have developed and compared other methods that are more accessible in forensic laboratories.

So we did a very substantial study to evaluate GC-MS and coupling GC-MS with a very cost-effective micro ECD detector. And those ones, when the GC-MS is coupled with micro ECD, can provide very comparable results to LC-MS. So the primary confirmatory methods for the organic gunshot residue are LC-MS and GC-MS.  

Gregory Dutton:  All right, excellent. Thank you, Tatiana. We are going to move along. So thank you again, Tatiana. And we will move next to our second speaker, Professor Shuna Ni. Shuna, it's all yours.

Shuna Ni: Yeah, thank you so much. Good afternoon, everyone. And I'm Shuna Ni from the University of Maryland. So fire damage will usually contain evidence, so it's a heating history. Today, I will discuss how controlled experiments, damage measurement, and machine learning can help us interpret that evidence more consistently. Next slide, please. Fire pattern on drywall or gypsum border surface in a building, including both visible surface damage changing and measurable subsurface damage, together they provide important physical evidence, although they may be obscured or difficult to interpret.

So the photo at the left shows a visible surface pattern, including sooting, charring, and discoloring, and their appearance depends on the fire, ventilation, and architecture finish, and some other factors as well. Beneath the surface, heating removes water from gypsum to do calcination and absorbing energy. In the end of NFPA 921, depth gauge is suggested to measure the calcination depth, while this is the depth of the <unintelligible> under the heating.

So this sub-surface recorder complements the visible patterns. Next slide, please. So three gaps motivated our workers. So drywall in a building, usually have a paint or other architectural finish, while most studies have rarely isolated finish as a controlled variable. And probing using depth gauge also needs appropriate pressure for each board type. Excessive pressure can penetrate intact material, while insufficient pressure can underestimate the damage.

NFPA 921's pressure range has no clear scientific basis. Lower pressure for lightweight board is wise but unquantified. Therefore, the depth gauge measurement is operator dependent and also not repeatable in one point. So finally, a visible interpretation of our pattern relies heavily on investigator judgment.

Quantitative methods could improve the consistency. So we try to adjust this gap through a fire test of architecture finish, calibrated probing, ultrasonic measurement, and the computational pattern analysis. Next slide, please. So, the first is how do architectural finishes impact the fire patterns? Next slide, please.

To isolate the effect of finish, we use a 1.8 meter enclosure lined with ultralight drywall and using a consistent foam fuel package. So heat flux gauge and thermocouples are used to characterize the heat exposure and also we're using a balance to check the mass loss. And in each test, the rear wall of this cube was covered by different architectural finish. And this consistent heating exposure condition and the instrumental plan can give us a comparison between different architectural finishes under similar experimental conditions.

Next slide, please. With this set up, we conducted 20 enclosure fire for reference test of drywall with level 2 surface and 16 tests with paper or wallpaper finish on level 4 or level 5 drywall surface. After each fire, we photograph the rear wall and the consistent lighting condition and measure the calcination depths at 36 locations. And this connects a visible fire pattern with subsurface damage. Next slide, please.

So, in the reference test, we found the exposure wallpaper burned without a significant surface flame spread. While on the painted walls, the paint ignited and the flame spread along the surface. As the paint burned away, the paper underneath became exposed and charred, and the paint over level four drywall produces similar overall geometry.

The sheen level mainly affects the local damage feature. For example, the gloss paint shows peeling along the edge of the severe burn areas. Next slide, please. Changing the color also affects local damage. We also tested the green eggshell paint in addition to the white one. And the green eggshell paint showed more extensively localized damage than the white eggshell paint.

And the difference of the formulation in these two different paints offers a possible explanation. However, the overall pattern geometry, even using colored eggshell paint, does not change it too much. Next slide, please. And the wallpaper, wall preparation had a very clear effect on the geometry. So the gloss paint, we have prepared over level 4 and also over level 5 drywalls.

And compared to the level 4 drywall surface, and the level 5 drywall surface have additional skim coat or joint compound over the whole drywall surface. While the test on gloss paint over level 5 drywall produced a more confined burning area than over level 4. The extra skim coat or joint compound shields the paper and absorbs heat, which provides explanation for the more localized damage.

Next slide, please. Moving from the paint to the wallpaper, we saw more substantial pattern change. For wallpaper, we observe the faster surface spread and a taller visible pattern, while for wallpaper with a primer, we also observe peeling and the attachment of the wallpaper produced highly variable pattern widths. And the extensive soot on the upper wall could coexist with limited underlying drywall damage.

A larger visible pattern in the boundary did not necessarily indicate equally extensive gypsum degradations. Next slide, please. So, to quantify the visible boundary, we convert our image by passing the image into gray scale and normalize them across the entire data set. We then use a threshold of 0.4 to trace the boundary or darkened region. While for the calcination, we map the probe depths and define the effected region relative to the indentation measured on unheated boards.

Next slide, please. So, here are the result dimensions, red for the visible pattern while gray for calcination. For a painted wall, the visible pattern generally corresponds closely with the calcination pattern. For wallpaper, with or without a primer, visible patterns extend above the calcinated region with limited underlying drywall damage in the upper region.

While for the wallpaper with a primer, widths do vary a lot between different tests. In one test, the wallpaper burned outside its main plume, widening the visible pattern without a comparable increase in the calcination width. So visible pattern size needed to be interpreted together with the finish and the underlying damage.

Next slide, please. So, here are some key findings. So, across this comparison, we found that paint sheen and color was mainly a local damage feature, while surface preparation and wallpaper affected the geometry of the fire pattern, while calcination adds information beneath those surface effects. So, the next question is how to measure the calcination depths reliably, which will come to our second objective of our research.

Next slide, please. So our second study is selecting the appropriate pressure for depth gauge measurement. Next slide, please. So to select a suitable probe pressure, we test the ultralight and the type X drywall from USG and CertainTeed against the two criteria. A suitable pressure must fully penetrate your full calcinated material while causing very little penetration into unheated material. Too little pressure can underestimate the damage depths. Too much can push into the intact gypsum and overestimate the calcination depths. Next slide, please. To turn these criteria into a pressure value, we fully calcinate a specimen at 400 degrees Celsius for 90 minutes, verifying dehydration with the TGA tester. Ten loading tests per type established 10 candidate pressures for 2 millimeter probes, and we then adjust this load to satisfy both criteria. Full penetration of calcium material and the minimum penetration of unheated gypsum.

So ultralight will require approximately 4.3 megaPascal for both manufacturers. Type X required 10.9 for USG and 7.5 for CT. This difference demonstrates why pressure must be selected for the product being measured. Next slide, please. So, with this pressure established, we follow the four steps shown here. First, we prepare the specimen and we get some gypsum cones.

And second, we heat them under 25, 50, and 75 kilowatts per square meter. And also we include both backing conditions. An open back condition approximates a hollow interior wall cavity, while insulated back represents contact with cavity insulation. So we compare these two different conditions, which can help us identify the effect of backing condition on calcination depth. After the heating, we remove the paper with brief propane torch exposure and send away the surface char and ash.

And in the last step, we measure the depth using the selected proper pressure using a loading machine. And the following plot compares the heat blast and the duration as accumulative heat exposure in megajoules per square meter. Next slide, please. So, this measurement reveals a simple relationship: calcination depth follows a linear increase and then plateaus when the boardis fully calcinated.

We see the test ranges follow the same trend of a given material and the backing condition. Also, rising branch depths can indicate accumulative exposure under matching conditions. Once the probe reaches its full thickness, additional heating cannot increase its reading. So depth no longer distinguishes higher exposures. While, for alternate insulation increases slope by about 43% and shifts the full calcination from roughly 22 to 15 megajoules per square meter. While type X remains nearly 31 under either backing conditions. Next slide, please.

So the back surface temperature can help us to explain the difference. Dehydration basically absorbs energy and slows down the temperature rise. Insulation retains the heat that could otherwise escape from the rear, back surface, accelerating the calcination in thin lightweight board ultralight. While type X is thicker and denser, so more energy is needed to heat and dehydrate it through its thickness.

Therefore, its back surface remains relatively cool for a longer duration, so insulation has a small effect. And the later temperature overshoot shown in these two plots comes from the back paper ignition after the calcination depth has already fully saturated. That's why it does not impact our calibration. Next slide, please.

When we develop this calibration and we're using a loading machine, we also want to check whether the handheld probing follows the same relationship between the calcination depth and heat exposure using a digital force gauge on the USG specimen. So what the results show is that the ultralight results agree closely with the calibration, while type X reaches the same plateau with a slightly steeper slope. Next slide, please. This test shows appropriate pressure improves the measurement consistency, and the calibration relates the depth to exposure within the testing condition. The important limitation is saturation. Once the board is fully calcinated, further heating cannot increase the measured depth. The depths alone, then, cannot distinguish high exposure. This motivated a complementary measurement of mature conditions. That is our objective three of our research. Next slide, please. So, to explain the complementary pressure measurement, our third study using the ultrasonic pulse velocity, UPV, and those results are preliminary and the further analysis is needed. Next slide, please. So, UPV generally measures the travel time of sound pulse across the board.

Heating altersthe gypsum structure changing how quickly the pulse travels. And we use the change as an indicator of the thermal damage of the material. In our setup, we're using a 54 kilohertz transducer under controlled contact pressure to keep acoustic contact consistent. Heating backing condition and the paper removal follows the depth test procedure, which is our second study.

So this measurement corresponds to a response along the whole body thickness, rather than just as a layer of the softer material, which is different from the depth gauge measurement. Next slide, please. So using this procedure, we found that the UPV decreased with exposure across all four product groups, most steeply at low to moderate exposure.

Both manufacturers show similar decay changes. At 25 kilowatts per square meter, velocity levels off near the dehydrated reference. At a higher exposure velocity, it continues below the reference. One possible explanation is that we integrate the incident heat flux, while degradation depends on the net absorbed energy. Surface temperature affects the heat loss, so equal incident heat exposure can deliver different absorbed energies. Heating rate, in addition to the <unintelligible> may also affect the microcrackings, so further experiments are needed to test these explanations. Next slide, please. So, this slide also shows the effect of backing condition on the UPV measurement. So, the insulation reduces the rear heat loss. So, compared to open backing, insulated specimens generally show lower UPV, while type X shows smaller differences.

This is consistent with the material dependent backing condition observed in the probe measurement. At high exposure, backing paper ignition can also add heat as well, too. So we are investigating its contribution to the UPV response. Next slide, please. So the UPV responds to change along the sound path through the board thickness, providing information that complements the probe depths.

We can repeat the measurement at the same location, and as the velocity changes most rapidly to the low and to moderate heat exposure. And our future work is to normalize and develop the normalization and calibration model. And we also would like to test this technique in the compact level fire test, and also to extend this research to high heat exposure. Next slide, please. So far, we have examined the surface pattern and the material change. Our final study tried to connect the pattern classification with the surrounding three-dimensional scene. And in this part of the work, we developed a framework for automated and quantitative image-based fire pattern analysis. Next slide, please.  

This framework has four steps. First, an investigator guides extraction of a fire pattern image. Second, numerical shape features supports a classification. Third the room scene is segmented to wall and objects. The final step is the classified patterns are placed in a three-dimensional scene, and a scene graph records its relationship to furniture and opening. Next slide, please.

So, the first step is extracting a pattern. So, automatic segmentation is usually a struggle with overlapping burning and the damaged object in a scene image. Here, our target has a few more prompts to help the Segment Anything Model, which is a segmentation algorithm isolated as an intended pattern. And then the extractive contour of a pattern can support quantitative analysis.

Next slide, please. So, once a pattern is extracted the contour becomes a numerical shape description. We load a chord through the chord centroid and divide its length by the longest chord. Repeating this is to 360 degrees creates a size independent shape vector. And on the right, the real pattern becomes a segmented contour, and then an aspect ratio signature. So, classify using this vector together with peak and valley information for classification. Next slide, please. We use a random forest classifier with this shape feature. Due to the limited available real fire pattern image, we create a synthetic image data set. The synthetic data set contains 70,000 images and 70% for training and 30% for testing.

The accuracy on the synthetic test set was 93 percent. And then we also test the classifier on limited real fire pattern images, and the accuracy is reaching 83 percent. Next slide, please. To add spatial context to the extracted fire pattern, we're using the algorithm Mask 3D segmenter to segment the scanned scene into wall, floor and individual objects, and around each pattern to associate with the surrounding surface and the nearby furniture or openings. Next slide, please. With the surface and object identified, we projected the classifier pattern into the scene and record its spatial relationship in a scene graph. And this organizes evidence spatially, which can help evaluate a possible fire origin hypothesis.

And the first work is needed to validate how this spatial relationship in a scene graph can help determine the fire origin. Next slide, please. So, in this work, we use the prompt extraction, shape classification, and three-dimensional mapping, creating a more consistent description of fire pattern evidence.

And we still need a larger real fire pattern data set. And also, we should rely less on the manual prompts. So in our future work, we will integrate the calcination and the UPV map into the scene graph as well and develop a reasoning tool to examine what patterns form and how this scene graph can inform the origin hypothesis.

Next slide, please. So, this slide basically summarizes how each study we have done supports fire investigation. First, fire patterning should be interpreted in the context of architectural finishing. More extensive surface damage does not necessarily mean greater heat exposure or more extensive subsurface damage. Secondly, suitable probe pressure and the depth exposure relation is needed to help us to assess cumulative heat exposure better, so we can incorporate it to the pattern more accurately.

And the third is that UPV provides a repeatable measurement that reflects material change along the sound paths through the board’s whole thickness. It can make access high exposure after probing depth has reached its limit. Also further calibration and validation are needed. Finally, is that the automated classification and scene graph can help us describe a pattern consistently and relate them to nearby furniture and openings and supporting the evaluation of fire origin hypothesis.

Next slide, please. So, these are the references we used, and next slide, please. So before closing, I would thank my students and collaborators and the National Institute of Justice for supporting this research. And the findings and conclusions are those of the authors and do not necessarily represent the Department of Justice.

Thank you. Thank you so much for your attention.

Gregory Dutton: All right, thank you, Shuna. A reminder to everyone to enter your questions into the Q&A. We have a few questions now. One question. Excuse me. How can you tell whether fire damage came from the fire itself or from the type of wall finish? So I guess they're asking for kind of a little more detail on what you found with the paint and wallpaper finish results.  

Shuna Ni: Yeah, uh...It's very difficult to distinguish both of them, because usually, because the ones we were we using the foam fuel packages in those tests, so there's a flame impinging region, which is area where the drywall was in direct contact with the flames. So, in that area, usually you can observe the flame was spread over the paint, because the paint also has some combustible materials, so, and generally I don't think the architecture finish contributes too much to the to the heat, because the layer is relatively thin, right? Relatively thin. Well, and also the boundary of the fire pattern is mainly determined or still determined by the flame impinging, well, that is for paint, while for the wallpaper it's a different story because it's a primer we use, which can accelerate the peeling of the wallpaper under the heating. So we can observe the wallpaper peeled and detached and the flame was spread very fast over the surface. And this, usually in fire’s past, with the wallpapers, the flame impinging mainly generates a darker region, right? And we can see the paper cracking, so drywall surface paper cracking, and also we can observe the drywall calcination.

But above the impinging areas, you can see a lot of soot, which comes from the burning. Even though the wallpaper itself has already peeled, and we can find that underneath the drywall surface paper is not burned, and they only have some soot. So this is mainly because in the wallpaper, you can distinguish the two differences, but for paint, it's very different. Difficult to distinguish whether the damage comes from the paint or whether it comes from the flame impinging.  

But generally for paint, the heat from the paint itself is very limited. Yeah.

Gregory Dutton: All right, thank you. Another question. Were there other factors like weather and aging of the walls that affected calcination and burn patterns and depth? If yes, did they cause any difference or hindrance in the result?  

Shuna Ni: Yeah. For the aging, we didn't study this pattern. We assume the gypsum is relatively stable as time goes. But the weather, for example, moisture, it really impacts because when the gypsum absorbed water, it becomes softer. There are already some studies on the impact of water on the calcination depths. So, so usually when you observe more moisture the measurement will be larger, but this can be reduced as the time goes when the drywall becomes dry again and the calcination depth can come back to its, it's not a complete comeback. It will be closer to its original. Yeah, I'm not sure whether I'm still, yeah, yeah. Okay.

Gregory Dutton: Yeah, interesting question, because I suppose with fire investigation, there's always the issue of fire suppression interfering with your evidence. Yeah. Another question here. What was the fire accelerant used in this study? And would you say that the fire pattern would change depending on the type of accelerant or fuel?  

Shuna Ni: You mean the fuel, right? The fuel we use is a polyurethane form, so it's a type of foam, and we’re also using the haptic tube as igniters. And in our study, we're using consistent fuel packages. We didn't change the fuel type. No.

But there are some other studies which focus on different fuel type, for example, solid fuel or liquid fuel. Usually solid fuel, which will burn less stable, so the pattern will have more variability, you know, will be if you're testing the same part or under the same fuel, the pattern can be different. While for the liquid fuel, usually the heat release rate is relatively high.

So you can see the burning usually the flame is higher, the fire pattern relatively is higher. But because liquid fuel, the burning process is relatively consistent, so the repeatability of the test is much better.  

Gregory Dutton: All right, thank you, Shuna. For the sake of time, we need to move forward, but thank you so much. We will now hear from our third speaker, Meredith Coon. Meredith, the floor is yours.  

Meredith Coon: Yeah. Thank you. Hi, yes, can you all hear me? Yes. Okay, great. Thank you for having me. I really appreciate the opportunity to present the research that we've been working on.

So this particular award involved several deliverables. I'm going to try to cover as much as I can in the 20 minutes, but please feel free to reach out to myself or Dr. Thomas Busey at Indiana University to cover anything we might have missed in this talk. We presented this at a few other forensic conferences, so happy to accept any questions offline if we can't get to everyone.

So our technique is referred to, we're calling it Base-LR. It is an interface for forensic laboratories to generate likelihood ratios on latent prints. So I am the subject matter expert of this particular study. I am a latent print examiner. If you can go to the next slide, please.

So just a disclaimer, these are our own and are only based on our training, analyses, and experiences. I am a practitioner, so I'm putting this disclaimer in the slide to indicate this doesn't reflect  any past or current employers’ opinions. Next slide, please. So I'm going to, thank you all, National Institute of Justice, for helping fund this research.

So we're at the conclusion of our three-year grant and we were presenting presentations to the International Association for Identification. That's where this slide comes from. Next slide, please. So our first publication for this grant has been published. It will come out in the December 2026 edition of Forensic Science International - Synergy, and it is currently available now.

It's free and open source and available to anyone who would like to read it or reference the first publication of the series. Next slide, please. So what is it? The first part of our research was a black box study. So it includes open source, freely available images for research and laboratory use. So this is 513 latent print impressions, 513 triplet pairs.

So each of the latent fingerprints have been matched up with corresponding mated and non-mated exemplars, which were pre-selected, categorized, and labeled for the use in both research and laboratory practice. So we have performance data of 104 examiners on a subset of 90 of these triplet pairs out of that total 513 impressions. So not all triplet pairs actually have performance data of practitioners.

Next slide, please. So, for example, the image all the way on the left here, you'll see the watermark at the top of the screen. AEK is the initials used to track a particular subject. The other letters there indicate the latent and the processing method and the finger position of the person who left the impression behind.

What's different about this study is our graduate students. So thank you once again to our graduate students who assisted with this project. They actually recorded the impressions as they contacted a surface. So for example, if they were to hold this cup here, a photograph of the actual fingers were tracked to help maintain the actual ground truth positioning of the fingers, as well as the position of the participant who left the impression.

So the subject matter expert, which was me, I helped aggregate all three of these images and these would have been provided to our subjects during the performance phase. So our participants were given one latent impression, they were asked to analyze this impression, and then they were partnered with another participant who would have gotten the opposite exemplar. So one person would receive the mated exemplar, which is the middle image, and another person would have received the non-mated. So we have an even 50/50 split of our participants and whether they received a mated or non-mated image pair. So next slide, please. I performed a star and color ranking on all of these. I'm not going to go into all the specific details as to what these are because they're very, very detailed and are included the OSF site includes a spreadsheet which ranks all of the latent impressions and the exemplars for their ability to reach different decisions. This is entirely a subjective decision criterion, but it might be useful for laboratories to help create a test environment for any practitioners who are either in training or they want to test any specific criteria for their laboratory or for research.

Next slide, please. So the star system that I created for the purposes of this study was also compared against the rankings from LQ metric. So there is a fairly good correlation between what is considered a high score. LQ metric determines the overall quality and ability of an examiner to reach a decision in the forensic comparison space.

So the higher the value of LQ metric, the more inherent information is available to reach decisions in latent current evidence. So there is a correlation here between my star ratings and a more objective computer measurement tool. Next slide, please. So here's once again, the LQ metric scores across all of the image pairs in this study.

The median was 67. And for those familiar with other error rate studies in latent print evidence, we're referring to black box 2. This is the Hicklin 2026 data. That study had a median of 47. So overall, the quality of the impressions, the latent impressions in this study were of higher quality than those which were observed in the Black Box 2 or the Hicklin 2026, that study. Next slide, please. So each triplet was also given a color rating according to criteria on whether or not I felt different groups of practitioners would reach certain types of decision making. For example, if I ranked a triplet pair yellow, it was in my opinion that I would expect inconclusive to be the most likely outcome of all participants with that particular latent, whether they were given a mated or non-mated trial.

So there was one interesting, one interesting aspect of the evidence received. I did not observe any triplets where I felt an exclusion decision or a support for different sources would be more likely than identification decision. This is not a very common finding in the latent print field, and therefore it was not labeled with a color scheme.

So if anyone goes into detail here and wonders the asymmetry of the decisions and the color criteria, I'm happy to answer those questions as well. So figure 6 and a corresponding discussion section of the paper details what these triplet color keys indicate, just once again for a selection of research purposes. Next slide, please.

So here you can also see the proportion of the color and the star ratings to kind of indicate the overall quality and information available in the latent impressions. Next slide, please. This is the ordered probit model data. So the faint lines indicate each individual image pair and all of the decision criteria are in faint black vertical lines there. Our participants were given a five conclusion source criteria, which at the time was the current ASB definition and criteria column categorical labels were the ones that were utilized in the research.

So the likelihood ratio at the source identification threshold, which is the threshold right around that value of four and a half on our x-axis here, was 74.36. That's our likelihood ratio. The largest likelihood ratio in the study observed was 1.6 million. There were 31 unanimous source identifications and 17 unanimous source exclusions. There were no false positives in this research.

There were false exclusions, so things that were mated pairs that should have reached, gone over into the right side here. The blue are the mated pairs and the red are the non-mated pairs. And the bold lines are the sum and renormalized curves of all image pairs as an aggregate compared to one another.

So at the top here, Black Box 2025, this is the data for that. So next slide, please. This is one of the charts that is found in the supplementary data. The highlighted row indicates where most practitioners were willing to use a term like source identification, where you have as more and more examiners are willing to use a term like source identification, we enter into majority support.

So the second to last column here, majority support being true, 20 examiners reached a term of identification, six said inconclusive similarities, six said inconclusive, three said inconclusive with dissimilarities, and one said source exclusion. At this point, we would expect a forensic laboratory, the odds would be very good that this laboratory would report a decision like source identification, and this is our threshold point here. This image pair number 35 would receive a 74 would be the likelihood ratio. Next slide, please. Can you click again? Thank you. So when we use a term like identification, or in this particular study, the term source identification, we have to discuss how fact finders might update their beliefs about the facts in a case.

And current quantitative measurements are also offered by other authors, such as Neumann et al., and they give estimates of the strength from 100,000 to 10 billion. These statistical methods are based upon minutiae, so they're actually utilizing databases and the angles of minutiae to perform a computer-based calculation of the placement of these minutiae. Some of the weaknesses of the existing statistical methods are they rely upon minutiae base characteristics only, while latent print examiners will describe the utilization of the overall pattern, the ridge texture and flow, and other anatomical functions such as creases or wrinkles or other forms of characteristics that are discriminable between individuals of the skin.

But the Neumann et al. statistics and another technique called FRStat, these two methods are relying upon minutiae only. So the goal of this research was to try to encapsulate more of the information that examiners might be utilizing other than minutiae. And the numbers above here, the 105 and 1010, When you observe the error rate studies performed by other authors and put the strength of the evidence of it, overall error rate study performed the Black Box 1 and 2 studies with Hicklin and Ulery, they put the strength of evidence in the 100 to 1000 likelihood ratio range.

So this was kind of at a disconnect between what the Neumann data was offering and what the error rate studies were kind of illustrating as a discipline. Next slide, please. So our approach is BASE-LR. It relies on human examiners to compute the strength of evidence on a sample of images with known ground truth. Can you click ahead for me? So we want to provide the strength of evidence for these images and develop and validate an interface that extends the strength of evidence to operational casework.

So we've done all of these things. The tool to use is free and open source. It is self-validating and does not require exposure of casework materials outside of a laboratory setting. So next slide, please. Sorry, we need to click ahead. Thank you. So this, could you go back? Thank you. So this BASE-LR stands for Benchmark Anchored Sorting Estimation Likelihood Ratios.

So that is the URL. The link is live. So if anyone wants to play along at home or go to this link at a later date, you can kind of walk through how the tool would work. I'm going to give a quick refresher or a quick update to you all. I've only got a few minutes here, so I don't want to get too much into detail, but this tool is currently live.

So next slide, please. So the goal for each image pair in the database is to not only test those image pairs that we discussed at the beginning, those triplets, those combinations of those 104 latent print examiners across 90 image pairs, we were assessing how those latents performed across the group of practitioners. So if we take a whole group of practitioners and all of their different opinions across those five conclusion scales, we can get a relative strength of support for how much information those examiners feel is available to reach decisions.

So some practitioners would look at an impression and say, every examiner felt that this was belonging in the source identification category. But as I showed in some of the tables there, some image pairs result in less conclusive decisions and a larger spectrum of decision making. So the discussion here is a likelihood ratio less than 100, like that 73, for example, doesn't seem to square up with the Newman statistics of 105 or 10 10, but these likelihood ratios are still significant in this space.

So comparing forensic disciplines and likelihood ratio calculations can have a really different impact on the justice system. Next slide, please. So the goal here was to generalize this to casework.

So once again, thank you for the funding because a lot of laboratories don't have the resources in order to develop a tool like this. So the key step that we had here is we don't actually ask the practitioners to provide the specific information that they used to reach their decision. We did not ask examiners to, we weren't actually collecting minutiae data at all.

So examiners would reach a decision of those five possible criteria. And then across the group, we asked practitioners to compare comparisons. So an examiner would be given a latent print and a known print. They would be asked to assess it in whatever way they, whatever criteria they feel is relevant to reach decision making like they would in normal case work.

And then we would provide another latent and another known and another latent and another known. And after they have a whole set of those pairs, we ask examiners to sort those pairs across their computer screen. Next slide, please. So we asked examiners to conduct six comparisons, and they had to keep mentally in their mind. They could take notes if they wanted to, and they could zoom in on the pairs once they were all placed on the screen.

The examiners were then given a graphical interface, so just a nice big white screen, to place each image pair such that the horizontal location corresponded to their belief about the relative strength of the evidence. So the further to the right on the screen meant more support for the same source proposition and image pairs which were further to the left indicated the least amount of support for the same source proposition. Next slide, please. So this is an example of what that interface might look like. The little blue tick marks are there so that the examiners can indicate the placement and the relative spatial arrangement of the image pairs to one another. So for example, on the right side of the screen where it says pair 157, the latent print is on the left and the known print is on the right and they are stuck together.

They are a single image that the examiner can drag right to left, and they can order them from right to left across the screen after they have spent enough time in a zoomed in state with the latent and the known, and then they're given this interface to provide a sorting technique. Next slide, please. So to validate this process, we asked examiners to conduct an additional casework analysis. These were all ground truth. We just changed the name on this to refer to a casework one, and they were asked to place it relative to the six image pairs that were their benchmarks. Next slide, please. So, next slide.

So this blue highlighted one would indicate the casework pair. So they would be given all six of the other images in the background. They would click done with sorting. And the horizontal position of all of the benchmarks are locked in place at that point. So this new one that is highlighted in blue, the examiner can move it right to left across the screen to indicate where it falls in relationship to the other image pairs.

Next slide, please. Next slide. Oh, yes. If you can continue to click through, sorry about that. Could you go back for a second? To the previous slide. Thank you. So we know the strength of evidence for the casework image pair because it is part of the black box study. The goal was to compare its horizontal location to the computed strength of support based upon the group of practitioners who would have received this image pair and based upon where this individual examiner might have placed this image pair in this study.

Next slide. Thank you. So 64 latent print examiners who were currently eligible to testify in the United States, they had previously completed our part one black box study. They were given two conditions. One is the free sort condition where they were asked to place the six benchmark pairs first, and the second was the fixed sort position where these benchmark pairs were placed prior to the examiner being able to move them.

So they were locked in place and they could only be moved vertically. Next slide, please. So all examiners completed two sets of these six benchmark image pairs. The free sort did their own sort and the fixed sort had them locked. This is kind of repetitive, sorry. Can you click the next slide? So they were repeated with six new benchmark image pairs and five new casework pairs. Next slide, please.

So our hypothesis was that we were trying to determine between the two groups whether the fixed sort or the free sort condition might result in better analysis and interpretation of this method. Next slide, please. So you can see here that the benchmark pairs, this is a free sort subject. So subject 307, they were given the ability to free sort, and they were provided a regression line with their benchmark pairs, which are marked here in black.

The casework pairs that they received later are marked in red to indicate their overall regression. So if you can click to the next slide. Yep. If you can keep clicking, thank you. Yep. So the estimated strength of each image pair is based upon the regression line. So there is a plus or minus factor of about 10 to each of these likelihood ratios. Next, please. Just an order of magnitude. So there's a deviation between the actual and estimated strength.

You can continue forward. Thank you. So there's an additional advantage to the free sort condition. We can use this to actually prune bad subjects. So if a participant doesn't understand or doesn't perform to the level of the group of participants, we can look at their regression line and indicate that they might not work well with this interface.

So next slide, please. So, we're looking to see whether the fixed group or the free sort were more accurate, and how strong the regression coefficients were in the free sort condition. Next slide, please. So overall, our subjects performed very well at this task. They performed best when they were given the free sort opportunity.

Next slide. So virtually all values are above the .85, and those few subjects that were below the .85 because they were given two sets of image pairs, They did not repeat the low R values. So there is some indication that the pairs that were used in the particular interface had a specific performance factor when using this tool. Next slide, please.

If you can click one more time, I believe there's a box here. Yes. So the free sort condition involves the image pairs listed there, 002m, 003n. So two non-mated pairs were provided to every four mated pairs. The free sort had the lower error and they have a lower root mean square error and performed better on this particular task.

Next slide, please. So if the user doesn't agree with the group consensus on the image pairs that they were provided, the software tool does not provide them a watermarked certificate, which is on the next slide. This is how the interface would look, is the examiner would not actually expose their casework or their laboratory evidence to the system. It provides a little thumbnail for them. Next slide, please. So this is a certificate that would be provided. The user ID can be indicated as well as the date and time.

Practitioners can also provide in the note section what relevant case is associated with their particular laboratory exercise. So the correlation R-value is how well this examiner has sorted the benchmark pairs. The interpolated mu value and the likelihood ratio are referring specifically to the image pair that the examiner has in front of them at their desk. And the support for the same source proposition statement appears below to indicate how much the likelihood ratio value is. And the examiners can then export this into their case notes. Next slide, please. So here we have the, how accurate are the computed strength of evidence values? We are about one order of magnitude within the true value for the samples that we provided these practitioners.

So overall, the subjects performed very well at this task. Next slide, please. So we would recommend this tool is best suited for borderline same source cases. It could be used for exclusions, but likelihood ratios less than one might be confusing. There is a large need for statistical literacy in both the friction ridge examiners and in the juries themselves.

So this might, it's a clear need for more communication research and understanding the relevance of these particular likelihood ratio reportings and how they might impact the larger judicial system. The order-probit model also does not do very well with unanimous decisions. So if all practitioners reach a decision like source identification or source exclusion, the model tends to approach infinity and therefore is, doesn't, it has poor, it's poorly constrained at the far ends. So we would recommend that this tool be used when practitioners do not anticipate a unanimous decision or when inconclusive is appearing in the laboratory analysis.

So most of the likelihood ratios following these guidelines would be within the 100s to 1000s. So this would help provide a little bit more granularity on those more difficult decisions. Next slide, please. So this builds upon a community-based database and quantifies the strength of support. It can be reported directly or used to augment traditional reporting scales.

And we can now estimate the strength of the evidence on something that examiners repeatedly describe as a more holistic form of reporting as opposed to a minutia-based form of reporting. Next slide, please. So, of course, there are challenges like likelihood ratio interpretation. Replacing categorical conclusion would involve the rethinking of validation studies, proficiency testing, and perhaps even verification. But our argument is that having a likelihood ratio system of reporting is an alternative to existing categorical opinions, which some have argued are a posterior decision as opposed to a likelihood ratio decision.

And the goal is to not take away the authority of the fact finder, like the jury or the judge. Next slide, please. So this is just our slide here for practitioners. If anyone here is a practitioner and has any interest in continuing with doing any other additional research with Dr. Busey's lab, that is the QR code. At this point, this is essentially the end of the presentation, if we have any time for questions.

Gregory Dutton: All right, thank you, Meredith. So we will have time for a couple of questions. A reminder to everyone to submit your questions into the Q&A. I did want to make a note: It came up in a couple of questions.

Recordings of the presentation will be posted within the next few days. So look for that. Meredith helpfully gave us some links. So we posted links to her publication. I believe it's open access. So you can see it free of charge and then the link to the BASE-LR tool. One question:  is the tool already used in any forensic labs?

I know it's relatively new. Yes, it is. And then a follow up to that: how would the likelihood ratio values be presented in the courtroom?

Meredith Coon: So the tool is just now being launched. We just got the website up and running and we've just started presenting. All of the publications associated with this research have not fully come out yet. So to our knowledge, no one has utilized this tool yet in court.

I know some AFIS vendors were very happy to get access to these open source images because it's hard to find ground truth data. So the concept of reporting likelihood ratios in the courtroom can be taken in a lot of different directions. Our job is not to advise policies specifically. Some laboratories are already reporting on likelihood ratios, especially in some European countries, even in categorical pattern evidence decision making.

You can look at DNA as an example of a discipline that reports in likelihood ratios. Kind of our baby step suggestions would be to partner up likelihood ratios with categorical reporting. Other techniques such as the FRStat, it provides a global similarity statistic that is not technically a likelihood ratio. Two.

Henry Swofford helped to develop that tool and he's happy to answer questions about the FRStat, but that technique was developed and it is typically reported in addition to a categorical decision. So they can kind of, they can kind of provide a little bit of a granularity or spectrum to an otherwise categorical decision. Is one option.  

Gregory Dutton:  Excellent. Yeah, I know that just the availability of this new data set of images is really an exciting thing. I was wondering, could you remind me again, I thought that the certificate was interesting. Remind me, what were the conditions you said that a certificate would not be issued? Yeah. Yes.

Meredith Coon: So the examiner has to be able to sort the, they have to be able to sort all the image pairs across the screen from left to right within a 0.9, 0.85 regression line. So they have to get the non-mated pairs on the correct on the left side of the screen and the mated pairs on the right side of the screen and they have to sort them in a similar relative position to the black box study participants. So looking at the overall distributions of examiners' opinions on the strength of the evidence. So a practitioner has to create a user account, which is free to do, and Dr. Busey's team will help someone create a user account. Then they have to be able to analyze and sort all those impressions across the screen. Yeah.

within a 0.9% regression line, and then they will be given the opportunity to generate a certificate. So if they don't meet those criteria, the certificate just has a big watermark across and it won't generate the numerical value. So the likelihood ratio and mu values will not appear on the certificate. We don't anticipate that this certificate process at this regression line would be that. Most examiners performed well at this task.

So it's very interesting to see examiners being able to verbalize and illustrate that they feel there are different variations of strength of evidence, even within a same source criteria that is otherwise categorical. So I found this data very exciting.

Gregory Dutton: All right, excellent. Thank you, Meredith. We will have to move on. We'll now hear from our final presenter of the day, Professor Ranjan Maitra. Professor Mitra, the floor is yours.

Ranjan Maitra:  Thanks. So, the title of my presentation is Quantitative Matching of Forensic Evidence Fragments Using Fracture Surface Topography and Statistical Learning. So, I'm the second author on this on this group here, and together we are a team of material scientists and engineers, statisticians and practitioners who have come together in this long-term work, and we acknowledge and thank NIJ for the funding. Next slide, please. And so here we have the standard disclaimer that only we are responsible for our findings. And I'd like to go on to the next slide, please. So, so we will, so I will talk about our motivation, our objective, hypothesis, proposed framework, how we generated samples and measured fracture surface topography. Then I will illustrate the performance of this technique and we will have some and we'll have some concluding remarks.

Next slide, please, 140. So our story started from the National Academy of Sciences report, that's about 17 years old now, but that suggested that we should establish, provide a scientific basis and reliability of many forensic tools. With our expertise in fracture mechanics and statistics, we felt that there's a lot of information hidden on the fracture surface that can support the uniqueness of identification, and that is what we'll try to focus on here.

So, for example, next slide, please. So, for example, on a macro scale, we can try to solve a jigsaw puzzle of fitting parts together, but the details of the topography here carry much more information that is unique to every fracture event and that is what we should, we will try to show and exploit. Next slide please, 142. So, microscopically, when we do comparisons, what we try to do, for example, is to match a texture in terms of ridges or particular microscopic features. So, next slide, please, 143. So what we rely on here is to go down to the microstructure and try to find what scale we can take this topology and compare it together.

And the way we'll do that is by looking at the fracture surface through imaging. So we do have a couple of issues here. First is how we actually do the measurements, and second, what scale we are going to run the measurement at. Next slide, please, 144. So for a quick overview of fracture mechanics, this is a statistician understanding it.

With a material load fracture, what happens is a material load fractures when the stresses of forces increase at a critical distance from the crack tip, and that critical distance is about two grain size for brittle material like stainless steel, for example, cutlery steel that is metallic but semi-brittle. So anyway, while it is stochastic in nature, what we have is we have to raise the stress ahead of the initiation crack at about two grain sizes in diameter, and that really is the starting idea for setting the scale of our measurement. So next slide, please. Then another thing that really happens in fracture is the stream that the crack follows. So it is stochastic, of course, but it is influenced by how big the grain of the material is. That is one driver. Fracture event, how the material, how this fracture event happens, how it is loaded, and how each, because each random loading has a unique characteristic depending on whether we are bending, whether we are twisting, shearing, etc.

Further, did it get exposed to environment? For example, was corrosion, was corrosion involved that can dissolve or change the topology, as well as the intrinsic resistance of material. How does this grain pooling or damping happen. Next slide please, 146. So when we think of a material breaking here in a simple representation, let's think of a bar I'm pulling on it.

So we can see the dimples coming out of the plane. But if I bend or twist, we see the dimple is going to be arising at an angle. And the surface has a lot of details unique to the event that that triggers this fracture.

Next slide, please, 147. As an example, we put a simple pull test on a sample on a specimen steel sample, and the crack runs through it. But what will happen is that it will read different signatures at different scales. So at the 2 millimeter magnification here, we start to see a big striation that is corresponding to how the crack is running or making those like fish scale patterns if we go down to the details.

And we find that we start to get like these fissures and those are related to the micro mechanism or further down we get more details that is account that is akin to the macro mechanism of failure. So the question is, can we utilize these different scales to arrive at its signature or a protocol to enhance a fracture match? With regard to this particular picture, we showed a picture to show this picture to our forensic collaborator, John Vandercock, and asked him what magnification he would suggest for identification. He said, OK, at most I'll take A&B in the top two figures.

But we show here that going down in the scale, we have more leverage and more details and that we can uncover that is unique to the process and that we can use for identification. Next slide, please, 148. For example, let us consider the topography of a cracked surface. We have a half millimeter by half millimeter view here and our topology here is measured by roughness.

So what we are putting here on the horizontal scale is a measure of roughness, that's what we call it a point-to-point correlation, as a function of the change of roughness, which is delta H. Okay, we find that the roughness changes based on the window that we are measuring it on, and it is proportional to that window. So this relationship identifies the scale where we can start to get unique texture on the surface versus something that reflects the random properties <unintelligible> of the process getting fractured, the surface getting fractured. Next slide please, 149. And for this particular sample, it's about like 30 microns.

This correlates directly to our grain size. For this material here, l0 is about two grain size, with each grain size being about 25 to 30 microns. So really, to be able to measure that scale, what this means is that we have to measure a characteristic scale of about twice the grain size, so about 50 to 70 microns.

And then we have to measure at least 10 periods of that feature in order to avoid aliasing. So if that is the case, that immediately determines our field of view, which should be about 10 periods of this characteristic length. So that means we have to have about 10 to 20 grain size in our field of view to get this unique transition from a self-affined to a unique texture.

So that is where we have to look for the surface at the surface to catch the unique identifier for this particular event. Next slide, please, 150. Simply put, if we take the data along one line and we use a bandpass filter, we have a high pass that will be the high frequency content.

This will be the grain distribution or the sub-mechanism of failure. But we have a lower frequency, that is the green ones, these are the green ones here, and that is like an indication of the river marks. Or the or the macro, sorry, or the macro fracture event that is unique to the event that we are looking for. Next slide, 151. So we will take a Fourier transform of this to put the spectral decomposition and see the range to focus on. So let's go to the next slide, 152 please. So here is the implementation protocol. We first generate a sample. We have a control set, which is a stainless steel alloy that is used in, for example, knives. So we have actually a knife set, a hacksaw blade set, and we also tried to do this on silicone replicas, casts, to see how things work out and how the fracture topography gets replicated.

So we tried different fracture modes, meaning that we pulled the sample to break or bend it or twisted it or did both. And so these can really generate different topographies. And we did imaging with confocal microscopy and spectral analysis using frequency space. And then we developed a statistical classifier to identify what is a match and what does not, what does not, what is a non-match.

Next slide, 153, please. So the first one of the uniform sample here that we did is we broke it in using a three-point bend where we can get the two pairs and we start, then we start to image them. Or we have a set of knives. So in this case, we have a set of knives that were consecutively manufactured and each of those were broken. Next slide, 154 please. And the way we did it, we put it in a vice, then we went back and forth until the knife tip broke, or we tried to twist it and to get a different topography. This picture here is of a failure under bending. We see it is just like we have a clear texture here.

Next slide, please, 155. For bend and twist, we will get the river marks, and if it is only a twist, we see the features are really different. There are approximate characteristic ridges here. So those spacings of those features are essentially what we will use for identification. Next slide, please. A quick illustration here. We have a little fixture with a comparative microscope.

We put the base on the tip. We align an image. And so we have one set of images on the base and another on the tip. And actually, we do multiple images, which we'll come to in a little bit. Next slide, 157, please. So we did a series of overlapping images, and we studied what should be the overlapping, how much should be the overlap, so we get some reduction in the noise of imaging.

That is why we are using overlapping images to reduce the noise in imaging. So we used 75% overlap, but we did nine overlapping images. So we will see, probably we do not need, we did not need to do nine images, and fewer would be fine.

We'll see that in a little bit. So next slide, please, 158. So analysis protocol has an imaging part and a model training part and a classification part, the specimen classification. So next slide please, 159. So with regard to the imaging, we'll move to the frequency space in order to reduce issues of alignment.

And that gives us more freedom in handling the samples as well as identifying those ranges of frequencies where we can get unique sample characteristics. There are some standard image processing filters that have to be applied in order to use the frequency space. This stuff is all standard in image processing, and I will not talk more about that here. Next slide, 160, please. Then, in the training, we have a set of characterized samples that are known matches, and we computed correlation of aligned pairs of matches for these matches and every other pair, which is a known non-match, right?

Thus, we got, and then we got, sorry. a statistical model for the match and the non-match distribution. Then after training the model, we do a classification of a specimen into a match or a non-match, and since we have multiple images per sample and different overlap amounts, our model will also account for that.

Next slide, 161, please. So, we will use a matrix-variate-t distribution to model the higher complexity levels of the data. This is because for every sample, we have a matrix of two sets of frequency ranges and nine images, or whatever, however many images we decide ultimately. A match probability is then calculated and used to classify a specimen as a match or a non-match.

Next slide, please. 162. So operationally, we image and do the first wave transform, and then we decide the range of the frequency into a frequency band corresponding exactly with what we did in one dimension in the illustration, where this will be the range of unique features, and after that, will be the range that will identify the micro mechanism of fracture.

And that is what we'll use in our identifications. So next slide, 163, please. So to show how we do this, we get a representation of correlation between these different bands. And continuing on, next slide, 164 please. We show here, we show here those correlations. So for the 5 to 10 frequency range, you can see that the reds are the non-matches and the blue are the matched pairs.

And we see that we have full separation between them, right? If we go for the 10 to 20 micron range, which is the next figure, second figure, while we still preserve separation, there's a little bit of overlap. And so these are the two ranges that we use for our bands for comparison. If we go higher, it starts to get less distinguishable between match and non-match and we get more overlap because this is the basis of the micro mechanism of failure.

Next slide, 165 please. So in our analysis, the first batch we did, we had 81 matches and 648 non-match pairs. Basically, the 81 are the matches and everything else is a non-match pair. And the way we settled on that is we decided to have a set of overlapping images for every sample that we automated our framework to do our imaging here. And we are showing here a set of base tip pairs, how fractured, how they correspond to each other with a given overlap. Then once we take a spectra and calculate the correlations for those frequency ranges of 5 to 10, and 10 to 20 microns in diameter.

Next slide, 166 please. So when we take those correlations and put them on a scatter plot to represent the match and the non-match over the two frequency ranges, a good point here is that sometimes images can have noise in the signal or in the imaging. For example, we could have a missing section, for instance, as in a valley or something in the <unintelligible>, right? So, our decision is not based on one image, but based on the collective response of the set of overlapping images that we have. Next slide, please, 167. So, in this figure, in this figure, which is basically the same as the, similar to the last slide, we see that about seven or eight of them are within the distinct cluster, but maybe one or two images in the matches have a low correlation.

So we increase our signal for identification by combining them by using all of them. We have the set of match, non-match with nine overlapping images and two frequency bands. So each comparison pair has a two by nine cross correlation measurement matrix, which is why we use the matrix variety distribution to model the Fisher transformation of these correlations.

To model match and non-match and based on that, we can get to a match probability, which is what we can use to identify whether a specimen is a match or a non-match. All right, next slide, 168, please. Let us now see how this technique is applied. So we have multiple sets of knives, like we could have a bending first set, a second set, or in the next slide, 169, a tensile or bending fracture of stainless steel tool set. Next slide, 170. So we will see how things work with regard to the degrees of freedom. That is a parameter here in our matrix-variate-t distribution. So here what we have done is we have trained on each of the four sets in the previous two slides and predicted match, non-match for all of the others.

And we see while there are differences in the results, we get very good separation between the match and non-match with log odds in terms of log odds ratios here. These are all in the log odds scale, which means that the current predicted matches are mostly being made with between 0.99 and 0.99 recurring to 8 to 9 places of probability of match or non-match either way. So fairly similarly from the non-matches. So we see very good results here and that we also see that the degrees of freedom are not that crucial a parameter to estimate exactly. We should get a reasonable ballpark. Next slide, 171, please. Then we started to ask, okay, we have nine images there. How many images do we really need?

So we let the number of images go as a variable, and we found that onwards from five images with about 50% overlap, this was enough, about enough to really give us good separation between match and non-match, and this would be insensitive to the imaging errors that we can get into. So going on, we have now done it with 50% overlap and five overlapping images. Next slide, please. So this is our slide 172. So when the sample set is reimaged, we wanted to see how repeatable is the process, right? So what we did was we repeated the same imaging with different operators and the mean and distribution of the cluster doesn't really change that much, which is really, which shows us that this entire process is really very tolerant. And that really, that the framework is really tolerant, very tolerant of operator effect. Of course, here the operators went through training, because if they aren't trained, then we can get imaging artifacts of alignment and things like that. Next slide, 173, please. We now decided to try a different experiment, which is how things could change if we mix the process of fracture. So, what we see, and we see that the result is fairly similar, as in the next slide, 174, please.

So here with the lighter colors are of the bent sample and the darker colors of the twisted fracture sample. So you can see that basically they are pretty much wherever they were for the other sample, right? Anyway, so next slide, 175, please. So indeed we see that the correct classification is largely achieved on the twisted sample when we train using the bending and the tensile strength sample.

These show that the classification results are repeatable when applied to new sets of samples or trained on new samples across material classes. Next slide. Excuse me, 178 please. Similarly, we see that we do the same thing with the hacksaw blade set provided to us by our forensic collaborator, Lauren Claytor. We tried the same experiment and what we get in the next slide, 179, is we see that our classifier did really very well. Here the matches are simply the 11 facing edges of this hacksaw blade, which has been chopped into pieces. And everything else is a non-match, and there are 220 of them. Next slide, 180, please. In terms of classification, we get a five to nine separation. between match and non-match. So what that means is we get a very high number that is a green one, and if that does not match, then we get complete separation with a different number and color here.

So next slide, 181, please. We also looked at whether we could use silicone replicas to do the sample generation and the training. So, we took one and obtained a cast iron replica, and we imaged the, sorry, silicone replica, and we imaged the replica. So, next slide, please, 182. So, we have here the original base and the original tip and the topography of each of them imaged and the replica of the tip.

We see that we get the low frequencies reproduced well, and perhaps the high frequencies are not reproduced that much. But it doesn't matter to us, because we are looking for the low range of frequencies, because that is where we see that we have full separation. Next slide, 183, please. Once we get to the micro mechanism of failure, base and tip start to overlap and the replicas start to not be reproducing the behavior well, but as I said before, it is still producing full separation in the part that we are interested in. Next slide, 184 please. And here are the results. So what we have here is that the red is the original base and original tip. The green is the original, sorry, the green and blue are the ones where the replica is done with the original base and the replica itself with the tip.

So we can see that it does pretty well here also. So next slide, 185 please. So then we took the process further and we tried to do a porcelain and ceramic characterization. We identified the scale that will be where we will see this unique texture. This is going to be about 70 to 80 microns.

That's higher than we saw in metal, but that's because of the nature of the ceramic and the porcelain. And we get full identification here. So that is the summary of known match and non-match, and we perfectly classify them. And next slide, please, 186. Sorry, you should have. 184. So, so next, next slide, please, 185. Yeah, so based on the level of this posterior probability, we can start to classify predictions further as extremely strong or very strong, or a strong or moderate, or counter that into sorry, or into or classify that further into likely very strong, or even that it does very weak evidence in favor of a match or a non-match. So, next slide, please.

So we also experimented with a case study separate from the standard set of plates that we said, okay, if I have a porcelain mug that broke, then we tried to identify the section here. And going to the next slide, 188. I'm sorry, I'm sorry, it's 186. I made a mistake. Sorry. Let me go back one.

Sorry. So moving on, we worked more into different textures, into different topology, and it can address, what we can see is that we can address a very diverse range of material and topology and provide good identification with that. We are continuing on with this work, and in the next slide, in 189, I'll have the summary.

So in summary, what we have provided here is we have provided here a promising forensic comparison framework for fracture with a primary objective to support and supplement expert judgment with a mathematically and statistically sound defensible defense decision framework. And we've also accounted for or tried to obtain a proper sample size for statistically significant comparisons. We feel that operator training is important in terms of good imaging skills of alignment, elimination, and proper magnification selection for our methodology to work well. And for the work, we will be also looking at applications to different classes of material and fracture.

We also plan to look into environmental degradation, such as with corrosion and things like that. Possible further investigations will be into plastics, auto parts, wood, tape, etc. With that, next slide, please, is our list of references pertaining to this work and I thank you and I welcome your questions.

Gregory Dutton: All right, thank you. So let's move into the Q&A. Reminder, submit your questions into the Q&A box. We have a few questions already. One questioner anticipated the last part of your presentation. Maybe you can expand on it. So the results you were showing were mostly on stainless steel, but their question is, is the model dependent on the material composition, its fracture and failure mechanisms, or could it be applied to other non-metal materials like polymers or bones?

Ranjan Maitra: So, that's an excellent question. I do not know about bones, but that's an excellent question. We have tried it on materials, so what it seems like is, so basically our original thinking was that metal of one kind would be transferable to metal of another kind, and of the same kind, sorry, and so on and so forth, right? But it seems to be far better than that. So it seems to matter really on grain size. And so what that means is we have used as long as the grain size is similar, it seems to work very well. So that's pretty good. With regard to polymers, I think there is a problem because of how polymers are.

And we don't think we actually, I'm sorry, not polymers, I'm sorry, sorry, sorry. With regard to glass, there is actually a problem, and that will require more work. We are not completely sure about the frequency ranges. It will need more investigation. Interesting. I hope that answers the question.

Gregory Dutton: I think it does. Another great question. For forensic practitioners to apply this model, would it require a 3D confocal microscope?  

Ranjan Maitra: Um, yes, I believe that we will need the imaging environment. Yes, that's correct. Right.

Gregory Dutton: So I believe there are a number of instruments that are available out there, but it would, yeah, it would require some sort of 3D topography imaging instrument.

Ranjan Maitra: Yeah, for some image, yeah. Yeah, I'm not very familiar with all the instruments that are available. I'm the statistician, but yeah, we will need some ways of imaging. Perhaps not confocal, I do not know the answer to this question. But yes, some imaging will be definitely needed, yes. Okay, yeah.

Gregory Dutton: Right. I think there are others. Confocal is just one type. Another great question. Is there a minimum sample size required to conduct the comparison examination before error rates are substantially affected? For example, if only a small piece is recovered instead of a whole knife tip.  

Ranjan Maitra: So that is an excellent question. So in our, so when we did this knife tip and base, right, we basically had a fairly small portion and we found that 50% overlap of that was good enough, right?

So that basically gave us five images, five overlapping images into that. So I am not 100% sure how small is very small, right? It's kind of difficult to come up with this answer, I think that in my, based on the experience on this on this project, what I what I would say is, if it's very small, it would still work if you can take a number of images to basically account for the fact that there's noise in there, right? So, for that, I think five is good enough, actually, and that's it, and that would be fine. Now, if you can only take like two images or something, I do not know.  

Gregory Dutton: Can you remind us that the size, the length scales of the images?

Ranjan Maitra: So the length scale of the image, one second, sorry, I have to go back. Oh, I guess I have to go back here. I think mine. Sorry, let me see.

Gregory Dutton: I recall that it's... It's related to the grain size, correct? You want to have...

Ranjan Maitra: Yes, it is related to the grain size. That is true. And the grain size was like, so basically, so it is five to, hang on one second. Sorry, I'm just pulling it up. It says there.

I guess. So it is basically the field of view was, it's written.

Gregory Dutton: So. It's not critical. It's something I think we can look up later, but...  

Ranjan Maitra: Sure, I mean I have it all, one second. So it is right here. It is about, yeah, 400 to 500 micrometers. Well, that was the field of view that was used.  

Gregory Dutton: Okay, excellent. So relatively small if you're thinking only five of those might be adequate.

Ranjan Maitra: Yeah, yeah, yeah, yeah, that's right.  

Gregory Dutton: And, and they're overlapping, so do the math. Ranjan Maitra: They are overlapping, that's correct. They are overlapping because we want to get as much information as possible.

Gregory Dutton: All right, excellent. Unfortunately, we don't have time for any more questions. We've reached the end of our session.

Thank you, Professor Maitra. We really appreciate it. Thank you to all of our speakers. And that's a wrap for today's webinar. Thank you so much to everyone for spending time with us today. If you enjoyed this and you'd like to hear more about NIJ supported research, we have sessions every day this week. So tomorrow's session is on forensic anthropology and forensic pathology.

You can register, if you haven't yet, for each day's session at NIJ.gov. Again, a recording will be posted in the days to come. For any follow-up questions, please contact [email protected]. And we look forward to seeing you later this week. Thank you. Goodbye.  

Ranjan Maitra: Thank you.