Contributed Poster Presentations: Section on Statistics in Defense and National Security

Computer model emulation --approximating expensive simulations with surrogate models-- has become an essential tool in uncertainty quantification and scientific computing. Various methods, including Gaussian processes, basis function expansions, and deep learning, have been developed to improve prediction accuracy and computational efficiency. However, their relative performance varies across different problem settings, making systematic evaluation crucial. In this work, we present an extensive and reproducible comparison of 11 emulation methods across 40 simulated and 25 real-world datasets. To facilitate standardized benchmarking, we introduce duqling, an R package designed for organizing and evaluating emulator performance on common test functions and real-world applications. This study provides practical insights into emulator effectiveness and offers a robust framework for future method development and comparison. 

To interpret the value of forensic evidence resulting from paired item data, the common source identification framework asks: do the items share a common unknown source or come from two different unknown sources? This question can be addressed using a variety of forensic statistics techniques, including the usual Two-Stage, Likelihood Ratio, and Bayes Factor approaches. Contrastive learning methods address the question using two major components: a method for quantifying the similarity (or dissimilarity) of pairs of evidence items, and a method for determining the best separation of within-source or between-source comparisons. Contrastive learning methods are particularly useful when the data derived from the evidence is high-dimensional or complex. In this case, the contrastive learning algorithms take advantage of high-performing artificial intelligence and machine learning tools to avoid specifying complicated probability models for the usual forensic statistics approaches. In this presentation, a contrastive learning algorithm framework is developed for complex evidence and applied to data from aluminum powder particles recovered from two pre-blast improvised explosive devices. 

The biometric capacity of a system refers to the number of biometric profiles that can be enrolled in the system while maintaining a given level of identification accuracy. In this work, we will discuss indirect ways of estimating the biometric capacity in terms of the random match probability between biometric profiles when the profiles are only observed in a noisy environment. Specifically, if the between-profile comparison distribution is indistinguishable from the within-profile comparison distribution, then the two profiles are considered to be indistinguishable and should be "associated" in the corresponding network. However, when only using samples from biometric profiles instead of the actual profiles themselves, we can only arrive at a probabilistic assessment of whether or not this type of association actually exists. We will explore various strategies for presenting network graphs for pairwise associations when the association can only be estimated, and demonstrate how the resulting network graphs can be used to characterize the biometric capacity of a system. We will use a handwriting biometric system to illustrate the proposed methodologies.