Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “explainable neural networks”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 145 records · Page 8
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Understanding Twinning and Deformation in High Entropy Alloys

A combination of high strength and high ductility has been observed in multi-principal element alloys due to twin formation attributed to low stacking fault energy (SFE). In the pursuit of low SFE alloys, a key bottleneck is the lack of understanding of the composition–SFE cor- relations that would guide tailoring SFE via alloy composition. Using density functional theory (DFT), we show that dopant radius, which have been postulated as a key descriptor for SFE in dilute alloys, does not fully explain SFE trends across different host metals. Instead, charge density is a much more central descriptor. It allows us to (1) explain contrasting SFE trends in Ni and Cu host metals due to various dopants in dilute concentrations, (2) explain the large SFE variations observed in the literature even within a given alloy composition due to the nearest neighbor environments in “model” concentrated alloys, and (3) develop a machine learning model that can be used to predict SFEs in multi-elemental alloys. This model opens a possibility to use charge density as a descriptor for predicting SFE in alloys. Furthermore, a descriptor-less machine learning (ML) model based only on charge density images extracted from density functional theory (DFT) is developed to predict stacking fault energies (SFE) in concentrated alloys. The model is based on convolutional neural networks (CNNs) as one of the promising ML techniques for dealing with complex images and data. Identification of correct descriptors is a key bottleneck to develop ML models for predicting materials properties. Often, in most ML models, textbook physical descriptors such as atomic radius, valence charge and electronegativity are used as descriptors which have limitations because these properties change in concentrated alloys when multiple elements are mixed to form a solid solution. We illustrate that, within the scope of DFT, the search for descriptors can be circumvented by electronic charge density, which is the backbone of the Kohn-Sham DFT and describes the system completely. The performance of our model is demonstrated by predicting SFE of concentrated alloys with an RMSE and R2 of 6.18 mJ/m2 and 0.87, respectively, validating the accuracy of the proposed approach.

36 MATERIALS SCIENCE↗
DOE OSTI2022

Out of Distribution Detection with Neural Network Anchoring

This is code to reproduce and build on OOD detection from the paper "Out of Distribution Detection with Neural Network Anchoring". Our goal here is to exploit heteroscedastic temperature scaling as a calibration strategy for out of distribution (OOD) detection. Heteroscedasticity here refers to the fact that the optimal temperature parameter for each sample can be different, as opposed to conventional approaches that use the same value for the entire distribution. To enable this, we propose a new training strategy called anchoring that can estimate appropriate temperature values for each sample, leading to state-of-the-art OOD detection performance across several benchmarks. Using NTK theory, we show that this temperature function estimate is closely linked to the epistemic uncertainty of the classifier, which explains its behavior. In contrast to some of the best-performing OOD detection approaches, our method does not require exposure to additional outlier datasets, custom calibration objectives, or model ensembling. Through empirical studies with different OOD detection settings - far OOD, near OOD, and semantically coherent OOD - we establish a highly effective OOD detection approach.

Thiagarajan, Jayaraman↗
DOE OSTI2021

Flow over an espresso cup: inferring 3-D velocity and pressure fields from tomographic background oriented Schlieren via physics-informed neural networks

Tomographic background oriented Schlieren (Tomo-BOS) imaging measures density or temperature fields in three dimensions using multiple camera BOS projections, and is particularly useful for instantaneous flow visualizations of complex fluid dynamics problems. We propose a new method based on physics-informed neural networks (PINNs) to infer the full continuous three-dimensional (3-D) velocity and pressure fields from snapshots of 3-D temperature fields obtained by Tomo-BOS imaging. The PINNs seamlessly integrate the underlying physics of the observed fluid flow and the visualization data, hence enabling the inference of latent quantities using limited experimental data. In this hidden fluid mechanics paradigm, we train the neural network by minimizing a loss function composed of a data mismatch term and residual terms associated with the coupled Navier–Stokes and heat transfer equations. We first quantify the accuracy of the proposed method based on a two-dimensional synthetic data set for buoyancy-driven flow, and subsequently apply it to the Tomo-BOS data set, where we are able to infer the instantaneous velocity and pressure fields of the flow over an espresso cup based only on the temperature field provided by the Tomo-BOS imaging. Moreover, we conduct an independent PIV experiment to validate the PINN inference for the unsteady velocity field at a centre plane. To explain the observed flow physics, we also perform systematic PINN simulations at different Reynolds and Richardson numbers and quantify the variations in velocity and pressure fields. Furthermore, the results in this paper indicate that the proposed deep learning technique can become a promising direction in experimental fluid mechanics.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Predicting wind-driven spatial deposition through simulated color images using deep autoencoders

Abstract For centuries, scientists have observed nature to understand the laws that govern the physical world. The traditional process of turning observations into physical understanding is slow. Imperfect models are constructed and tested to explain relationships in data. Powerful new algorithms can enable computers to learn physics by observing images and videos. Inspired by this idea, instead of training machine learning models using physical quantities, we used images, that is, pixel information. For this work, and as a proof of concept, the physics of interest are wind-driven spatial patterns. These phenomena include features in Aeolian dunes and volcanic ash deposition, wildfire smoke, and air pollution plumes. We use computer model simulations of spatial deposition patterns to approximate images from a hypothetical imaging device whose outputs are red, green, and blue (RGB) color images with channel values ranging from 0 to 255. In this paper, we explore deep convolutional neural network-based autoencoders to exploit relationships in wind-driven spatial patterns, which commonly occur in geosciences, and reduce their dimensionality. Reducing the data dimension size with an encoder enables training deep, fully connected neural network models linking geographic and meteorological scalar input quantities to the encoded space. Once this is achieved, full spatial patterns are reconstructed using the decoder. We demonstrate this approach on images of spatial deposition from a pollution source, where the encoder compresses the dimensionality to 0.02% of the original size, and the full predictive model performance on test data achieves a normalized root mean squared error of 8%, a figure of merit in space of 94% and a precision-recall area under the curve of 0.93.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2021

Investigations of water/oxide interfaces by molecular dynamics simulations

Water/oxide interfaces are ubiquitous on earth and show significant influence on many chemical processes. For example, understanding water and solute adsorption as well as catalytic water splitting can help build better fuel cells and solar cells to overcome our looming energy crisis; the interaction between biomolecules and water/oxide interfaces is one hypothesis to explain the origin of life. However, knowledge in this area is still limited due to the difficulty of studying water/solid interfaces. As a result, research using increasingly sophisticated experimental techniques and computational simulations has been carried out in recent years. Although it is difficult for experimental techniques to provide detailed microscopic structural information, molecular dynamics (MD) simulations have satisfactory performance. In this report, we discuss classical and ab initio MD simulations of water/oxide interfaces. Generally, we are interested in the following questions: How do solid surfaces perturb interfacial water structure? How do interfacial water molecules and adsorbed solutes affect solid surfaces and how do interfacial environments affect solvent and solute behavior? Finally, we discuss progress in the application of neural network potential based MD simulations, which offer a promising future because this approach has already enabled ab initio level accuracy for very large systems and long trajectories.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2024

Development of a Full-Scale Connected U-Net for Reflectivity Inpainting in Spaceborne Radar Blind Zones

CloudSat’s Cloud Profiling Radar is a valuable tool for remotely monitoring high-latitude snowfall, but its ability to observe hydrometeor activity near the Earth’s surface is limited by a radar blind zone caused by ground clutter contamination. This study presents the development of a deeply supervised U-Net-style convolutional neural network to predict cold season reflectivity profiles within the blind zone at two Arctic locations. The network learns to predict the presence and intensity of near-surface hydrometeors by coupling latent features encoded in blind zone-aloft clouds with additional context from collocated atmospheric state variables (i.e., temperature, specific humidity, and wind speed). Results show that the U-Net predictions outperform traditional linear extrapolation methods, with low mean absolute error, a 38% higher Sørensen–Dice coefficient, and vertical reflectivity distributions 60% closer to observed values. The U-Net is also able to detect the presence of near-surface cloud with a critical success index (CSI) of 72% and cases of shallow cumuliform snowfall and virga with 18% higher CSI values compared to linear methods. An explainability analysis shows that reflectivity information throughout the scene, especially at cloud edges and at the 1.2-km blind zone threshold, along with atmospheric state variables near the tropopause, are the most significant contributors to model skill. This surface-trained generative inpainting technique has the potential to enhance current and future remote sensing precipitation missions by providing a better understanding of the nonlinear relationship between blind zone reflectivity values and the surrounding atmospheric state.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2022

Learning Global Proliferation Expertise Evolution Using AI-Driven Analytics and Public Information

Detecting and anticipating global proliferation expertise and capability evolution from unstructured, noisy, and incomplete public data streams is a highly desired, but extremely challenging task. Here, in this article, we present our pioneering data-driven approach to support the non-proliferation mission to detect and explain the evolution of proliferation expertise and capability development globally from terabytes of publicly available information (PAI), focusing on our knowledge extraction pipeline and descriptive analytics. We first discuss how we fuse nine open-source data streams, including multilingual data, to convert 4 TB of unstructured data to structured knowledge and encode dynamically evolving proliferation expertise representations—content and context graphs. For this, we rely on natural language processing (NLP) and deep learning (DL) models to perform information extraction, topic modeling, and distributed text representation (aka embedding) learning. We then present interactive, usable, and explainable descriptive analytics to refine domain knowledge and present it in a human-understandable form. Finally, we introduce future work avenues that will leverage our dynamic knowledge representations and descriptive analytics to enable predictive and prescriptive inferences to achieve real-time domain understanding and contextual reasoning about global proliferation expertise and capability evolution.

98 NUCLEAR DISARMAMENT, SAFEGUARDS, AND PHYSICAL P↗
DOE OSTI2023

Data-driven predictions of the time remaining until critical global warming thresholds are reached

Leveraging artificial neural networks (ANNs) trained on climate model output, we use the spatial pattern of historical temperature observations to predict the time until critical global warming thresholds are reached. Although no observations are used during the training, validation, or testing, the ANNs accurately predict the timing of historical global warming from maps of historical annual temperature. The central estimate for the 1.5 °C global warming threshold is between 2033 and 2035, including a ±1σ range of 2028 to 2039 in the Intermediate (SSP2-4.5) climate forcing scenario, consistent with previous assessments. However, our data-driven approach also suggests a substantial probability of exceeding the 2 °C threshold even in the Low (SSP1-2.6) climate forcing scenario. While there are limitations to our approach, our results suggest a higher likelihood of reaching 2 °C in the Low scenario than indicated in some previous assessments—though the possibility that 2 °C could be avoided is not ruled out. Explainable AI methods reveal that the ANNs focus on particular geographic regions to predict the time until the global threshold is reached. Our framework provides a unique, data-driven approach for quantifying the signal of climate change in historical observations and for constraining the uncertainty in climate model projections. Given the substantial existing evidence of accelerating risks to natural and human systems at 1.5 °C and 2 °C, our results provide further evidence for high-impact climate change over the next three decades.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2022

Understanding the Physics Representation of Deep Learning Models in Environmental Applications

Deep learning (DL) models have been popular in earth and environmental modeling and analysis, which exhibit huge potential in capturing and reconstructing the non-linearity of relevant environmental processes. They are extensively used as analytical tools or emulators for multiple domains (atmosphere, land surface, ocean, and biogeochemistry). Despite their success, their internal working mechanism remains largely unknown. Such a lack of knowledge hinders the identification of physically consistent models that are fully adaptive to non-stationary climate, as well as the development of physics-informed machine learning such as physics-informed neural network (PINN). To establish preliminary knowledge and framework of such physics representation evaluation, this project focuses on an improved understanding of DL models in the environmental applications. DL models are increasingly applied to environmental modeling and prediction. However, they have been evaluated mostly from a performance perspective, and there is a gap in understanding how they represent the known physics internally. Such knowledge is especially critical when applying DL models under climate change conditions, where new inputs are likely outside the ranges of the training datasets. In this project, we reveal how the known physical processes are represented within DL models from both statistical and mechanistic perspectives. Leveraging the traditional model evaluations that focus more on the accuracies of predictions, we establish a framework that examines both the accuracy and physics representation of DL models. This analysis framework can identify DL models that make the correct predictions based on correct physics, thus enhancing the existing explainable artificial intelligence (explainable-AI) portfolio. It lays a foundation for developing novel metrics to evaluate the emerging DL models in environmental applications. This knowledge also informs the development of physics-informed DL models by revealing the direct connections between the known physical processes and specific model components or structures.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2022

Quantifying subsurface parameter and transport uncertainty using surrogate modelling and environmental tracers

Here, we combine physics-based groundwater reactive transport modelling with machine-learning techniques to quantify hydrogeological model and solute transport predictive uncertainties. We train an artificial neural network (ANN) on a dataset of groundwater hydraulic heads and 3 H concentrations generated using a high-fidelity groundwater reactive transport model. Using the trained ANN as a surrogate model to reproduce the input–output response of the high-fidelity reactive transport model, we quantify the posterior distributions of hydrogeological parameters and hydraulic forcing conditions using Markov chain Monte Carlo calibration against field observations of groundwater hydraulic heads and 3 H concentrations. We demonstrate the methodology with a model application that predicts Chlorofluorocarbon-12 (CFC-12) solute transport at a contaminated field site in Wyoming, United States. Our results show that including 3 H observations in the calibration dataset reduced the uncertainty in the estimated permeability field and infiltration rates, compared to calibration against hydraulic heads alone. However, predictive uncertainty quantification shows that CFC-12 transport predictions conditioned to the parameter posterior distributions cannot reproduce the field measurements. We found that calibrating the model to hydraulic head and 3 H observations results in groundwater mean ages that are too large to explain the observed CFC-12 concentrations. The coupling of the physics-based reactive transport model with the machine-learning surrogate model allows us to efficiently quantify model parameter and predictive uncertainties, which is typically computationally intractable using reactive transport models alone.

58 GEOSCIENCES↗
DOE OSTI2024

Characterizing climate pathways using feature importance on echo state networks

The 2022 National Defense Strategy of the United States listed climate change as a serious threat to national security. Climate intervention methods, such as stratospheric aerosol injection, have been proposed as mitigation strategies, but the downstream effects of such actions on a complex climate system are not well understood. The development of algorithmic techniques for quantifying relationships between source and impact variables related to a climate event (i.e., a climate pathway) would help inform policy decisions. Data-driven deep learning models have become powerful tools for modeling highly nonlinear relationships and may provide a route to characterize climate variable relationships. In this paper, we explore the use of an echo state network (ESN) for characterizing climate pathways. ESNs are a computationally efficient neural network variation designed for temporal data, and recent work proposes ESNs as a useful tool for forecasting spatiotemporal climate data. However, ESNs are noninterpretable black-box models along with other neural networks. The lack of model transparency poses a hurdle for understanding variable relationships. We address this issue by developing feature importance methods for ESNs in the context of spatiotemporal data to quantify variable relationships captured by the model. We conduct a simulation study to assess and compare the feature importance techniques, and we demonstrate the approach on reanalysis climate data. In the climate application, we consider a time period that includes the 1991 volcanic eruption of Mount Pinatubo. This event was a significant stratospheric aerosol injection, which acts as a proxy for an anthropogenic stratospheric aerosol injection. Furthermore, we are able to use the proposed approach to characterize relationships between pathway variables associated with this event that agree with relationships previously identified by climate scientists.

black-box models↗
DOE OSTI2022

Fraction of broad absorption line quasars in different radio morphologies

ABSTRACT In this study, we investigated the orientation model of Broad Absorption Line (BAL) quasars using a sample of sources that are common in Sloan Digital Sky Survey (SDSS) Data Release (DR)-16 quasar catalogue and Very Large Array (VLA)-Faint Images of the Radio Sky at Twenty Centimeters (FIRST) survey. Using the radio cut-out images from the FIRST survey, we first designed a deep-learning model using convolutional neural networks (CNN) to classify the quasar radio morphologies into the core-only, young jet, single lobe, or triples. These radio morphologies are further sub-classified into core-dominated and lobe-dominated sources. The CNN models can classify the sources with a high precision of >98 ${{\ \rm per\ cent}}$ for all the morphological sub-classes. The average BAL fraction in the resolved core, core-dominated, and lobe-dominated quasars are consistent with the BAL fraction inferred from radio and infrared surveys. We also present the distribution of BAL quasars as a function of quasar orientation by using the radio core-dominance as an orientation indicator. A similar analysis is performed for HiBALs, LoBALs, and FeLoBALs. All the radio morphological sub-classes and BAL sub-classes show an increase in BAL fraction at high orientation angles of the jets with respect to the line of sight. Our analysis suggests that BAL quasars are more likely to be found in viewing angles close to the equatorial plane of the quasar. However, a pure orientation model is inadequate, and a combination of orientation and evolution is probably the best way to explain the complete BAL phenomena.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2022

Decoding the protein–ligand interactions using parallel graph neural networks

Abstract Protein–ligand interactions (PLIs) are essential for biochemical functionality and their identification is crucial for estimating biophysical properties for rational therapeutic design. Currently, experimental characterization of these properties is the most accurate method, however, this is very time-consuming and labor-intensive. A number of computational methods have been developed in this context but most of the existing PLI prediction heavily depends on 2D protein sequence data. Here, we present a novel parallel graph neural network (GNN) to integrate knowledge representation and reasoning for PLI prediction to perform deep learning guided by expert knowledge and informed by 3D structural data. We develop two distinct GNN architectures: $$\hbox {GNN}_{\mathrm{F}}$$ GNN F is the base implementation that employs distinct featurization to enhance domain-awareness, while $$\hbox {GNN}_{\mathrm{P}}$$ GNN P is a novel implementation that can predict with no prior knowledge of the intermolecular interactions. The comprehensive evaluation demonstrated that GNN can successfully capture the binary interactions between ligand and protein’s 3D structure with 0.979 test accuracy for $$\hbox {GNN}_{\mathrm{F}}$$ GNN F and 0.958 for $$\hbox {GNN}_{\mathrm{P}}$$ GNN P for predicting activity of a protein–ligand complex. These models are further adapted for regression tasks to predict experimental binding affinities and $$\hbox {pIC}_{\mathrm{50}}$$ pIC 50 crucial for compound’s potency and efficacy. We achieve a Pearson correlation coefficient of 0.66 and 0.65 on experimental affinity and 0.50 and 0.51 on $$\hbox {pIC}_{\mathrm{50}}$$ pIC 50 with $$\hbox {GNN}_{\mathrm{F}}$$ GNN F and $$\hbox {GNN}_{\mathrm{P}}$$ GNN P , respectively, outperforming similar 2D sequence based models. Our method can serve as an interpretable and explainable artificial intelligence (AI) tool for predicted activity, potency, and biophysical properties of lead candidates. To this end, we show the utility of $$\hbox {GNN}_{\mathrm{P}}$$ GNN P on SARS-Cov-2 protein targets by screening a large compound library and comparing the prediction with the experimentally measured data.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2022

How Many Elements Matter?

Some studies of stars' multielement abundance distributions suggest at least 5–7 significant dimensions, but others show that many elemental abundances can be predicted to high accuracy from [Fe/H] and [Mg/Fe] (or [Fe/H] and age) alone. We show that both propositions can be, and are, simultaneously true. We adopt a machine-learning technique known as normalizing flow to reconstruct the probability distribution of Milky Way disk stars in the space of 15 elemental abundances measured by APOGEE. Conditioning on T eff and $\mathrm{log}\,g$ minimizes the differential systematics. After further conditioning on [Fe/H] and [Mg/Fe], the residual scatter for most abundances is σ [X/H] ≲ 0.02 dex, consistent with APOGEE's reported statistical uncertainties of ~0.01–0.015 dex and intrinsic scatter of 0.01–0.02 dex. Despite the small scatter, residual abundances display clear correlations between elements, which we show are too large to be explained by measurement uncertainties or by the finite sampling noise. We must condition on at least seven elements to reduce the correlations to a level consistent with the observational uncertainties. Our results demonstrate that cross-element correlations are a much more sensitive probe of a hidden structure than dispersion, and they can be measured precisely in a large sample even if the star-by-star measurement noise is comparable to the intrinsic scatter. We conclude that many elements have an independent story to tell, even for the mundane disk stars and elements produced by the core-collapse and Type Ia supernovae. The only way to learn these lessons is to measure the abundances directly, and not merely infer them.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2024

Thinking Bayesian for plasma physicists

Bayesian statistics offers a powerful technique for plasma physicists to infer knowledge from the heterogeneous data types encountered. To explain this power, a simple example, Gaussian Process Regression, and the application of Bayesian statistics to inverse problems are explained. The likelihood is the key distribution because it contains the data model, or theoretic predictions, of the desired quantities. By using prior knowledge, the distribution of the inferred quantities of interest based on the data given can be inferred. Because it is a distribution of inferred quantities given the data and not a single prediction, uncertainty quantification is a natural consequence of Bayesian statistics. The benefits of machine learning in developing surrogate models for solving inverse problems are discussed, as well as progress in quantitatively understanding the errors that such a model introduces.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2024

A Science Gateway for the Repeatable Analysis of Machine Learning Predicted Gravity Anomalies

In recent years, deep learning has become an increasingly popular alternative for modeling in geoscience applications due to its scalability and efficiency. However, the interpretability, compute, data volume, and hyperparameter tuning requirements of deep learning models make development and monitoring difficult. Furthermore, model explainability and communicating results obtained by these models to users or domain experts is a challenge, as domain experts in geoscience also need to have a deep understanding of how those models function in order to support their scientific works. Here, we describe a science gateway and machine learning pipeline for predicting gravity anomalies from geophysical data. The gateway, built on open-source technologies, provides a holistic view of the pipeline through interactive visualizations aimed at enabling efficient exploratory data analysis. The repeatability, reproducibility, and monitoring capabilities of this overall system allow us to iterate and analyze at scale. Using this pipeline and gateway, we can repeatedly produce accurate high-resolution gravity anomaly datasets. By describing the underlying technologies, implementation, and results, here we provide a foundation for the broader adoption of science gateways into cross-cutting geoscience and machine learning research projects as a means to improve the scientific discovery and collaboration in the geophysics and computational sciences community.

58 GEOSCIENCES↗
DOE OSTI2025

A comparison of probabilistic generative frameworks for molecular simulations

Generative artificial intelligence is now a widely used tool in molecular science. Despite the popularity of probabilistic generative models, numerical experiments benchmarking their performance on molecular data are lacking. Here, in this work, we introduce and explain several classes of generative models, broadly sorted into two categories: flow-based models and diffusion models. We select three representative models: neural spline flows, conditional flow matching, and denoising diffusion probabilistic models, and examine their accuracy, computational cost, and generation speed across datasets with tunable dimensionality, complexity, and modal asymmetry. Our findings are varied, with no one framework being the best for all purposes. In a nutshell, (i) neural spline flows do best at capturing mode asymmetry present in low-dimensional data, (ii) conditional flow matching outperforms other models for high-dimensional data with low complexity, and (iii) denoising diffusion probabilistic models appear the best for low-dimensional data with high complexity. Our datasets include a Gaussian mixture model and the dihedral torsion angle distribution of the Aib9 peptide, generated via a molecular dynamics simulation. We hope our taxonomy of probabilistic generative frameworks and numerical results may guide model selection for a wide range of molecular tasks.

Artificial intelligence↗
DOE OSTI2020

How Low Can You Go? Using Synthetic 3D Imagery to Drastically Reduce Real-World Training Data for Object Detection

Deep convolutional neural networks (DCNNs) currently provide state-of-the-art performance on image classification and object detection tasks, and there are many global security mission areas where such models could be extremely useful. Crucially, the success of these models is driven in large part by the widespread availability of high-quality open source data sets such as Image Net, Common Objects in Context (COCO), and KITTI, which contain millions of images with thousands of unique labels. However, global security relevant objects-of-interest can be difficult to obtain: relevant events are low frequency and high consequence; the content of relevant images is sensitive; and adversaries and proliferators seek to obscure their activities. For these cases where exemplar data is hard to come-by, even fine-tuning an existing model with available data can be effectively impossible. Recent work demonstrated that models can be trained using a combination of real-world and synthetic images generated from 3D representations; that such models can exceed the performance of models trained using real-world data alone; and that the generated images need not be perfectly realistic (Tremblay, et al., 2018). However, this approach still required hundreds to thousands of real-world images for training and fine tuning, which for sparse, global security-relevant datasets can be an unrealistic hurdle. In this research, we validate the performance and behavior of DCNN models as we drive the number of real-world images used for training object detection tasks down to a minimal set. We perform multiple experiments to identify the best approach to train DCNNs from an extremely small set of real-world images. In doing so, we: Develop state-of-the-art, parameterized 3D models based on real-world images and sample from their parameters to increase the variance in synthetic image training data; Use machine learning explainability techniques to highlight and correct through targeted training the biases that result from training using completely synthetic images; and Validate our results by comparing the performance of the models trained on synthetic data to one another, and to a control model created by fine-tuning an existing ImageNet-trained model with a limited number (hundreds) of real-world images.

97 MATHEMATICS AND COMPUTING↗