Defense Notices

Artificial Intelligence (AI) is being widely used in diverse domains such as industrial automation, traffic control, precision agriculture, and smart cities for major heavy lifting in terms of data analysis and decision making. However, the AI life- cycle is a major source of greenhouse gas (GHG) emission leading to devastating environmental impact. This is due to expensive neural architecture searches, training of countless number of models per day across the world, in-field AI processing of data in billions of edge devices, and advanced security measures across the AI life cycle. Modern applications often involve multitasking, which involves performing a variety of analyzes on the same dataset. These tasks are usually executed on resource-limited edge devices, necessitating AI models that exhibit efficiency across various measures such as power consumption, frame rate, and model size. To address these challenges, we introduce a novel neural network architecture model that incorporates a layer sharing principle to optimize the power usage. We propose a novel neural architecture, Layer Shared Neural Networks that merges multiple similar AI/NN tasks together (with shared layers) towards creating a single AI/NN model with reduced energy requirements and carbon footprint. The experimental findings reveal competitive accuracy and reduced power consumption. The layer shared model significantly reduces power consumption by 50% during training and 59.10% during inference causing as much as an 84.64% and 87.10% decrease in CO2 emissions respectively. 

The accurate prediction of small molecule binding affinity and toxicity remains a central challenge in drug discovery, with significant implications for reducing development costs, improving candidate prioritization, and enhancing safety profiles. Traditional computational approaches, such as molecular docking and quantitative structure-activity relationship (QSAR) models, often rely on handcrafted features and require extensive domain knowledge, which can limit scalability and generalization to novel chemical scaffolds. Recent advances in language models (LMs), particularly those adapted to chemical representations such as SMILES (Simplified Molecular Input Line Entry System), have opened new ways for learning data-driven molecular representations that capture complex structural and functional properties. However, achieving both high binding affinity and low toxicity through a resource-efficient computational pipeline is inherently difficult due to the multi-objective nature of the task. This study presents a novel dual-paradigm approach to critical challenges in drug discovery: predicting small molecules with high binding affinity and low cardiotoxicity profiles. For binding affinity prediction, we implement a specialized graph neural network (GNN) architecture that operates directly on molecular structures represented as graphs, where atoms serve as nodes and bonds as edges. This topology-aware approach enables the model to capture complex spatial arrangements and electronic interactions critical for protein-ligand binding. For toxicity prediction, we leverage chemical language models (CLMs) fine-tuned with Low-Rank Adaptation (LoRA), allowing efficient adaptation of large pre-trained models to specialized toxicological endpoints while maintaining the generalized chemical knowledge embedded in the base model. Our hybrid methodology demonstrates significant improvements over existing computational approaches, with the GNN component achieving an average area under the ROC curve (AUROC) of 0.92 on three protein targets and the LoRA-adapted CLM reaching (AUROC) of 0.90 with 60% reduction in parameter usage in predicting cardiotoxicity. This work establishes a powerful computational framework that accelerates drug discovery by enabling both higher binding affinity and low toxicity compounds with optimized efficacy and safety profiles. 

Compiler optimizations are critical for performance and have been extensively studied, especially for C/C++ language compilers. Our overall goal in this thesis is to investigate and compare the properties and behavior of optimization passes across multiple contemporary, state-of-the-art (SOTA)  C/C++ compilers to understand if they adopt similar optimization implementation and orchestration strategies. Given the maturity of pre-existing knowledge in the field, it seems conceivable that different compiler teams will adopt consistent optimization passes, pipeline and application techniques. However, our preliminary results indicate that such expectation may be misguided. If so, then we will attempt to understand the differences, and study and quantify their impact on the performance of generated code.

In our first work, we study and compare the behavior of profile-guided optimizations (PGO) in two popular SOTA C/C++ compilers, GCC and Clang. This study reveals many interesting, and several counter-intuitive, properties about PGOs in C/C++ compilers. The behavior and benefits of PGOs also vary significantly across our selected compilers. We present our observations, along with plans to further explore these inconsistencies in this report. Likewise, we have also measured noticeable differences in the performance delivered by optimizations across our compilers. We propose to explore and understand these differences in this work. We present further details regarding our proposed directions and planned experiments in this report. We hope that this work will show and suggest opportunities for compilers to learn from each other and motivate researchers to find mechanisms to combine the benefits of multiple compilers to deliver higher overall program performance.

Stimming refers to repetitive actions or behaviors used to regulate sensory input or express feelings. Children with developmental disorders like autism (ASD) frequently perform stimming. This includes arm flapping, head banging, finger flicking, spinning, etc. This is exhibited by 80-90% of children with Autism, which is seen in 1 among 36 children in the US. Head banging is one of these self-stimulatory habits that can be harmful. If these behaviors are automatically identified and notified using live video monitoring, parents and other caregivers can better watch over and assist children with ASD.
Classifying these actions is important to recognize harmful stimming, so this study focuses on developing a deep learning-based approach for stimming action recognition. We implemented and evaluated four models leveraging three deep learning architectures based on Convolutional Neural Networks (CNNs), Autoencoders, and Vision Transformers. For the first time in this area, we use skeletal joints extracted from video sequences. Previous works relied solely on raw RGB videos, vulnerable to lighting and environmental changes. This research explores Deep Learning based skeletal action recognition and data processing techniques for a small unstructured dataset that consists of 89 home recorded videos collected from publicly available sources like YouTube. Our robust data cleaning and pre-processing techniques helped the integration of skeletal data in stimming action recognition, which performed better than state-of-the-art with a classification accuracy of up to 87%
In addition to using traditional deep learning models like CNNs for action recognition, this study is among the first to apply data-hungry models like Vision Transformers (ViTs) and Autoencoders for stimming action recognition on the dataset. The results prove that using skeletal data reduces the processing time and significantly improves action recognition, promising a real-time approach for video monitoring applications. This research advances the development of automated systems that can assist caregivers in more efficiently tracking stimming activities.

Just as the internet-of-things leverages machine learning over a vast amount of data produced by an innumerable number of sensors, the Internet of Catalysis program uses similar strategies with catalysis research. One application of the Internet of Catalysis strategy is treating research papers as datapoints, rich with text, figures, and tables. Prior research within the program focused on machine learning models applied strictly over text.

This project is the first step of the program in creating a machine learning model from the images of catalysis research papers. Specifically, this project creates a dataset of faradaic efficiency graphs using transfer learning from pretrained models. The project utilizes FasterRCNN_ResNet50_FPN, LayoutLMv3SequenceClassification, and computer vision techniques to recognize figures, extract all graphs, then classify the faradaic efficiency graphs.

Downstream of this project, researchers will create a graph reading model to integrate with large language models. This could potentially lead to a multimodal model capable of fully learning from images, tables, and texts of catalysis research papers. Such a model could then guide experimentation on reaction conditions, catalysts, and production.

Significant advances in information and networking technologies have transformed Cyber-Physical Systems (CPS) into networked cyber-physical systems (NCPS). A noteworthy example of such systems is smart grid networks, which include distributed energy resources (DERs), renewable generation, and the widespread adoption of Electric Vehicles (EVs). Such complex NCPS require intelligent and autonomous control solutions. For example, the increasing number of EVs introduces significant sources of demand and user behavior uncertainty that can jeopardize grid stability during peak hours. Traditional model-based demand-supply controls fail to accurately model and capture the complex nature of smart grid systems in the presence of different uncertainties and as the system size grows. To address these challenges, data-driven approaches have emerged as an effective solution for informed decision-making, predictive modeling, and adaptive control to enhance the resiliency of NCPS in uncertain environments.

As a powerful data-driven approach, Multi-Agent Reinforcement Learning (MARL) enables agents to learn and adapt in dynamic and uncertain environments. However, MARL techniques introduce complexities related to communication, coordination, and synchronization among agents. In this PhD research, we investigate autonomous control for smart grid decision networks using MARL. First, we examine the issue of imperfect state information, which frequently arises due to the inherent uncertainties and limitations in observing the system state.

Second, we focus on the cooperative behavior of agents in distributed MARL frameworks, particularly under the central training with decentralized execution (CTDE) paradigm. We provide theoretical results and variance analysis for stochastic and deterministic cooperative MARL algorithms, including Multi-Agent Deep Deterministic Policy Gradient (MADDPG), Multi-Agent Proximal Policy Optimization (MAPPO), and Dueling MAPPO. These analyses highlight how coordinated learning can improve system-wide decision-making in uncertain and dynamic environments like EV networks.

Third, we address the scalability challenge in large-scale NCPS by introducing a hierarchical MARL framework based on a cluster-based architecture. This framework organizes agents into coordinated subgroups, improving scalability while preserving local coordination. We conduct a detailed variance analysis of this approach to demonstrate its effectiveness in reducing communication overhead and learning complexity. This analysis establishes a theoretical foundation for scalable and efficient control in large-scale smart grid applications.

This project presents the development of a custom Convolutional Neural Network (CNN) designed to classify object states—such as sliced, diced, or peeled—in cooking environments. Recognizing fine-grained object states is essential for context-aware manipulation but remains challenging due to visual similarity between states and a limited dataset. To address these challenges, I built a lightweight CNN from scratch, deliberately avoiding pretrained models to maintain domain specificity and efficiency. The model was enhanced through data augmentation and optimized dropout layers, with additional experiments incorporating batch normalization, Inception modules, and residual connections. While these advanced techniques offered incremental improvements during experimentation, the final model—a combination of data augmentation, dropout, and batch normalization—achieved ~60% validation accuracy and demonstrated stable generalization. This work highlights the trade-offs between model complexity and performance in constrained environments and contributes toward real-time state recognition with potential applications in assistive technologies.

The growing dependence on electronic systems in consumer and mission critical domains requires engineers who understand the inner workings of digital hardware. Yet many students bypass hardware electives, viewing them as abstract, mathematics heavy, and less attractive than software courses. Escalating workforce shortages in the semiconductor industry and the recent global chip‑supply crisis highlight the urgent need for graduates who can bridge hardware knowledge gaps across engineering sectors. In this thesis, I have developed FPGA‑based games, embedded in inclusive curricular modules, which can make hardware concepts accessible while fostering interest, self‑efficacy, and positive outcome expectations in hardware engineering. A design‑based research methodology guided three implementation cycles: a pilot with seven diverse high‑school learners, a multiweek residential summer camp with high‑school students, and a fifteen‑week multidisciplinary elective enrolling early undergraduate engineering students. The learning experiences targeted binary arithmetic, combinational and sequential logic, state‑machine design, and hardware‑software co‑design. Learners also moved through the full digital‑design flow, HDL coding, functional simulation, synthesis, place‑and‑route, and on‑board verification. In addition, learners explored timing analysis, register‑transfer‑level abstractions, and simple processor datapaths to connect low‑level circuits with system‑level behavior. Mixed‑method evidence was gathered through pre‑ and post‑content quizzes, validated surveys of self‑efficacy and outcome expectations, focus groups, classroom observations, and gameplay analytics. Paired‑sample statistics showed reliable gains in hardware‑concept mastery, self‑efficacy, and outcome expectations. This work contributes a replicable framework for translating foundational hardware topics into modular, game‑based learning activities, empirical evidence of their effectiveness across secondary and early‑college contexts, and design principles for educators who seek to integrate equitable, hands‑on hardware experiences into existing curricula.

Although the Center for Remote Sensing of Ice Sheets (CReSIS) performs several radar calibration steps to produce Operation IceBridge (OIB) radar depth sounder data products, these datasets are not radiometrically calibrated and the swath array processing uses ideal (rather than measured [calibrated]) steering vectors. Any errors in the steering vectors, which describe the response of the radar as a function of arrival angle, will lead to errors in positioning and backscatter that subsequently affect estimates of basal conditions, ice thickness, and radar attenuation. Scientific applications that estimate physical characteristics of surface and subsurface targets from the backscatter are limited with the current data because it is not absolutely calibrated. Moreover, changes in instrument hardware and processing methods for OIB over the last decade affect the quality of inter-seasonal comparisons. Recent methods which interpret basal conditions and calculate radar attenuation using CReSIS OIB 2D radar depth sounder echograms are forced to use relative scattering power, rather than absolute methods.

As an active target calibration is not possible for past field seasons, a method that uses natural targets will be developed. Unsaturated natural target returns from smooth sea-ice leads or lakes are imaged in many datasets and have known scattering responses. The proposed method forms a system of linear equations with the recorded scattering signatures from these known targets, scattering signatures from crossing flight paths, and the radiometric correction terms. A least squares solution to optimize the radiometric correction terms is calculated, which minimizes the error function representing the mismatch in expected and measured scattering. The new correction terms will be used to correct the remaining mission data. The radar depth sounder data from all OIB campaigns can be reprocessed to produce absolutely calibrated echograms for the Arctic and Antarctic. A software simulator will be developed to study calibration errors and verify the calibration software. The software for processing natural targets and crossovers will be made available in CReSIS’s open-source polar radar software toolbox. The OIB data will be reprocessed with new calibration terms, providing to the data user community a complete set of radiometrically calibrated radar echograms for the CReSIS OIB radar depth sounder for the first time.

With the increasing congestion of the usable RF spectrum, it is increasingly necessary for communication and radar systems to share the same frequencies without disturbing one another. To accomplish this, research has focused on designing a class of non-repeating radar waveforms that appear as noise at the receiver of uncooperative systems, but the peak power from high-power pulsed systems can still overwhelm nearby in-band systems. Therefore, to minimize peak power while maximizing the total energy on target, radar systems must transition to operating at a 100% duty cycle, which inherently requires Simultaneous Transmit and Receive (STAR) operation.

One inherent difficulty when operating monostatic STAR systems is the direct path coupling interference that can saturate a number of components in the radar’s receive chain, which makes digital processing methods that remove this interference ineffective. This thesis proposes a method to reduce the self-interference between the radar’s transmitter in receiver prior to the receiver’s sensitive components to increase the power that the radar can transmit at. By using a combination of tests that manipulate the timing, phase, and magnitude of a secondary waveform that is injected into the radar just before the receiver, upwards of 35.0 dB of self-interference cancellation is achieved for radar waveforms with bandwidths of up to 100 MHz at both S-band and X-band in both simulation and open-air testing.