Boiling is a highly efficient heat transfer mechanism central to various technologies and industries.
A critical issue in boiling, known as the boiling crisis (Critical Heat Flux - CHF), occurs when bubbles accumulate and block heat transfer, leading to dangerous overheating.
This phenomenon can cause catastrophic events like nuclear reactor meltdowns, highlighting the importance of accurately detecting CHF.
Current computer vision models detect CHF effectively within the same domain (e.g., similar experimental setups).
However, these models fail when applied to images from different domains.
My PhD research focuses on solving the domain shift problem to improve CHF detection across diverse scenarios.
My dissertation proposes a 4-phases solution to address this:
Phase I: We introduce a novel cross-domain classification framework to adapt a pre-trained classifier to new domains without any additional annotation effort. To the best of our knowledge, this work is the first of its kind to address the domain shift in computer vision CHF for detection.
Phase II: We introduce DIPS: a novel UI2I evaluation metric designed specifically for the task of cross-domain classification. The metric outperformed SOTA metrics in the task of cross-domain classification.
Phase III: We introduce SequenceSync-GAN: novel UI2I model aiming at temporal sequential consistency preservation during translation. The model outperformed SOTA models for the task of CHF cross-domain classification.
Phase IV: We introduce BubbleSync-GAN: a novel UI2I model aiming at physical properties consistency preservation during translation. The model outperformed SOTA models for the task of CHF cross-domain classification.
We present the Smart Driving System, a pioneering framework leveraging multi-modal sensing (MMS) to evaluate driving behavior for the early detection of Alzheimer's disease and related dementias (AD/ADRD). This project is a collaborative effort between several ASU labs from the School of Electrical, Computer, and Energy Engineering, the School of Engineering for Matter, Transport, and Energy, and the School of Computing and Augmented Intelligence at Arizona State University (ASU). It also involves key contributions from the ASU-Mayo Clinic Medical Devices and Methods Laboratory, the Center for Bioelectronics and Biosensors at ASU's Biodesign Institute, and the Barrow Neurological Institute's Healthy Brain Clinic in Phoenix, Arizona. Additionally, TF Health Corporation, operating as Breezing Co., plays a significant role in this partnership.
Advanced Driving Performance Metrics: We introduce new features that assess cognitive decision-making based on driving actions such as turns, sudden braking, and path deviations.
Metabolic Rate Integration: By incorporating the driver's metabolic rate, we estimate energy expenditure as a potential indicator of brain activity and age-related cognitive decline.
Environmental Contextualization: The system accounts for environmental factors, such as passenger count and overall comfort, to enhance the cognitive assessment.
AI-Powered Cognitive Differentiation: We combine these engineered features using machine learning to differentiate between healthy cognition and mild cognitive impairment (MCI). This lays the foundation for an AI-driven solution aimed at early detection of cognitive decline.
Innovative Application: To our knowledge, this is the first reported effort towards the development of a Smart Driving Device and App. Our goal is to integrate these into vehicles and validate their effectiveness in detecting MCI through comprehensive pilot studies.
The PIRE project aims to tackle the challenges of operating District Heating and Power (DHP) systems in a way that is efficient, human-centered, and socially justifiable to combat climate change with a focus on energy equity and resilience. This project aims at integrating AI techniques with an understanding of human needs and behaviors to enable an efficient, human-centered, resilient, and socially justifiable operation of district- and community-scale heat pumps systems that promote and support regional scale adoption of building decarbonization. This PIRE brings together a diverse, yet highly interdisciplinary and synergistic team led by Texas A&M University (TAMU), with Drexel University (DU), Arizona State University (ASU), Temple University (TU), KTH Royal Institute of Technology (KTH) and Blekinge Institute of Technology (BTH) in Sweden, and Aalborg University (AAU) in Denmark as partner institutions (Figure 2).
Interdisciplinary Collaboration: Assemble a global team of engineers, data scientists, urban planners, and social scientists to explore how DHP systems can reduce building carbon footprints, enhance climate resilience, and promote energy equity.
AI Advancement: Develop scalable AI solutions that improve DHP operation efficiency by leveraging data-driven energy diagnostics, community-informed data modeling, and adaptive controls, while also advancing our understanding of human interaction with the built environment.
Human-Centric Operation Strategies: Create innovative, human-centered DHP operation strategies that align with community needs and support broader environmental, economic, and social goals.
Community Engagement: Engage government and community stakeholders in guiding AI modeling to ensure it addresses energy injustices and supports the transition to low-carbon energy.
Education and Knowledge Sharing: Establish a platform for shared knowledge, education, and resources across U.S. and international partners to prepare the next generation of professionals with diverse backgrounds and expertise in DHP systems.
Biomedical datasets, particularly those used for disease detection, often contain inherent heterogeneity, not only within disease cohorts but surprisingly also within supposedly homogeneous healthy cohorts due to inter-subject variability. To address this challenge, our project proposes the concept of a **"Healthy Bio-Core (HBC)"**, a novel framework designed to identify and select a homogeneous subset of healthy multivariate time series (MTS). By leveraging dynamic time warping (DTW), Gaussian distribution modeling, and ROCKET classification accuracy, the HBC method systematically extracts representative healthy instances, thereby enhancing the discriminative power of subsequent classification models. Our empirical analysis demonstrates that this targeted selection of homogeneous healthy subjects significantly improves classification performance across multiple biomedical datasets, emphasizing the potential of the HBC approach for improved precision in biomedical diagnostics and predictive analytics.