Research Projects
A core theme unifying our research is the exploration of the powerful synergy between AI and Uncertainty Quantification (UQ), aiming to build next-generation intelligent systems that understand their own limitations and provide reliable insights for safer infrastructure and communities, featuring both fundamental research and applied industry collaborations.
Efficient and Robust Infrastructure Defect Detection with AI and Foundational Models
Working in close collaboration with our industrial partners RaSpect, we are developing AI and foundation model driven apprach to assess infrastructure defects from drone captured images. The methodology developed represents a significant improvement in applicability across a diverse range of communities and robustness over annotation noises, enabling foundational vision models (e.g., Segment Anything Model) to specialize in crack segmentation with minimal cost.
Neural-parameterized, GPU-accelerated Tropical Cyclone and Storm Surge Modeling
Tropical cyclones (TC) have caused significant financial losses and jeopardized the lives of millions of people who live along the world's coastlines. With rising coastal population and climate change, the coastal community worldwide is becoming more vulnerable to TCs. We aim to develop a family of innovative techniques to rapidly and accurately account for uncertainties into the TC storm surge risk, empowered by neural-parameterized techniques that could be accelerated on state-of-the-art GPU computing platforms.
Interaction between Uncertainty Quantification and AI
We explore how UQ methods can enhance AI models by rigorously assessing the confidence in their predictions and improving their robustness, particularly when dealing with noisy, incomplete, or novel data. Concurrently, we leverage powerful AI techniques, such as deep learning and surrogate modeling, to tackle the computational challenges often associated with traditional UQ methods, making complex uncertainty analysis more feasible. Ultimately, this project aims to develop next-generation AI systems that are not only accurate but also reliable and trustworthy, capable of quantifying their own uncertainties for safer and more informed decision-making in complex real-world applications.
Physics-informed Machine Learning for Wave Energy Modeling
This research focuses on developing and applying Physics-informed Machine Learning (PiML) techniques to improve the modeling of wave energy systems. By embedding fundamental physical laws directly into the machine learning model's architecture and training process, we aim to overcome the limitations of both traditional, computationally expensive numerical simulations and purely data-driven approaches which may lack physical consistency or require vast datasets. The goal is to create accurate, computationally efficient, and physically consistent surrogate models for predicting wave interactions with energy converters, power takeoff dynamics, and overall energy yield, ultimately accelerating the design, optimization, and control of wave energy technologies.
For a list of recent research ouputs, visit our publications page.