Physics-guided Neural Computational Methods

Physics-guided neural computational methods have been proposed to combine principles from physics and machine learning to enhance the predictive capabilities and generalizability of purely data-driven models. We develop methods for incorporating physics-based knowledge into the deep learning framework to provide the learning process with strong inductive biases, using examples such as wave scattering behaviors and resonance-driven mechanisms in metamaterials. Our study delivers the potential to address various scientific and engineering challenges that require real-time, accurate, interactive, and interpretable analyses.

Generative AI for Desired Properties

One of the significant advantages of neural computational methods lies in their synergistic utilization in the generative design of materials and engineering systems, departing from the discontinuous nature of conventional numerical approaches throughout the entire problem-solving pipeline. We develop deep learning and reinforcement learning methods tailored for design generation and recommendation to address multiple aspects of generated designs, including diversity, fidelity, generalizability, and manufacturability.

Super-resolved Sound Source Localization

Our works on super-resolved sound source localization pushes the boundaries of conventional physical resolution limits in auditory perception, enabling us to “see” and pinpoint the precise origin of sound with unprecedented accuracy. This capability finds applications in a multitude of fields, including acoustic source tracking, human-robot interaction, and immersive technologies of augmented reality (AR) and virtual reality (VR). Super-resolved sound source localization promises to revolutionize how we perceive the auditory world, opening doors to exciting possibilities.

Making the Invisible Visible

We also focus on the development of AI-driven methods aimed at rendering the invisible visible. This innovative approach empowers us to deploy advanced imaging methods that can unveil concealed and otherwise undetectable patterns. We develop novel approaches that expedite non-invasive imaging applications across diverse fields, such as optoacoustic-based biomedical imaging to visualize internal structures and functions of biological tissues without invasive procedures. Furthermore, we propose the methods for materials engineering, facilitating the recognition of hidden defect patterns and monitoring crack growth within structures.

Data-driven Advanced Fabrication

Data-driven advanced fabrication endeavors to optimize various process variables and conditions through machine learning techniques, ultimately leading to the attainment of desired outcomes in complex fabrication processes. Our research centers on the application of machine learning-driven approaches to unravel the intricate connections between process parameters and resulting properties. For instance, this includes delving into the molecular dynamics of ions to ensure the creation of precise deep RIE etching profiles and fine-tuning cooling conditions to achieve the targeted heat transfer distribution during the solidification of steel.