Yijin Wang L&S Math & Physical Sciences
Adaptive Strategies for PINNs: Addressing Spectral Bias in Fluids
This project aims to enhance the performance of Physics-Informed Neural Networks (PINNs) in simulating fluid dynamics, specifically addressing the challenge of spectral bias which can skew simulation accuracy. Our approach involves developing adaptive training strategies that empower PINNs to more precisely model a diverse range of fluid behaviors, from steady, predictable patterns to dynamic, turbulent flows. By integrating principles from the neural tangent kernel theory, we are meticulously fine-tuning the training processes of these networks. This involves recalibrating how PINNs learn from data, focusing on a balanced assimilation of both low- and high-frequency phenomena. The primary goal is to improve the precision of these simulations, thereby enhancing their application in scientific and engineering contexts. This refinement could lead to more reliable predictive models, supporting advanced research and practical applications in fluid dynamics.
Message To Sponsor
I'm sincerely grateful for the opportunity to engage with this project and extend my understanding of Physics-Informed Neural Networks and fluid dynamics. Your support fuels my passion for discovery and innovation in a field that fascinates me deeply. Thank you for being a pillar of encouragement in my academic and research endeavors.