DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
M.Tech Research Thesis Defense
Speaker : Ms. Diya Nag Chaudhury
S.R. Number : 06-18-01-10-22-23-1-23885
Title : “A physics-informed framework for super-resolution of fluid flows”
Thesis examiner : Dr. Ganesh Kiran Vaidya, Dept. of Mathematics, IISc.
Research Supervisor : Prof. Sashikumaar Ganesan
Date & Time : June 03, 2026 (Wednesday) at 11:30 AM
Venue : # 102 CDS Seminar Hall
ABSTRACT
The reconstruction of high-resolution flow fields from low-resolution data remains a persistent challenge within fluid dynamics. Super-resolution using deep learning is a valuable tool for enhancing the visualization of flow fields, allowing for the recovery of detailed features from low-resolution data. This work introduces the Bicubic Fourier Neural Operator (Bicubic FNO), which combines the benefits of both bicubic interpolation and FNO. We compare our architecture against standard FNO and two convolutional neural network (CNN) models reported in the literature for super-resolution: Super-resolution Convolutional Neural Network (SRCNN), widely used for natural images, and the Downsampled Skip-Connection/Multi-Scale (DSC/MS) model, designed for reconstructing high-resolution turbulent flow fields.
We tested our model on two-dimensional decaying homogeneous isotropic turbulence (HIT), wall-bounded channel flow, and the Stanford flame dataset. Our results demonstrate a 10.47% improvement over the standard FNO and a 19.59% improvement over the DSC/MS model in terms of Peak Signal-to-Noise Ratio (PSNR) for 2D HIT. We achieved a 63% reduction in Mean Squared Error (MSE) compared to pure bicubic interpolation, confirming the FNO’s ability to recover fine-scale structures.
We extend our research to integrate physical principles into the framework, ensuring that the model can reconstruct physically consistent flow fields even with sparse data. Incorporating this domain knowledge into the network provides an additional 14% improvement in MSE, indicating that physical knowledge effectively complements data-driven learning.
We find that physics constraints, when carefully formulated and weighted, can significantly enhance learned models. To ensure physical consistency, we enforce a divergence-free constraint that guarantees local mass conservation, a principle fundamental to incompressible flow physics. We then transition to a self-supervised approach, which opens new possibilities for applications where generating high-fidelity training data is computationally prohibitive, such as direct numerical simulations (DNS) at high Reynolds numbers. Finally, we employ Explainable AI (XAI) to examine how the model learns physical structures, respects conservation laws, and processes turbulence phenomena.
This research highlights the efficacy of the Bicubic FNO and its extension with physical constraints, demonstrating that we can extract meaningful turbulent features while significantly reducing the reliance on massive datasets.
ALL ARE WELCOME
