LEVERAGING PHYSICS-INFORMED NEURAL NETWORKS FOR BIFURCATION ANALYSIS IN THERMO-FLUID CONVECTION: PREDICTING MULTICELLULAR FLOW IN HIGH-ASPECT RATIO CAVITIES

Abstract

Thermo-fluid convection is a phenomenon observed across various spatial scales, from large-scale ocean currents to small, enclosed cavities, such as those found in insulating glazing units (IGUs). In these systems, convective flows undergo a transition from laminar to turbulent regimes as the Rayleigh number varies, which has significant implications for heat transfer. In high-aspect ratio cavities, such as IGUs, this transition manifests as a bifurcation from a unicellular to a multicellular flow pattern within the laminar regime, thereby increasing the overall heat transfer. The Rayleigh number is a dimensionless quantity that measures the influence of buoyancy forces relative to viscous forces in a fluid. In IGUs, as the Rayleigh number increases, the convective flow changes from a simple unicellular pattern to more complex multicellular patterns. This bifurcation is important because it marks a transition that affects the efficiency of the heat transfer process. To study this complex phenomenon, traditional computational methods like Computational Fluid Dynamics (CFD) are commonly used. However, CFD simulations can be computationally expensive and time-consuming, especially when trying to model a range of parameters like different Rayleigh numbers. This thesis investigates the bifurcation response of convective flows in high-aspect ratio cavities using a Physics-Informed Neural Network (PINN) framework. The PINN leverages the Navier-Stokes equations to model the flow dynamics, while its loss function integrates both data loss, derived from limited Computational Fluid Dynamics (CFD) snapshot data, and physics loss, which includes governing equations and boundary conditions. Despite the constraint of limited data, the surrogate PINN model demonstrates the capability to accurately predict the multicellular bifurcation behavior. This work highlights the potential of PINNs in developing computationally efficient surrogate models for exploring the bifurcation regime in thermo-fluid convection as key control parameters, such as the Rayleigh number, are varied. The findings provide valuable insights into optimizing heat transfer in IGUs and other high-aspect ratio cavities, contributing to advancements in energy-efficient building technologies