Session: K8-01: FUNDAMENTALS OF MACHINE LEARNING IN HEAT TRANSFER

Paper Number: 131314

131314 - Inferring Flow Boiling Interfacial Shear Stress Using Physics Informed Neural Networks From Control Volume Models 

Abstract: 

Flow boiling has been identified as a promising method for heat acquisition in thermal management technologies. Common issues preventing flow boiling technology from being widely used are variable heat transfer characteristics as well as large pressure drops along the flow boiling channels. These complications arise from the phase change phenomenon and can be rather difficult to predict and characterize. A common approach to predict these phenomena are semi-empirical models, due to their simplicity and basis in well-established physics. These types of models employ a control volume approach where the conservation of mass, momentum, and energy are applied to solve for average quantities like mean vapor and liquid velocities. They are promising as they directly embed physics into solutions while also utilizing correlation-based methods to directly compute complex flow quantities like shear stress. These methods also act as closure models to make the system well-posed and computable. As a result of utilizing correlations to fill in for hard to measure quantities, semi-empirical models act as a prelude to data driven methods in physical modelling. Physics informed neural networks (PINNs) are a recent advancement in the data-driven science and engineering field that also accurately models physics problems using data-driven techniques. In this paper we replace the semi-empirically derived interfacial shear stress term in the momentum equations, which often underpredicts the interfacial shear stress, with an interfacial shear term informed by a PINN. We utilize Bayesian methods to interpolate between known test cases to expand a relatively small data set. Then we train a PINN to learn the void fraction, pressure, and interfacial shear stress solution to a set of semi-empirical equations for a separated flow boiling model. In general, we find that the PINN is better at fitting the interfacial shear stress term than well-known correlations, with the added benefit of directly being able to include data into the modelling process. This paper demonstrates the effectiveness of data-driven solutions to differential equations for inferring complex flow quantities in underdetermined systems, like these semi-empirical models. Furthermore, applications of PINNs to flow boiling systems will aid in the advancement of thermal management technologies by improving model reliability and incorporation of experimental data.

Presenting Author: Logan Pirnstill Case Western Reserve University

Presenting Author Biography: Logan Pirnstill is a 3rd year graduate student and PhD candidate at Case Wester Reserve University. His research interests include numerical analyis, machine learning, physics informed machine learning, two-phase flows and thermal management, and numerical heat transfer.

Authors:

Logan Pirnstill Case Western Reserve University
Chirag Kharangate Case Western Reserve University