Date of Award
2024
Type
Thesis
Major
Master of Science
Degree Type
MASTER OF SCIENCE
Department
Earth & Space Science
First Advisor
Mohammad Jafari, Ph.D.
Second Advisor
Mahmut Reyhanoglu, Ph.D
Third Advisor
Abiye Seifu, Ph.D.
Abstract
This thesis presents a machine learning (ML) based approach to both model and estimate the parameters of an electric pump system that is used in liquid-propellent rocket engines. This is accomplished by utilizing Physics-Informed Neural Networks (PINNs) and inverse Physics-Informed Neural Networks (iPINNs). In an extreme system such as a liquidpropellent rocket engine, accurate control and efciency is imperative for successful performance. The electric pump system is inherently nonlinear and its dynamics are governed by ordinary diferential equations (ODEs). While entirely approachable, these nonlinear dynamics create a challenge in achieving precise control and accurate parameter estimations. Utilization of a PINN and iPINN is highly efective for complex/nonlinear systems such as this as they allow the combination of data-driven learning with physics-based constraints. PINNs and iPINNs are used to help with complex issues that stem from traditional numerical methods, such as computational cost and limited accuracy when data is limited. This study is divided into 2 main stages, the frst being the PINN and the second being the iPINN. To start, a PINN is trained to predict the system’s behavior based on established known parameters. That trained system is then used in the iPINN to estimate the system’s parameters. During the iPINNs training, the parameters are constantly adjusted ensuring that they respect the constraints set by the systems ODE. The thesis concludes with a comparative analysis of the fnal parameters learned by the iPINN and the established known/target values. The iPINN ended up achieving parameter estimates that are within a 1% margin of error when compared to the known values. This ultimately confrms that the created iPINN is capable of learning and subsequently predicting the system’s dynamics. The obtained results demonstrate the power and dependency of PINNs and iPINNs for system modeling and parameter estimation while also laying the foundation for further application in control systems.
Recommended Citation
McFarland, Logan, "Physics-Informed Neural Networks (PINNS) and Inverse PINNS for the Modeling and Parameter Estimation of Electric Pumps in Liquid-Propellant Rocket Engines" (2024). Theses and Dissertations. 549.
https://csuepress.columbusstate.edu/theses_dissertations/549
