Georgia Institute of Technology
Development Of Physics-Informed Machine Learning Framework with Gaussian Process Regression and Its Application to Paper Production Systems
Abstract
dc:description.abstractThree distinct methodologies are proposed for integrating physics-informed insights into Gaussian Process Regression (GPR). These approaches aim to enhance crucial components of GPR: hyperparameter estimation, the mean function, and the kernel function. They are designed to tackle common challenges in conventional Machine Learning (ML) methods, such as overfitting and limited predictive performance, which are particularly evident in scenarios with sparse or incomplete data. The primary objective is to surpass the ‘black box’ nature of standard ML models, infusing them with physics-based knowledge, thereby elevating predictive accuracy in complex systems. The first strategy introduces a physics-based penalization method for hyperparameter estimation in GPR, aimed at enhancing prediction accuracy and reducing the risk of overfitting. The second strategy focuses on the enhancement of the GPR mean function through the integration of Physics-informed Neural Networks (PINNs), ensuring a more accurate alignment with the system’s underlying physical principles. The final strategy proposes a new physics-infused kernel for GPR, aimed at quantifying and addressing deviations from physical constraints to refine the model's covariance relationship. Collectively, these methodologies demonstrate significant advancements in the capability of GPR to model and represent the dynamics of complex systems. Physics-informed machine learning strategies are also applied to the paper production system, demonstrating improved performance and uncovering true dynamics that conventional data-driven methods previously failed to detect. This integration of domain-specific knowledge with data-driven predictions offers a novel approach to advancing machine learning in scientific and engineering domains.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Chemical and Biomolecular Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kim, Jinhyeun
- Advisors dc:contributor.advisor
-
- Boukouvala, Fani
- Paynabar, Kamran
- Committee members dc:contributor.committeemember
-
- Luettgen, Christopher
- Realff, Matthew
- Joseph, Roshan
- Aidun, Cyrus
Subjects
dc:subject × 3Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/76832