Virginia Tech
Scientific Machine Learning for Engineering Systems: Optimal Control, Prognostics, and Uncertainty Quantification
Abstract
dc:description.abstractThis dissertation investigates how scientific machine learning can support engineering systems across two complementary tasks: optimal control and uncertainty-aware prognostics. In the control setting, the objective is to compute actions that steer governed dynamics toward desired outcomes while respecting physical laws and optimality conditions. In the prognostic setting, the objective is to forecast degradation-related quantities, estimate remaining useful life, and communicate predictive uncertainty in a form that supports maintenance decisions. The central argument of the dissertation is that machine learning becomes more useful for engineering when it is organized around the structure of the decision problem itself. The dissertation is composed of four studies. The first study introduces Control Physics-Informed Neural Networks for open-loop PDE-constrained optimization. The method embeds forward dynam- ics, adjoint equations, and first-order optimality conditions into a one-stage learning framework that jointly predicts state and control. The second study extends the control line of work to closed-loop, infinite-horizon feedback control by learning value functions from the Hamilton-Jacobi-Bellman equation and recovering policies from value gradients, with ensemble strategies used to improve robustness. The third study benchmarks five major uncertainty-quantification methodologies for turbine gas temperature prediction and evaluates them using coverage and interval-width metrics. The fourth study develops a multi-task scientific machine learning framework for engine health management that jointly predicts turbine gas temperature, delta turbine gas temperature, remaining useful life, and survival-related quantities while reporting uncertainty intervals on real fleet data. Taken together, these works show that scientific machine learning can bridge active intervention and passive monitoring across the engineering lifecycle. Across all four studies, domain knowledge is used not only to interpret model outputs, but to shape the learning objective, the representation, and the notion of reliability itself. The resulting perspective connects optimal control, prognostics, and uncertainty quantification within one coherent framework for engineering decision support.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Computer Science & Applications
- Department dc:contributor.department
- Computer Science and#38; Applications
- Grantor dc:publisher
- Virginia Tech
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Barry-Straume, Jostein Alf
- Chairs dc:contributor.committeechair
-
- Sandu, Adrian
- Son, Chang Min
- Committee members dc:contributor.committeemember
-
- Onufriev, Alexey
- Stilwell, Daniel J.
- Attia, Ahmed Mohamed Mohamed
- Karpatne, Anuj
Subjects
dc:subject × 7Rights
dc:rights- Statement dc:rights
-
- Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Other
- vt_gsexam:46916