Modeling techniques for dynamic systems using nonlinear physics and machine learningwith application
Abstract
The essential task of modeling is prediction, i.e., charting out the future states of a dynamic system including important nonlinear, aspects such as oscillations, instabilities, bifurcations, and undesirable behavior that affect mission-critical performance. In m,any engineering systems, high fidelity or "full physics" models are either not available, or even when available, are computationall,y too expensive to evaluate or impractical, especially for tasks such as control, diagnostics, prognostics and automation. Moreover,, it is not possible to model all phenomena that we observe with the physics-based tools we have currently. On the other hand, data,driven models using machine learning (ML) and artificial intelligence (AI) are increasingly being used for predictive models, but th,ey have proved to be poor at generalizations and extrapolations especially in changing environments and large connected architecture,s. Finally, many of the properties that we know exist in real systems such as symmetry and conservation laws are often violated in M,L algorithms as they are simply not aware of the need to enforce such principles. In general, this leads to lack of certifiability,and trust in ML-based algorithms hindering their use in engineered systems.In order to mitigate the above problems, the proposed pro,ject will develop several methods to integrate physics into AI-based methods, in particular for modeling and diagnostics of interdis,ciplinary systems. We call this "physics-informed machine learning". We believe that this is a very promising approach that will,result in higher accuracy and more robustness than can be obtained by using either physics or machine learning alone. We will also,explore new and exciting paradigms for collecting data non-intrusively with ultra-high resolution video with potential applications,for ship systems.While noting that this research effort sits squarely within the basic research portfolio, a particular area of inte,nded application will be Power Electronic Power Distribution Systems (PEPDS), which is a relatively new concept and represents a rev,olutionary change in the way power, energy and control are distributed. A key aspect of PEPDS is "Model is the Specification" which, uses the model itself as the specification for design. This requires a very high-fidelity model with accurate one-to-one correspon,dence between the model and real-time behavior. Such models need to be developed and validated for the many components and their in,terplay. Working closely with other researchers, we will develop advanced hybrid models including efficient low-dimensional ones fo,r these systems. In the presence of numerous uncertainties, these models still need to be sufficiently robust to serve as a sound b,asis for the simulations that analyze and predict ship-wide impact. Models will also aid accurate diagnostics that will not only id,entify and accurately estimate degradation and damage but also effectively differentiate natural and nonlinear perturbations from th,e deviations in behavior caused due to damage or cyber incursions. In addition, these models make many tasks such as uncertainty pr,opagation, optimization and control feasible when in fact they would have been normally considered computationally intractable.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 08, 2022
- Source ID
- N000142212480
Entities
People
- C. Nataraj
Organizations
- Office of Naval Research
- United States Navy
- Villanova University