Automating Parameter Estimation of Differential Equation Models for Process Systems via Mechanism-Constrained and Data-Driven Hybrid Modeling
Author(s)
Bradley, William T.
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
A mechanistic model properly formulated can serve to predict complex relationships otherwise explainable only through exhaustive experimentation. However, building these models for dynamic systems can be difficult. A lack of domain knowledge requires the modeler to test multiple possible formulations until the data is explained well by the model. Even if formulated correctly, computational tools to estimate parameters may be inefficient or altogether lacking. In contrast, the recent explosion of data and software tools to exploit data has led many to prefer a data-driven approach to model-building. These data-driven models are easy to formulate and fast to implement yet often lack the interpretability required to make useful predictions. This work proposes and evaluates several novel mergers of data-driven and mechanistic modeling paradigms with the aim of accurate, automated estimation of the parameters of mechanistic differential equations. A rationale for each method is given by showcasing its ability to interpolate data and estimate parameters for scenarios where standard methods are inaccurate, inefficient, or simply intractable. Finally, this work demonstrates how improved tools for automated model fitting can assist in the greater problem of automated model formulation.
Sponsor
Date
2022-05-03
Extent
Resource Type
Text
Resource Subtype
Dissertation