Model structures and structural identifiability: What? Why? How?

TL;DR

This paper discusses the importance of structural global identifiability in modeling physical systems, explaining its theoretical basis and demonstrating how to test for it to ensure reliable parameter estimation and predictions.

Contribution

It provides an overview of SGI theory, clarifies key distinctions, and demonstrates testing methods through example applications.

Findings

SGI is crucial for reliable parameter estimation.

Testing for SGI can prevent non-unique model predictions.

Examples illustrate practical testing approaches.

Abstract

We may attempt to encapsulate what we know about a physical system by a model structure, $S$. This collection of related models is defined by parametric relationships between system features; say observables (outputs), unobservable variables (states), and applied inputs. Each parameter vector in some parameter space is associated with a completely specified model in $S$. Before choosing a model in $S$ to predict system behaviour, we must estimate its parameters from system observations. Inconveniently, multiple models (associated with distinct parameter estimates) may approximate data equally well. Yet, if these equally valid alternatives produce dissimilar predictions of unobserved quantities, then we cannot confidently make predictions. Thus, our study may not yield any useful result. We may anticipate the non-uniqueness of parameter estimates ahead of data collection by testing $S$ for structural global identifiability (SGI). Here we will provide an overview of the importance of SGI, some essential theory and distinctions, and demonstrate these in testing some examples.