Speakers
Description
In this second part of the session on Catalyst.jl, we focus on inverse problems, i.e. estimating parameters in models via fitting to data. Using Catalyst models as examples, we will demonstrate how Julia packages can be composed to generate complete model fitting pipelines.
First, we will show how structural identifiability can be assessed for any model using StructuralIdentifiability.jl. Next, we will demonstrate how Catalyst.jl integrates with PEtab.jl to define and solve inverse problems by fitting models to time-series data. This includes showing how PEtab.jl can be used to create parameter problems with a wide range of features, such as events/callbacks, simulation conditions (where the model is simulated under different control parameters for different measurements), complex observable models linking model output to data, and steady-state initialization. We will then show how to fit models using robust global optimization methods such as multistart parameter estimation, and how to assess practical identifiability using profile likelihood analysis.
Finally, we will cover how to define and perform parameter estimation for scientific machine learning (SciML) models that combine mechanistic and data-driven components using Catalyst.jl and PEtab.jl. As these models are often challenging to train, we will also demonstrate how to leverage state-of-the-art training strategies implemented in PEtab.jl to improve parameter estimation performance.
