2025-10-02 –, Jean-Baptiste Say Amphitheater
Language: English
When using parametric dynamical models with partial observations, one important task is to infer the parameter values and the states from the measurement data. One of the obstacles to a successful estimation can be the lack of structural identifiability. This means that an unambiguous estimation is impossible due to the very structure of the model, regardless of the quality and quantity of data. We will present StructuralIdentifiability.jl, a Julia package capable of detecting identifiability issues, exploring their nature, and suggesting model transformation improving identifiability.
Structural identifiability/observability refers to a property of a parametric dynamical model that determines if the model parameters/states can be inferred uniquely in the absence of noise. Lack of structural identifiability may pose problems when calibrating model from data, so assessing this property is an important step in model development.
StructuralIdentifiability.jl is a Julia package developed since 2021 and now allowing different types of identifiability analysis for ODE models with rational dynamics (and also some functionality for discrete-time models). The main functionality includes analyzing different type of identifiability (such as local and global) and, if the model is not structurally identifiable, finding quantities which are nevertheless identifiable and using them to propose a new model. The package has interfaces to interact with ModelingToolkit and Catalyst (the latter by T. Loman).
The goal of the presentation will be to present the functionality of the package through a series of case studies. We will highlight the way different types of analyses can interact with each other and the efficiency/detailedness tradeoffs to consider when dealing with a practical model. We will also explain how to interpret and exploit the results of indentifiability analysis performed by the software. The presentation will conclude with a survey of upcoming features and development directions for the package. We expect that the talk will be of interest for researchers using Julia in the context of dynamical modeling.
Most of the presented features are based on the first version developed together with R. Dong, C. Goodbrake and H. Harrington and recent works with A. Demin and C. Rackauckas.
- PhD in Mathematics from Moscow State University in 2016;
- Postdoc in Johannes Kepler University (Linz, Austria) in 2016-2017 and in New York University in 2017-2019;
- Assistant Professor in Computer Science at the Higher School of Economics (Moscow) in 2019-2020
- Since 2020: Assistant Professor in Computer Science at École Polytechnique, Institute Polytechnique de Paris
Research interests: symbolic computation for differential and difference equations