Speaker
Description
Hybrid dynamical models are increasingly used to describe complex biological systems. Neural differential equations, such as PINNs, UDEs and Neural ODEs, combine mechanistic structure with data-driven components to model complex dynamical systems. However, this flexibility often weakens practical identifiability, leading to overfitting, redundant neural components, poorly constrained parameters and limited extrapolation.
We introduce identifiability-aware Neural ODEs (iNODEs), a framework that uses practical identifiability as a guiding principle for hybrid model formulation and neural architecture selection. Neural components are embedded explicitly within the differential equations as parametric multilayer perceptrons, enabling closed-form sensitivities and systematic assessment of parameter uncertainty and redundancy. Data availability and identifiability metrics guide an automated architecture search that selects parsimonious models and retains only neural components supported by the available data.
We evaluate the framework on benchmark nonlinear systems such as Lotka-Volterra and SIR models, including cases with latent parameters and partial observability. iNODEs yield more stable parameter estimates, reduced uncertainty and improved generalization compared with standard approaches. Identifiability-guided design thus provides a robust route toward reliable, interpretable and parsimonious neural differential equation models.
