arXiv · 2611.23291
Structure-Preserving Universal Differential Equations for Epidemic Dynamics
Abstract
Universal differential equations (UDEs) combine mechanistic epidemic models with learned neural closures, but generic additive corrections may violate essential epidemiological properties. We introduce a class of structure-preserving epidemic UDEs in which learned closures parameterize intercompartmental fluxes and the force of infection factorizes through infected compartments. This architecture guarantees positivity, population boundedness, and invariance of the disease-free set. We derive a rank-one next-generation matrix and a closed-form basic reproduction number determined by the learned rates at the disease-free equilibrium. Under explicit conditions, we establish results on threshold dynamics, disease-free and endemic equilibria, stability, and uniform persistence. For a neural SEIR model, center manifold analysis characterizes the direction of bifurcation through derivatives of the learned transmission rate. We further establish fundamental identifiability limitations: closures that agree on observed epidemic trajectories can yield different reproduction numbers and bifurcation types. Quantitative bounds relate uncertainty in the reproduction number to closure regularity and the distance of observations from the disease-free equilibrium, while monotonicity provides one-sided bounds. A numerical example involving three closures illustrates these results. Our findings provide a rigorous framework for epidemiologically consistent scientific machine learning and clarify which dynamical properties can be recovered from epidemic data.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
ABdallah Alsammani, Maia Martcheva. 2026-10-29. Structure-Preserving Universal Differential Equations for Epidemic Dynamics. https://searcharxiv.com/abs/2611.23291
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.