ECMS 2026 - Grimstad, Norway - June 23-26, 2026. A Sensitivity-Driven Sampling Reduction Method for Probabilistic Approximations of ODEs

These slides were presented at ECMS 2026 in Grimstad, Norway (June 23-26, 2026). The work addresses the high computational cost of generating training data for Dynamic Bayesian Networks (DBNs) used as probabilistic approximations of ODE models. We rely on global sensitivity rankings to identify key direct and indirect dynamical dependencies, allowing us to define reduced sampling supports without exhaustive simulations.

Traditional equation-based sampling often focuses only on direct dependencies, potentially missing critical system interactions. We therefore propose a sensitivity-driven framework that incorporates selected indirect influences to maintain the model's fidelity while drastically reducing the required simulation data.

On benchmark models of progressively higher dimensionality, the sensitivity-driven strategy maintains the DBN's structural and probabilistic fidelity relative to the full ODE dynamics, offering a highly efficient alternative to full dataset construction.