Replication Data for: CODE: A global approach to ODE dynamics learning

The dataset contains trained model results (JLD2 files) from three data-driven ODE learning approaches — arbitrary polynomial chaos expansion (aPCE/chaos), neural networks (Lux), and Gaussian process kernel regression (GP) — applied to two benchmark dynamical systems: Lotka–Volterra (predator–prey): 750 experiment runs + 234 orthonormal comparison runs (984 JLD2 files, ~13 GB) Damped Pendulum: 45 experiment runs including neural network model files (60 JLD2 files, ~88 MB) Each JLD2 file stores one fully trained model together with its predictions, extrapolations, out-of-distribution results, training losses, and the corresponding experiment configuration.

Experiments systematically vary training data size (N_data), observation noise level, and random seed. Reproduction: Clone the companion repository and run the plot scripts to reproduce all 25 paper figures. Repository: github.com/NilsWildt/CODE-Global-ODE-learning