Structure-Preserving Neural ODEs via Nonstandard Finite Difference Discretization

What the math gives to ML

The paper turns a positivity constraint on continuous dynamics into an explicit gain-loss parameterization and a nonstandard finite-difference update whose denominator prevents negative states for any step size. This is directly transferable to latent Neural ODEs, diffusion-like state models, epidemic models, biological simulators, and architectures whose hidden variables represent concentrations, counts, probabilities, or intensities. The most promising implementation is to replace an unconstrained vector-field block with two softplus networks for production and destruction rates, then integrate them using the rational update and compare robustness under deliberately coarse solver steps.