‹ back to 2026-07-16

RTS Smoother-Guided Learning of Physics-Based Neural Differential Models

arXiv cs.LG Physics-Informed ML Ahmet Demirkaya, Georgios Stratis, Tales Imbiriba, Zachary D. Danziger, Deniz Erdogmus 2026-07-16

TL;DR - A hybrid neural–physics framework that learns unknown parts of an ODE system from partial, noisy measurements by alternating between RTS-smoother state estimation and neural-network parameter fitting. It matters because it recovers missing dynamics while preserving interpretable mechanistic structure.

Keeps known ODE terms explicit and represents unknown terms with a neural network, retaining mechanistic interpretability.
Uses a two-stage iterative scheme: (1) infer latent states via a Rauch–Tung–Striebel smoother with parameters fixed, (2) fit NN parameters via backpropagation on the smoothed trajectories, repeating until convergence.
Designed for partial state observation, where only a subset of variables are measured and some equations are unknown.
Evaluated on linear, nonlinear, and stiff benchmark systems, reporting improved latent-state reconstruction and long-horizon prediction.

view merged work →