Adaptive Stability-Constrained Neural Differential Equations for Controlled Dynamical Systems with Unknown Inputs

What the math gives to ML

The paper provides a concrete contraction-based robustness mechanism for neural differential equations: jointly learn a vector field and a positive-definite, state-input-dependent Riemannian metric, then penalize violations of a sampled differential contraction inequality. The important transferable detail is that the metric derivative contains both state and input contributions, including the term \(\partial M/\partial u\,\dot u\), which is often omitted in neural stability regularizers. This yields an incremental input-to-state bound: trajectory discrepancies decay exponentially when inputs and disturbances match, and otherwise remain bounded by explicit gains. The most useful transfer is a metric-constrained NODE training loss with an online estimate of \(\dot u\), tested by measuring the predicted contraction boundary and rollout separation decay.