Integrating Deep Learning and Contraction Theory for Robust Nonlinear State Estimation via Unsupervised Scientific Machine Learning

What the math gives to ML

The paper provides a constructive way to learn nonlinear observer corrections and contraction metrics jointly by minimizing a matrix partial differential inequality (MPDI) residual at collocation points. The transferable asset is a data-independent exponential convergence certificate: if the learned observer Jacobian is uniformly contracting in a positive-definite metric, state-estimation errors decay exponentially without requiring a known equilibrium or nominal trajectory. The most useful neural-network transfer is a neural state-space observer whose correction network and metric network are trained with a positive-definiteness parameterization and a pointwise contraction penalty, followed by an empirical certification sweep that tests the predicted contraction-rate boundary and noise amplification.