Convex Hybrid Modeling bridges ML accuracy with interpretability via operator theory

In a new paper on arXiv, researcher Wentao Tang introduces a convex hybrid modeling approach that aims to reconcile the accuracy of machine learning with the interpretability required for decision-making in process control. Traditional linear models are favored in control because they are structurally simple and easy to optimize, but they often fail to capture complex nonlinear dynamics. Tang leverages operator theory—which can represent any nonlinear system as a nonparametric operator—and combines it with convex optimization to produce surrogate models that remain interpretable.

The framework is built around three progressive settings: (i) regularization around a known reference model, (ii) restriction to an interpretable subspace, and (iii) restriction to a nonlinear interpretable manifold. The most general setting uses operator-theoretic reparameterization to view the system as a kernel-based mixture of interpretable models. Tang demonstrates the method on both static and dynamic examples, showing how the convex formulation yields efficient, physically meaningful models. The full 19-page paper (with 6 figures) is submitted to the 2027 Foundations of Computer Aided Process Operations / Chemical Process Control conference, and the shorter version highlights its potential for real-world industrial applications.