Formal Certification of Surrogate Models for Cyber-Physical Systems Verification

In this short paper, we propose a method based on Statistical Model Checking to formally verify the prediction accuracy of surrogate models of Cyber-Physical Systems learned from simulation data. We show how surrogate models, trained with any desired Machine Learning algorithm and certified via our approach, can aid simulation-based formal verification techniques by greatly reducing the overall total number of model simulations needed. Our preliminary experimental evaluation over a Modelica model of a water pumping system shows that the proposed approach is viable in real-world scenarios.