Design Space Approximation with Gaussian Processes

Abstract The design and operation of process systems are required to meet multiple constraints related to production schedules, product quality, safety, economic performance and environmental footprint. These constraints define the set of feasible design and/or operational parameters which is called process Design Space (DS). In most instances, process constraints are defined as functions of state variables of the system in which case the full-scale process model must be solved for their verification which can be a computationally demanding task for large-scale nonlinear models. This a challenge for online applications such as model-predictive control or real-time optimisation. In this study we present a computationally efficient method of evaluating the feasibility of a set of model parameters using a surrogate indicator function of the DS through Gaussian Process approximation of deterministic inequality model constraints. The method allows finding a compromise between the computational effort required and the level of confidence reflecting the accuracy of DS approximation.