Acyclic modular flowsheet optimization using multiple trust regions and Gaussian process regression

This paper presents an algorithm to optimize process flowsheets using Gaussian processes regression and trust regions. We exploit the modular structure of the flowsheet by training separate Gaussian processes (GPs) for each module based on data generated by a process simulator. These GPs are embedded into an optimization model, whose outcome is used to adapt the position and size of the trust region at each iteration. A complication that arises because of the multiple trust regions is that the optimization problem may become infeasible, in which case a feasibility (restoration) problem is invoked. An inherent advantage of this approach is that it removes the need for simulating the complete flowsheet at any point. We demonstrate these ideas on the case-study of an extractive distillation system in order to minimize its total annualized cost (TAC). The performance shows a robust strategy to address flowsheet optimization problems without recycles.