Mestres A.M. Development of surrogate models for distillation trains

Barcelona: Universitat Politechnika de Catalunya, 2021. — 73 p.The computational time required to solve optimization problems in rigorous simulation programs is usually unaffordable, raising the need to use surrogate models. The development of these approximate models is a challenge that needs to handle the computational burden and risk of overfitting. In the present work, tools, and procedures to build, train, and validate an Artificial Neural Network (ANN) are developed to build simplified models of rigorous simulations. The proposed tools are tested with a case study that addresses the synthesis of separation trains for the products of polyethylene pyrolysis, focusing on the distillation columns of the process simulated with Aspen-HYSYS. Finally, two ANN models have been developed to simulate the behaviour of the column regarding a function that considers the costs of the simulation. Both models fit correctly and show good accuracies with respect to the surface studied. The purpose of this work is the development of a method based on artificial neural networks to obtain a surrogate model capable to describe the behaviour of the individual distillation columns so that they can be later used in the solution of the optimization of the separation train recovering the monomers obtained from the polyethylene pyrolysis. These surrogate models will efficiently mimic the results of the rigorous models obtained with Aspen HYSYS. The approach hereby adopted not only includes the creation and training of the model, but also the sampling of the data and its treatments to fit surrogate’s requirements. Deep learning concepts are applied in conjunction with basic distillation knowledge to avoid models with over fitting and under fitting.