A Methodology for The Optimal Surrogate Modelling of Digital Twins Using Machine Learning

Process simulation and digital twins are of paramount importance to design new processes or optimize already existing plants and equipment. The main drawback with current process simulation software is the computational time required to obtain a solution convergence towards a new steady-state. Especially when the input or the output to a system are perturbated. This procedure may take up to minutes in large systems or with strong non-linear recycles. Surrogate modelling of digital twins offers the possibility to speed up the time to convergence required by process simulators by substituting the fundamental or rigorous models with machine learning methods and models. In this work, a surrogate modelling methodology is described for extracting a useful amount of data in a domain near the nominal steady-state of the plant for which a digital twin has been created. For each process variable, a plethora of machine learning models are trained and compared. The best-performing models are chosen to predict the behaviour of such process variables. The application of the surrogate modelling framework thus created has been successfully applied to a steady-state simulation, i.e. digital twin, of an acid gas dimethylamine washing process at the Itelyum exhausted oil refinery in Pieve Fissiraga (LO), Italy.