Adaptive least-squares surrogate modeling for reaction systems

Surrogate modeling has been increasingly used to predict the behavior of a given system as an alternative to complex formulations that often lead to time consuming solutions and convergence issues. Surrogates are addressed herein to replace complex formulations for reactor systems within optimization problems. An adaptive sampling algorithm explores the solution space by iteratively building surrogates. Latin Hypercube Sampling is used for the experimental design (input data), and a first principles reaction formulation calculates the output data. Then, discrete least-squares regression minimizes the deviation between the surrogate response and the function being approximated. An optimization problem based on a reaction system is formulated, in which complex first principles equations are successfully replaced by the surrogates. The results indicate highly accurate predictions and near optimal solutions. Therefore, the surrogates can replace the rigorous model without significant loss in the solution quality and objective function. This methodology can potentially provide several benefits and improvements for real-time applications and for integrated optimization environments, in which the use of complex or rigorous models is not suitable.