Surrogate model optimization: a comparison case study with pooling problems of CO2 point sources

Abstract

In this work, we present a benchmark study to leverage the implementation of surrogate models (SMs) within mathematical optimization problems for the integration of carbon capture technologies within an industrial complex, focusing on the pooling of CO₂ streams to enhance efficiency and reduce capture costs. The SMs are built using data from rigorous process simulations in Aspen Plus, with each data point generated by solving equation-oriented optimization problems. We evaluate five different SMs approaches: Automated Learning of Algebraic Models for Optimization (ALAMO), Kriging, Radial Basis Functions (RBFs), Polynomials, and Artificial Neural Networks (ANNs). We assess their accuracy, computational efficiency, and optimization performance.

In “one-shot” optimization, ALAMO is the most computationally efficient, while Kriging requires the highest CPU time and may struggle with convergence. To improve solution reliability, we incorporate a trust-region filter (TRF) solution strategy. Within this framework, Kriging and ANN achieve the fastest convergence (two iterations), while ALAMO offers a good balance between efficiency and reliability. The RBF SM presents high accuracy on training data; however, it requires more iterations, increasing computational demand.