Core Concepts
Surrogate models and implicit function formulations can significantly improve the convergence reliability of chemical process flowsheet optimization problems compared to the full-space formulation, with a trade-off between solution accuracy and solve time.
Abstract
The paper presents alternative formulations for the optimization of chemical process flowsheets that leverage surrogate models and implicit functions to replace and remove, respectively, the algebraic equations that describe a difficult-to-converge Gibbs reactor unit operation. The authors compare the convergence reliability, solve time, and solution quality of an optimization problem among full-space, ALAMO surrogate, neural network surrogate, and implicit function formulations.
The key findings are:
Both surrogate and implicit function formulations lead to better convergence reliability compared to the full-space formulation, with low sensitivity to process parameters.
The surrogate formulations are faster at the cost of minor solution error, while the implicit function formulation provides exact solutions with similar solve time.
In a parameter sweep on the autothermal reformer flowsheet optimization problem, the full-space formulation solves 33 out of 64 instances, the implicit function formulation solves 52 out of 64 instances, the ALAMO polynomial formulation solves 64 out of 64 instances, and the neural network formulation solves 48 out of 64 instances.
The ALAMO surrogate formulation is the most reliable, while the implicit function formulation may be preferred when solution accuracy is critical or when generating enough simulation data to train an accurate surrogate model is computationally prohibitive.
The authors demonstrate the trade-off between accuracy and solve time that exists in current methods for improving convergence reliability of chemical process flowsheet optimization problems.
Stats
The full-space formulation solves 33 out of 64 instances.
The implicit function formulation solves 52 out of 64 instances.
The ALAMO polynomial formulation solves 64 out of 64 instances.
The neural network formulation solves 48 out of 64 instances.
Quotes
"Both surrogate and implicit function formulations lead to better convergence reliability, with low sensitivity to process parameters."
"The surrogate formulations are faster at the cost of minor solution error, while the implicit formulation provides exact solutions with similar solve time."