Data-driven optimization framework slashes distillation process optimization time

Distillation is one of the most energyintensive separation technologies, accounting for a significant share of industrial energy consumption. Optimizing distillation systems is essential for reducing costs and emissions, but the highly nonlinear models and mixed discretecontinuous decision variables make this a challenging mixedinteger nonlinear programming problem. Metaheuristic algorithms like genetic algorithms are widely used, but they frequently generate infeasible solutions due to simulation convergence failures, wasting substantial computational resources.

In a study published in ENG. Chem. Eng., researchers at Chongqing University and collaborators propose a multiobjective optimization framework (MODIDC) that systematically identifies and corrects promising infeasible solutions, significantly improving optimization efficiency.

The framework follows a standard NSGAII workflow but adds two key operations. First, a Gaussian process surrogate model is trained to distinguish Paretooptimal solutions from dominated ones based on decision variables. This model is then applied to infeasible solutions to identify those located near the Paretooptimal region in the design space—referred to as “promising infeasible solutions.” Second, for each promising infeasible solution, the decision variable with the largest average distance to Paretooptimal solutions is identified and corrected using uniform random sampling within the range observed in Paretooptimal solutions. This adaptive directed correction generates new candidate solutions with a higher likelihood of feasibility and improved performance.

The framework was validated on two complex distillation systems. The first case is a sidestream doublecolumn extractive distillation process for separating methanoltoluene azeotrope using triethylamine as entrainer. The optimization problem involved eight decision variables with two objectives: total annual cost and CO₂based greenhouse gas emissions. The second case is a fourcolumn extractive distillation process for recovering ethyl acetate and methanol from wastewater using dimethyl sulfoxide as entrainer, involving 16 decision variables with three objectives including total annual cost, CO₂ emissions, and process safety index.

In both cases, MODIDC consistently identified optimal or nearoptimal designs that matched or surpassed previously reported results, while significantly reducing computational effort. The framework achieved optimization time reductions of 35.3 % for the first case and 20.8 % for the second case compared to a conventional genetic algorithm.

The key insight is that many infeasible solutions are not random failures but lie close to optimal designs and can be salvaged with minor corrections. By identifying and exploiting these solutions, MODIDC reduces wasteful evaluations and concentrates computational effort on feasible, highquality solutions. The framework is applicable to other complex chemical processes where simulation convergence failures are common.