Mixed-Integer Linear Programming Formulation with Embedded Machine Learning Surrogates for the Design of Chemical...

Area of Science:

• Chemical Engineering
• Process Systems Engineering
• Machine Learning Applications

Background:

• Traditional and modular design methods struggle with rapid deployment of new energy and process technologies.
• Process family design offers an alternative strategy for developing adaptable systems.
• Large-scale optimization problems in process design are often computationally intractable.

Purpose of the Study:

• To develop and demonstrate a Machine Learning (ML) surrogate-based approach for process family design.
• To reduce manufacturing costs and deployment timelines for new process technologies.
• To overcome the limitations of traditional optimization methods for complex process design problems.

Main Methods:

• Formulated process family design as a Generalized Disjunctive Program (GDP).
• Transformed the GDP into a large-scale mixed-integer nonlinear programming (MINLP) problem.
• Developed piecewise linear ML surrogates using the Optimization and Machine Learning Toolkit (OMLT) to approximate nonlinearities.
• Generated an efficient mixed-integer linear programming (MILP) formulation using ML surrogates.

Main Results:

• Successfully applied the ML surrogate approach to design families of carbon capture and water desalination systems.
• Achieved optimal solutions in reasonable computational time for complex design problems.
• Obtained solutions comparable in quality to previously reported methods.

Conclusions:

• ML surrogate modeling provides an efficient method for solving large-scale process family design problems.
• This approach enables faster and more cost-effective deployment of new energy and process technologies.
• The methodology is applicable to diverse process systems with varying operational conditions.