Date of Graduation
Bachelor of Science in Chemical Engineering
The optimal operation of chemical processes provides the foundation for optimization problems to determine the most effective way to operate or design a given process. Chemical processes can be represented as nonlinear systems of equations with decision variables, resulting in a problem that can be solved through nonlinear solvers. The downfalls of nonlinear solvers create the need for improved methods of finding globally optimal solutions to the design or operation of a chemical process. The project will seek to evaluate the use of artificial neural networks to approximate nonlinear systems of equations for the purpose of optimizing chemical processes. The super critical carbon dioxide (sCO2) Brayton recompression cycle was selected as a surrogate chemical process. The process was approximated by a neural network with a rectifying linear activation unit (ReLU). The sCO2 power generation cycle involves discrete decisions and is nonlinear, however the mixed integer nonlinear programming problem can be approximated as a mixed integer linear programming (MILP) form because of the ReLU formulation of the neural network. The MILP formulation for the optimization of the ReLU approximation successfully modeled the locally optimal solution of the original nonlinear model, supporting the use of neural network approximations for complex chemical processes as well as the MILP approximation of the mixed integer nonlinear problem.
Artificial neural networks, nonlinear systems of equations
Watts, L. (2021). Machine Learning Representations for Optimization of Process Systems. Chemical Engineering Undergraduate Honors Theses Retrieved from https://scholarworks.uark.edu/cheguht/177