(370v) Machine Learning-Based Surrogate Modeling and Optimization: Application to Systematic Process Intensification Using Building Block Superstructure

Process intensification aims drastic improvement in process performance in several process metrics, such as size, energy efficiency, cost, operability, and safety, through utilizing the synergy between multiple process phenomena [1-2]. However, identification of intensified solutions at the conceptual design stage is a challenge as there can be a plethora of possible process configurations. While optimization-based process synthesis approaches provide systematic tools for process design, they request pre-postulated superstructures with fixed connectivity and equipment types [3]. This approach requires all process configurations and the alternatives to be specified beforehand, which, often times, may not lead intensification.

Recently proposed building block superstructure is a novel approach that does not require a priori postulation of process alternatives. Instead, it relies on physicochemical phenomena to automatically generate optimal intensified flowsheets by positioning building blocks on a two-dimensional grid [4-7]. By using either single or multiple building blocks, we can represent many physicochemical phenomena. The resultant superstructure is modeled with a mixed integer nonlinear programing (MINLP) model, in which continuous variables describe the mass and energy flow while binary variables determine the position of the active phenomena within the grid. Larger grid sizes can embed more flowsheet variants and may result in better solutions. However, as these building blocks are represented using a set of algebraic equations describing thermodynamics, material and energy transfer, etc., it contains several non-convex constraints. Accordingly, the resulting MINLP model suffers from combinatorial complexity and non-convexity as we utilize larger grid sizes. Hence, locating a globally optimal solution still remains a challenge.

To overcome this challenge, we propose the use of machine-learning driven optimization for building block superstructure optimization. Specifically, in this work we will explore fitting accurate but tractable surrogate models for representing components of the block superstructure. Surrogate-based optimization has been applied to superstructure optimization problems [8-9], but its application to the PI building block approach has not been fully investigated. Instead of explicitly using a set of complicated equations to represent a building block, we can construct neural network surrogate models to accurately approximate building blocks with reduced complexity. Several strategies to further reduce the complexity of the problem will be explored, such as combining multiple building blocks to reduce the degrees of freedom, or replacing a subset of the building block superstructure with surrogate models while handling other constraints explicitly. The resulting MINLP model of building block superstructure becomes more tractable, allowing us to obtain optimal solutions. A diverse set of case studies will be presented to demonstrate the proposed methodology and its performance towards locating optimized solutions to complex formulations.