(223a) Intelligent Process Development: Design and Validation of Separation Archetypes with Machine Learning

In the development of pharmaceutical manufacturing processes, quick and efficient design decisions need to be made to accelerate the development timeline and bring drugs to market quicker. While retrosynthetic planning for reaction pathways helps to accelerate some aspects of process development, it fails to address the separation or workup of products to meet purity specifications. This is critical, as a difficult or impossible separation can render a particular synthetic pathway useless or cost prohibitive. Unlike retrosynthetic planning tools, there is automated tooling for separations process design. We aim to create an automatic framework for the rapid evaluation of separations technologies by offering an architecture for expediting design decisions.

Consider the design of a staged equilibrium-based separations process. In this well-established practice, all stage-wise calculations are solved via a series of material balances, energy balances, and equilibrium equations. Together these equations are called MESH equations (i.e., Material balances, Equilibrium relationships, Summation, and Enthalpy balances). These series of algebraic expressions have deterministic solutions though they often require iteration to find the operating condition, say for example the reflux ratio (R_D), that meets the specified capital cost constraint, via the number of stages, N.

To automatically design a distillation process for a binary mixture (i.e., A,B), chemical properties of both components must be established in machine-readable inputs. Material balance specifications such as feed, bottoms, and distillate flow rates and composition, must be solved along with equilibrium relationships and the operating conditions (R_D) to determine the number of stages (N) required for the desired separation. Simulation software, such as ASPEN, is often used to generate the vapor-liquid equilibrium (VLE) relationship and solve the MESH equations subject to the specifications and constraints, to return design parameters. Here, we consider a workflow beginning with the featurization of binary components (A,B) that enables simulation-free distillation design. We focus on the development of a deep learning model that accelerates explication of the capital and operating cost requirements; see (Alam, McGill) for discussion of vapor-liquid equilibrium and predictive machine learning. Specifically, we encode the distillation simulation results into a neural network through the use of a data library for distillation. The machine learning model captures the behavior of the MESH equations to predict the necessary design criteria and compare optimal performance based on cost objective functions. Validation of the model will be shown with an implemented case study of amine/alcohol separations within a continuous pharmaceutical manufacturing process.