John Biffl, PhD Student

Co-Advisor: Dr. Douglas Frey

TITLE: Differential Equation Embedded Neural Networks for Modeling and Simulation of Chromatography and Related Systems

ABSTRACT: 

Chromatography is used in many industries to separate compounds of interest from impurities, but developing chromatographic processes is expensive and time consuming. When the chromatographic system is well understood, a small number of calibration experiments can generate a predictive mechanistic model which can be used to perform rapid process development in-silico, but many types of chromatography have incompletely described binding behavior, eliminating this option. By substituting the binding behavior in a mechanistic model for a small neural network acting as a universal function approximator, a predictive hybrid model can be created regardless of the knowledge of binding behavior. Additionally, the trained neural network can be extracted and used as a basis for symbolic regression to attempt recovery of human readable equations, which can contribute to scientific understanding and regulatory acceptance. 
In this seminar, a hybrid model for chromatography and the results of a proof-of-concept experiment to train it using simulated data are presented. Applications to related systems such as biosensors are also discussed, and future work motivated by expanding the applicability of predictive modeling to all types of chromatography is introduced. New drug development tools are churning out candidate molecules at an ever-increasing pace, and tools to optimize downstream processes must keep up. Fast, low resource process development is increasingly important, and this work to develop a universally applicable predictive model for chromatography would be an important part of a next generation toolbox.

Agenda

• 12:55 pm: Meeting room will open
• 1:00 pm: 45-min presentation will be open to the public with Q&A.
• Followed by a closed session with the committee and PhD Student.