Graduate Student Seminar | An informatics-based approach to sustainable manufacturing by the coalition of data, first-principles simulations, and systems modeling, by Anjana Puliyanda

Tuesday, January 3, 2023 4:00 pm - 5:00 pm EST (GMT -05:00)

The Chemical Engineering Department is hosting a special graduate seminar about an informatics-based approach to sustainable manufacturing by the coalition of data, first-principles simulations, and systems modeling. 

Biography : Anjana Puliyanda is presently a Postdoctoral researcher in the Department of Chemical and Biomolecular Engineering at the University of Delaware (USA), working alongside the groups of Dr Marianthi Ierapetritou and Dr Dionisios Vlachos, on a systems-enabled paradigm shift for sustainable manufacturing in a circular economy. Prior to this she earned her PhD in Process Control after a successful defense in January 2022, under the supervision of Dr Vinay Prasad and Dr Zukui Li, from the Department of Chemical and Materials Engineering at the University of Alberta (Canada). Anjana received her Bachelor of Technology in Chemical Engineering from the National Institute of Technology Karnataka (India), in 2017. She has secured competitive scholarships from MITACS and Alberta Innovates in funding support for research in graduate school, and is currently funded by the National Science Foundation.

Abstract

The development of sustainable manufacturing strategies in transitioning from fossil fuel-derived platform chemicals to their biomass-derived analogues is challenged by the (i) feedstock complexity, and (ii) multiple catalytic routes leading to a variety of product distributions, making the choice of catalysts and solvents crucial to minimize downstream separation costs. In this talk, I will present how data, models and simulations form an integrated approach in addressing these challenges in the context of my research background, followed by discussing prospective research plans that have germinated thereof.

Experimental spectroscopic data hold molecular-level information that could demystify feedstock complexity in making diagnostic decisions. The development of an end-to-end inferential machine learning framework comprising regularized spectral decompositions and graph-based causal techniques to identify the underlying species; hypothesize reaction pathways, reaction mechanism dynamics and kinetic parameter estimation purely from the spectroscopic data of complex reacting feedstocks shall be presented. Validation of the hypothesized pathways by domain knowledge, followed by a computational framework to map the resolved spectral signatures to compounds of a database will be shown to enable the mapping to real chemistry.

Aside from experimental data, first-principles molecular dynamics simulations hold information about the impact of solvent configurations on the biomass conversion kinetics but are computationally expensive to be used as a tool to optimize solvent environments. I shall present my findings from training a self-supervised 3D convolutional neural network that extracts latent features from the reactant and product simulation trajectories to assess solvent configuration changes, so that given an unseen reactant trajectory, one can predict if solvent molecules need to be accounted for, in simulating the product profile, thereby saving computational effort. The physical interpretability of the machine learning framework is assessed using saliency maps.

Finally, I shall focus on a systematic approach to formalize a plethora of literature findings on the conversion of biomass-derived sugars to commodity chemicals by developing a model to optimize adaptive in-silico experimentation. The literature-mined data of chemo-catalytic transformations has been enriched with quantitative catalyst and solvent descriptors to train a machine learning surrogate that intelligently guides future experimental efforts by sampling points in the design space that are expected to improve the predicted reaction yields. This is vital in changing the trial-and-error approach based on expert heuristics in the design of experiments.

In view of my research background, I shall present my future plans of investigating the potential of combining machine learning and mechanistic models in driving adaptive experimentation in (bio)manufacturing of therapeutics; the prospect of discovering collective variables via topology-based manifolds of molecular dynamics simulations for solvent screening at the process scale; and the incorporation of feasibility constraints to facilitate the inverse design of optimal catalyst and solvent candidates from the design space navigated by the surrogate models. The ability of experimentally validating the process formulations suggested by the coalition of data, simulations, and models, is projected to have a strong basis in achieving commercialization and technology transfer in the long-term future.