Sonakshi Gupta
Advisor: Prof. Rampi Ramprasad


will defend a doctoral thesis entitled,


From Literature to Data-Driven Design of Polymers and Their Blends

On


Monday, November 23 at 10:00 a.m.
MRDC Room 3515

or 

 Virtually via MS Teams: 

Teams Link

 

Committee
Prof. Rampi Ramprasad - School of Materials Science and Engineering (Advisor)

Prof. Chao Zhang - School of Computational Science and Engineering

Prof. Helen Xu - School of Computational Science and Engineering

Prof. Guoxiang (Emma) Hu - School of Materials Science and Engineering

Prof. Will Gutekunst - School of Chemical and Biochemical Engineering

Abstract
Polymer informatics has emerged as a powerful route to accelerating polymers design, but its reach is limited by the availability of reliable data. For many polymer properties, curated datasets contain only thousands of values, and for polymer blends and other multi-component formulations, whose properties depend jointly on composition and processing, shared datasets are largely absent. Much of the relevant knowledge already exists in the scientific literature, yet it is reported in heterogeneous, unstructured formats that are not readily machine-readable. This dissertation develops language-model-based frameworks that convert the polymer literature into reliable, informatics-ready data and use that data for materials design, for both neat polymers and polymer blends. First, automated extraction and validation pipelines were developed to build curated datasets from full-text articles, comprising more than one million property records for neat polymers and a dataset of polymer blend formulations and their mechanical properties. These datasets were then used to develop property prediction models, comparing fine-tuned language models with established fingerprint-based approaches for both neat polymers and blends. Finally, these capabilities were integrated into a literature-grounded design agent. Given a design objective, the agent identifies relevant formulations that have already been reported and proposes new candidates guided by the literature and the property predictors, directing experimental effort towards promising designs that have not yet been explored. Together, this work show that the scientific literature can become an active part of materials design, allowing decades of reported experimental knowledge to inform the development of new materials in a reliable an