ENHANCING POLYMER INFORMATICS: A STUDY ON THE IMPACT OF FINGERPRINTING METHODS AND MACHINE LEARNING MODELS

Materials science engineering design and selection require accurate polymerproperty prediction. This study will evaluate machine learning model performance prediction accuracy using single-task and multi-task learning methods. We examined eight polymer properties and added polymer genome (PG) fingerprints to Morgan fingerprinting for more accurate predictions. This work aims to increase polymer property prediction accuracy and efficacy. ML models evaluated include Gaussian Process Regression Single-Task Learning (GP-ST), Neural Network Single-Task Learning (NN-ST), Neural Network Multi-Task Learning 1 (NN-MT1), Multi-Task Learning-all and Neural Network 2 (NN-MT2-all), and Multitask Learning-sub with Neural Networks 2 (NNMT2- sub). Our results show that multi-task learning models outperform singletask learning models when predicting multiple attributes simultaneously. The neural network model shows the best performance in all evaluated attributes. Overall, PG fingerprinting consistently yielded lower root mean square error (RMSE) and normalized root mean square error (NRMSE) values than Morgan fingerprinting, indicating a more accurate representation of polymer structure. The work demonstrates the significant potential of multi-task learning and modified fingerprinting techniques for enhancing the prediction of polymer characteristics. Future polymer research may examine it as a crucial area of study. Additionally, multi-task learning is now widely available, and fingerprinting techniques may be modified.

Name thesis:degree_name

M.S., Mechanical Engineering

Level thesis:degree_level

Master of Science

Discipline thesis:degree_discipline

Mechanical Engineering

Grantor

Cornell University

Year dc:date.issued

2023

Author dc:creator

• Yang, Yu

Committee member dc:contributor.committeemember

• Yang, Rong

Language dc:language.iso

en

DOI

https://doi.org/10.7298/3110-q175

Dc Identifier Other

ProQuest Submission ID: 11874
ProQuest Publication ID: 30573707

OAI identifier oai:identifier

oai:ecommons.cornell.edu:1813/114459