Machine learning-enhanced metadata analysis for identifying polymer compositions

Abstract

In materials science, accurately predicting and identifying polymer compositions from limited experimental data could potentially help in advancing material design, sustainability, and faster development. Traditional methods for analysing aged or complex polymers are often time-consuming and costly, requiring extensive measurements. Machine learning along with meta-data analysis and curation of an un-biased dataset offers a promising alternative, enabling quicker and notably accurate identification of materials or material properties.

The aim is to integrate machine learning with metadata analysis to predict polymer-filler combinations from independently measured property values. As an introductory study, a machine learning pipeline using ensemble methods, including KNN, Random Forest and Neural Networks (MLP) with XGBoost as the final estimator combined via stacking, was developed. The model was fine-tuned using Bayesian optimization for efficient hyper-parameter tuning. All individual and combinations of models were evaluated for accuracy and computational efficiency.

Evaluation was done on a dataset of various polymer-filler combinations obtained from NanoMine, the performance was assessed using top-k accuracy metrics and cross-validation score, based on the ability to correctly identify likely compositions based on user input. The model predicts the top three polymer-filler combinations with associated confidence levels.

More information

Identifiers

ORCID for Sunny Chaudhary: orcid.org/0000-0003-2664-7083

ORCID for Orestis Vryonis: orcid.org/0000-0002-2862-4494

ORCID for Paul Lewin: orcid.org/0000-0002-3299-2556

Catalogue record

Export record