Tackling dataset curation challenges towards reliable machine learning: a case study on thermoelectric materials

TL;DR

This paper addresses challenges in curating reliable thermoelectric material datasets for machine learning by identifying errors in existing data, proposing a statistical filtering method, and creating a hybrid dataset to improve data quality.

Contribution

It introduces a novel error-based filtering method and a hybrid dataset creation workflow to enhance the reliability of thermoelectric material data for ML applications.

Findings

The proposed filtering method effectively reduces data inaccuracies.

The hybrid dataset shows improved consistency and quality over existing datasets.

Error detection can be generalized to other material property datasets.

Abstract

Machine Learning (ML) driven discovery of novel and efficient thermoelectric (TE) materials warrants experimental TE datasets of high volume, diversity, and quality. While the largest publicly available dataset, Starrydata2, has a high data volume, it contains inaccurate data due to the inherent limitations of Large Language Model (LLM)-assisted data curation, ambiguous nomenclature and complex formulas of materials in the literature. Another unaddressed issue is the inclusion of multi-source experimental data, with high standard deviations and without synthesis information. Using half-Heusler (hH) materials as an example, this work is aimed at first highlighting these errors and inconsistencies which cannot be filtered with conventional dataset curation workflows. We then propose a statistical round-robin error-based data filtering method to address these issues, a method that can be applied to filter any other material property. Lastly, a hybrid dataset creation workflow, involving data from Starrydata2 and manual extraction, is proposed and the resulting dataset is analyzed and compared against Starrydata2.