Extracting Material Property Measurement Data from Scientific Articles

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5393–5402 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 5393 Extracting Material Property Measurement Data from Scientific Articles Gihan Panapitiya*, Fred Parks, Jonathan Sepulveda and Emily Saldanha* Pacific Northwest National Laboratory, Richland, WA, 99354, USA {gihan.panapitiya, fred.parks, jonathan.sepulveda, emily.saldanha}@pnnl.gov Abstract Machine learning-based prediction of material properties is often hampered by the lack of sufficiently large training datasets. The ma- jority of such measurement data is embedded in scientific literature and the ability to auto- matically extract these data is essential to sup- port the development of reliable property pre- diction methods. In this work, we describe a methodology for an automatic property ex- traction framework using material solubility as the target property. We create an annotated dataset containing tags for solubility-related entities using a combination of regular expres- sions and manual tagging. We then compare five entity recognition models leveraging both token-level and span-level architectures on the task of classifying solute names, solubility val- ues, and solubility units. Additionally, we ex- plore a novel pretraining approach that lever- ages automated chemical name and quantity extraction tools to generate large datasets that do not rely on intensive manual effort. Finally, we perform an analysis to identify the causes of classification errors. 1 Introduction While the application of machine learning meth- ods for material property prediction holds great promise for material discovery and design across a broad range of applications, such methods are of- ten hampered by a lack of sufficiently large and di- verse training datasets. Typically, relevant measure- ments and information exist only in unstructured formats such as the published scientific literature and are not available in aggregated and standard- ized databases. The ability to automatically extract, process, and analyze large sets of material property data would represent a significant capability for the advancement of material design efforts. The development of predictive models for the solubility of organic molecules is one such use case that would support a wide range of application ar- eas including pharmaceutical, environmental, and energy storage applications. For example, molec- ular solubility is a key performance driver for re- dox flow battery (RFB) technologies which rely on energy-bearing redox active molecules that are dissolved in a liquid electrolyte. The solubility of a molecule determines its maximal concentration in the electrolyte and the resulting energy density of the system. To support the development of pre- dictive models for the discovery and design of new materials for these batteries, a comprehensive solu- bility measurement database is required. The ability to automatically parse the scientific literature for existing and newly published solubil- ity measurements would provide a significant ac- celeration to our ability to enlarge existing datasets, as well as keep predictive models up-to-date with the newest data. However, the extraction of numer- ical data from the materials science and chemistry literature presents several key challenges. Material property measurements are often highly sensitive to the specifics of the experimental conditions. Ad- ditionally, in comparison with tasks …

Content truncated to 500 words