Speaker
Description
Since 1960, MIT Haystack Observatory has supported atmospheric and geospace research through long running observational programs and the Madrigal distributed database, which provides access to incoherent scatter radar, total electron content, Fabry Perot interferometer, and other datasets. These resources have contributed to a large scientific literature, but identifying relationships among datasets, methods, phenomena, and conclusions across decades of publications remains difficult.
We analyzed a corpus of 951 papers published between 1960 and 2026 that used data available through Madrigal. We developed an interactive application that combines large language models, text embeddings, knowledge graphs, and graph based retrieval augmented generation to catalog and visualize this literature. Each paper is represented by a paper level embedding, while structured representations capture the Madrigal datasets used, scientific methods, phenomena studied, and principal conclusions. Full text was extracted with PyMuPDF and divided into smaller passages to support more detailed concept extraction. Recursive clustering organizes the resulting representations at multiple levels, allowing users to move between broad research areas and narrower groups of related studies.
Researchers can use the application to locate studies that use similar datasets or methods, trace how a topic has developed over time, identify representative case studies, and find methods applied to related problems in other research areas. The interface turns a static publication catalog into an explorable semantic map in which papers can be reorganized according to different scientific criteria.
The talk will describe the application and present results from task based evaluations with students, researchers, and domain experts, including how their feedback informed revisions to the clustering criteria and interface. Although Madrigal provides the initial testbed, the open source software and workflow can be adapted to other scientific archives and document collections.