Experimental Design Signatures Emerge from Topic Modeling of Graphene Oxide Literature

Abstract

Abstract The literature on graphene oxide (GO) comprises hundreds of thousands of publications spanning synthesis, characterization, and applications in diverse domains. Advancing GO materials requires a structured knowledge map derived from systematic mining of this extensive corpus. However, extracting meaningful insights from this vast, heterogeneous, sparse, and unstructured data remains a challenge. To address this challenge, Natural Language Processing techniques, including topic modeling, community discovery, and parameter extraction, are here employed to establish a framework. Latent Dirichlet Allocation (LDA) was applied to the full text of more than 120,000 articles, enabling the identification of eight macro-domains within the GO literature, and neural topic modeling was subsequently employed to characterize the granularity and heterogeneity of these domains. Parameter extraction was performed using Large Language Models (LLMs) on a subset of 8000 full-text articles, with a proto-ontology incorporated into the extraction prompt. Even with a broad prompt, an F1 score of 94% was achieved against manual validation. The extracted parameters were then clustered through community modeling and compared with the LDA-derived topics. At this intermediate resolution of the literature, we demonstrate the application of a recommendation system for relevant parameters according to the material family and application domain. The applicability of the framework at higher resolutions was further demonstrated through the construction of a data set of CO2 adsorption for graphene- and GO-based materials. This data set was subsequently employed in a downstream predictive case study using an Extremely Randomized Trees Regressor, with different training-test splitting strategies evaluated to assess in-domain and out-of-domain generalization at article-level resolution.