A Data-Driven Review of Geoscience Research Evolution Using SciBERT-Based Topic and Network Analysis

Abstract

This review provides a data-driven synthesis of the evolution of geoscience research from 2004 to 2023. A field-level bibliometric overview of 439,680 records indexed in the Web of Science Core Collection (Geosciences, Multidisciplinary) characterises broad publication patterns, while the in-depth structural and semantic analysis focuses on the 200 most-cited papers per year (4000 articles in total), to which keyword co-occurrence network analysis and SciBERT-based BERTopic modelling were applied. This two-tier design situates an analysis of the field's most influential literature within its full publication landscape, enabling systematic examination of both the structural and semantic dynamics of geoscience. The keyword-network analysis indicates that model, climate and water have consistently constituted the structural backbone of geoscience research, whereas the centrality of evolution, initially a dominant hub, declined steadily over the study period, while data- and technology-oriented themes, such as remote sensing, GIS and machine learning, have gained increasing prominence since the 2010s. Complementary semantic patterns identified through the SciBERT-based topic model reveal a parallel progression: early research emphasised geological and geotechnical processes; the 2010s were characterised by expanding attention to climate change, paleoenvironmental reconstruction and ecosystem dynamics; and the late 2010s onward show a growing focus on disaster response, resource management and advanced technological applications. In the most recent period, AI- and data-intensive approaches emerge as a dominant thematic direction. Together, this review identifies three major paradigm transitions in geoscience research—from geology-centred foundations, through climate- and environment-oriented expansion, to AI-driven and data-intensive Earth science. Trend and change-point analysis of annual topic prevalence in a single 2004–2023 topic model dates the statistically significant shift towards AI- and data-intensive research to 2013–2014, whereas the climate- and environment-oriented expansion proceeded gradually, without an abrupt transition. By offering a systematic and reproducible synthesis of long-term disciplinary change, this review complements conventional narrative reviews and provides a reference framework for future cross-domain research in the era of foundation models and scientific large language model applications.