Although the deceleration in scientific knowledge production has been well documented at the level of fields, the patterns of knowledge growth within subdisciplines remain poorly explored, primarily due to challenges in classifying papers at lower level. Our study addresses this gap by employing natural language process (NLP) tools to explore the growth of sedimentological knowledge which is a subdiscipline within geology. Utilizing SedBERT, a specialized Bidirectional Encoder Representations from Transformers (BERT) model, we accurately classify sedimentological papers, revealing an exponential growth in sedimentological publications over the past 120 years. Publications have doubled every 10.3 years between 1945 and 1980, and every 14 years from 1980 to 2021.We identify a significant paradigm shift during the 1950 s-1970 s, a period known as the 'Golden Age' of sedimentology, characterized by increased lexical diversity and myopic-referencing citation pattern. A subsequent decline in research diversity, driven by a 'follow-the-crowd' strategy, has led to a stagnation in knowledge expansion in the post-Golden Age. Our study illuminates the dynamic research landscape of sedimentology and offers a framework for analyzing the evolution of sub-disciplinary knowledge.