Low-Resource Topic Modeling for Domain Experts: A Pilot Study in CO₂ Mineralization Literature

Bibliometric topic modeling is essential for scientific literature synthesis, yet its adoption by domain-expert teams is often hindered by technical barriers. This pilot study overcomes these constraints by evaluating four topic-modeling pipelines (MalletLDA, ProdLDA, BERTopic+SciBERT, SciBERT-NVCTM) through a controlled benchmark on 286 CO₂ mineralization papers (2013–2025). Unlike conventional studies, our framework is implemented on consumer hardware by researchers without prior NLP training, integrating AI-assisted coding to bridge the gap between statistical complexity and domain-expert usability. We introduce a four-level ablation study to calibrate zero-shot LLM performance in innovation detection. Our results demonstrate that while specialized neural models offer superior semantic coherence, simpler pipelines remain highly effective for rapid exploration. We provide a decision-making framework that aligns pipeline selection with research goals, empowering non-specialists to perform reproducible, robust literature syntheses.