Wei Ren

Wei Ren


Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
5 September 2025 in npj Heritage Science
Verified
Journal Article
Open access
PDF file
JXJia Xing
Jia Xing
WRWei Ren
Wei Ren
DLDu Lei
Du Lei
LZLin Zhao
Lin Zhao
XQXue Qin
Xue Qin
BZBinbin Zhou
Binbin Zhou
The analysis and protection of grotto sculptures face growing urgency due to climate-driven deterioration and a shortage of domain experts. Here we present ChronoStyleNet (CSN) the world’s first domain-specific multimodal large model purpose-built for sculptural heritage. CSN is trained on 295 expert-annotated statues from the Feilai Peak Grottoes (covering 49% of the West Lake region and 13% of Zhejiang grotto sculptures) and 2.46 GB of archaeological literature. It achieves precise performance through targeted fine-tuning and structured prompting under limited data conditions. Evaluated on 22 Yuan-dynasty samples, CSN outperformed five mainstream multimodal large language systems within a six-dimensional ontology-aligned framework. This work establishes a scalable benchmark for domain-adaptive AI in cultural heritage, offering replicable methodology for other endangered monuments worldwide. CSN demonstrates that domain-adapted multimodal AI can deliver high-precision interpretation even with scarce data, providing new tools for digital preservation and scholarly research. It highlights AI’s potential to bridge expertise gaps and reshape conservation practices in vulnerable heritage contexts.

12 August 2026 in Research Square
Verified
Preprint
Open access
PDF file
ZYZike Yu
Zike Yu
JXJia Xing
Jia Xing
LZLin Zhao
Lin Zhao
DLDu Lei
Du Lei
LCLinfeng Chen
Linfeng Chen
WRWei Ren
Wei Ren
Specialist shortages constrain the documentation, dating, and interpretation of Buddhist statuary. We present ChronoStyleNet 2.0 (CSN2.0,Chinese Name: Zhijian Zaoxianghttps://grottoai.xyz/), a multimodal system spanning Chinese dynasties. Without additional parameter fine-tuning, CSN 2.0 combines structured prompting, tri-strategy fusion retrieval-augmented generation, and a domain ontology informed by CIDOC CRM. The system uses an expert-curated knowledge base of 3,642 entries and a dataset of 145 statues. We evaluated CSN 2.0 and four general-purpose multimodal models on 18 held-out sculptures across eight archaeological dimensions. CSN 2.0 obtained relatively high scores in stylistic feature extraction, dating, and linguistic consistency, but lower scores in cultural interpretation and evidence-to-conclusion reasoning. These results support its use as an exploratory, traceable aid for specialist analysis, not a substitute for expert judgment. The small test set, single-run outputs, automated scoring, and absence of component ablations limit generalization and causal interpretation.