Large Language Models in Chemistry and Materials Science: A Comprehensive Survey
6 affiliations
- 1Peking University Shenzhen Graduate School
- 2School of Materials Science and Engineering
- 3Peking University
- 4School of Computer Science
- 5Sinochem Digital Intelligence Technology Co., Ltd.
- 6Shanghai Artificial Intelligence Laboratory
Abstract
Large language models (LLMs) are expanding the scope of artificial intelligence in chemistry and materials science by connecting scientific knowledge, heterogeneous data, computational tools, and experimental workflows. This article presents a comprehensive survey of the foundations, representations, methods, applications, and evaluation of LLM-centered scientific systems. We develop a non-exclusive representation-by-mechanism taxonomy linking text, molecular and crystal structures, spectra, images, and numerical properties with model adaptation, cross-representation grounding, external knowledge and tools, and workflow coordination. Within this framework, we synthesize prompting, domain pre-training and fine-tuning, preference alignment and verifier-driven reinforcement learning, multimodal integration, and retrieval- and agent-based approaches. Applications span literature understanding, property prediction, molecular and materials design, synthesis planning, characterization, hypothesis generation, and closed-loop discovery. By linking method choices to input information, task definitions, and validation conditions, the framework clarifies which reported results can be compared and what scientific claims they support. We organize representative datasets and benchmarks and introduce an evaluation matrix connecting task performance with information fidelity, uncertainty, chemical and physical validity, and workflow reliability. Detailed supplementary cases document the settings and evidence underlying key comparisons. Finally, we identify research priorities in integrated scientific representations, efficient adaptation, reusable knowledge and tool interfaces, and verifiable human-machine collaboration. By linking methodological developments to scientific objectives, this survey provides a conceptual map and practical guidance for developing and assessing LLM-enabled chemical and materials discovery.
Related
- GVIM 2.0: A Multi-Agent Research System for Chemistry and Materials Science
- Systematic evaluation of LLM-based zero-shot model construction in chemistry and materials science: predictive performance and response reliability
- RECOB: Reliable Benchmarking of Experimental Optimization in Chemistry and Materials Science
- A Survey of AI for Materials Science: Foundation Models, LLM Agents, Datasets, and Tools
- Integrating Generative AI and Data Science into General Chemistry Laboratories
BibTeX
@misc{jinshi2026large,
title = {{Large Language Models in Chemistry and Materials Science: A Comprehensive Survey}},
author = {JINSHI LI and CHENGCHUN LIU and CHENGRUI WEI and YIMI WANG and BOXUAN ZHAO and ZHIYUAN YAN and WENDI CAI and WENCHAO WU and SHUANGYAN TIAN and ZIFU WANG and WEIDA WANG and SHUFEI ZHANG and JIE ZHU and FANYANG MO},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009184/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009184/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Large Language Models in Chemistry and Materials Science: A Comprehensive Survey
AU - JINSHI LI
AU - CHENGCHUN LIU
AU - CHENGRUI WEI
AU - YIMI WANG
AU - BOXUAN ZHAO
AU - ZHIYUAN YAN
AU - WENDI CAI
AU - WENCHAO WU
AU - SHUANGYAN TIAN
AU - ZIFU WANG
AU - WEIDA WANG
AU - SHUFEI ZHANG
AU - JIE ZHU
AU - FANYANG MO
PY - 2026
DA - 2026/09/21
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009184/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009184/v1
AB - Large language models (LLMs) are expanding the scope of artificial intelligence in chemistry and materials science by connecting scientific knowledge, heterogeneous data, computational tools, and experimental workflows. This article presents a comprehensive survey of the foundations, representations, methods, applications, and evaluation of LLM-centered scientific systems. We develop a non-exclusive representation-by-mechanism taxonomy linking text, molecular and crystal structures, spectra, images, and numerical properties with model adaptation, cross-representation grounding, external knowledge and tools, and workflow coordination. Within this framework, we synthesize prompting, domain pre-training and fine-tuning, preference alignment and verifier-driven reinforcement learning, multimodal integration, and retrieval- and agent-based approaches. Applications span literature understanding, property prediction, molecular and materials design, synthesis planning, characterization, hypothesis generation, and closed-loop discovery. By linking method choices to input information, task definitions, and validation conditions, the framework clarifies which reported results can be compared and what scientific claims they support. We organize representative datasets and benchmarks and introduce an evaluation matrix connecting task performance with information fidelity, uncertainty, chemical and physical validity, and workflow reliability. Detailed supplementary cases document the settings and evidence underlying key comparisons. Finally, we identify research priorities in integrated scientific representations, efficient adaptation, reusable knowledge and tool interfaces, and verifiable human-machine collaboration. By linking methodological developments to scientific objectives, this survey provides a conceptual map and practical guidance for developing and assessing LLM-enabled chemical and materials discovery.
ER -