Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

L^2M^3OF: A Large Language Multimodal Model for Metal-Organic Frameworks

Jiyu Cui, Fang Wu, Haokai Zhao, Minggao Feng, Xenophon Evangelopoulos, Andrew I. Cooper, Yejin Choi

TL;DR

This work addresses the challenge of MOF design by introducing L2M3OF, the first multimodal large language model tailored for MOFs that fuses a crystal-structure encoder with language understanding. Trained on MOF-SPK, a structure–property–knowledge dataset, L2M3OF outperforms leading closed-source LLMs across property prediction, structure extraction, description generation, and Q&A while using fewer parameters. The approach demonstrates the critical value of grounding LLMs in 3D structural representations and curated literature for porous material discovery, enabling more reliable and efficient materials design. The authors provide a reproducible framework with a public MOF-SPK dataset and detailed methodology to advance AI-assisted MOF discovery.

Abstract

Large language models have demonstrated remarkable reasoning capabilities across diverse natural language tasks. However, comparable breakthroughs in scientific discovery are more limited, because understanding complex physical phenomena demands multifaceted representations far beyond language alone. A compelling example is the design of functional materials such as MOFs-critical for a range of impactful applications like carbon capture and hydrogen storage. Navigating their vast and intricate design space in language-based representations interpretable by LLMs is challenging due to the numerous possible three-dimensional atomic arrangements and strict reticular rules of coordination geometry and topology. Despite promising early results in LLM-assisted discovery for simpler materials systems, MOF design remains heavily reliant on tacit human expertise rarely codified in textual information alone. To overcome this barrier, we introduce L2M3OF, the first multimodal LLM for MOFs. L2M3OF integrates crystal representation learning with language understanding to process structural, textual, and knowledge modalities jointly. L2M3OF employs a pre-trained crystal encoder with a lightweight projection layer to compress structural information into a token space, enabling efficient alignment with language instructions. To facilitate training and evaluation, we curate a structure-property-knowledge database of crystalline materials and benchmark L2M3OF against state-of-the-art closed-source LLMs such as GPT-5, Gemini-2.5-Pro and DeepSeek-R1. Experiments show that L2M3OF outperforms leading text-based closed-source LLMs in property prediction and knowledge generation tasks, despite using far fewer parameters. These results highlight the importance of multimodal approaches for porous material understanding and establish L2M3OF as a foundation for next-generation AI systems in materials discovery.

L^2M^3OF: A Large Language Multimodal Model for Metal-Organic Frameworks

TL;DR

This work addresses the challenge of MOF design by introducing L2M3OF, the first multimodal large language model tailored for MOFs that fuses a crystal-structure encoder with language understanding. Trained on MOF-SPK, a structure–property–knowledge dataset, L2M3OF outperforms leading closed-source LLMs across property prediction, structure extraction, description generation, and Q&A while using fewer parameters. The approach demonstrates the critical value of grounding LLMs in 3D structural representations and curated literature for porous material discovery, enabling more reliable and efficient materials design. The authors provide a reproducible framework with a public MOF-SPK dataset and detailed methodology to advance AI-assisted MOF discovery.

Abstract

Large language models have demonstrated remarkable reasoning capabilities across diverse natural language tasks. However, comparable breakthroughs in scientific discovery are more limited, because understanding complex physical phenomena demands multifaceted representations far beyond language alone. A compelling example is the design of functional materials such as MOFs-critical for a range of impactful applications like carbon capture and hydrogen storage. Navigating their vast and intricate design space in language-based representations interpretable by LLMs is challenging due to the numerous possible three-dimensional atomic arrangements and strict reticular rules of coordination geometry and topology. Despite promising early results in LLM-assisted discovery for simpler materials systems, MOF design remains heavily reliant on tacit human expertise rarely codified in textual information alone. To overcome this barrier, we introduce L2M3OF, the first multimodal LLM for MOFs. L2M3OF integrates crystal representation learning with language understanding to process structural, textual, and knowledge modalities jointly. L2M3OF employs a pre-trained crystal encoder with a lightweight projection layer to compress structural information into a token space, enabling efficient alignment with language instructions. To facilitate training and evaluation, we curate a structure-property-knowledge database of crystalline materials and benchmark L2M3OF against state-of-the-art closed-source LLMs such as GPT-5, Gemini-2.5-Pro and DeepSeek-R1. Experiments show that L2M3OF outperforms leading text-based closed-source LLMs in property prediction and knowledge generation tasks, despite using far fewer parameters. These results highlight the importance of multimodal approaches for porous material understanding and establish L2M3OF as a foundation for next-generation AI systems in materials discovery.
Paper Structure (21 sections, 5 equations, 11 figures, 3 tables)

This paper contains 21 sections, 5 equations, 11 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Dataset and model architecture
  4. MOF-SPK Dataset Construction
  5. Model architecture
  6. Training Objective
  7. Experiments
  8. Experimental results
  9. Property prediction.
  10. Structure extraction.
  11. Description generation.
  12. Question & Answering.
  13. Ablation study
  14. Conclusions
  15. Reproducibility Statement
  16. ...and 6 more sections

Figures (11)

  • Figure 1: An overview of the L2M3OF framework and its applicability in MOFs design.
  • Figure 2: An overview of model architecture and model training methods. (A) The architecture differences between L2M2OF and L2M3OF. (B) The schematic diagram of the group training method.
  • Figure 3: Performance comparison of Gemini-2.5-pro and L2M3OF on the tasks of description generation and a case study of application recommendation task.
  • Figure 4: Performance comparison of Gemini-2.5-pro and L2M3OF in the tasks of question&answer.
  • Figure 5: Data analysis of structure–property–knowledge database for crystal materials. (A) Elemental distribution in the dataset. (B) The token quantity density distribution of the CIFs in the dataset. (C) The distribution of properties of crystal material in the dataset. (D) The application distribution of the crystal material in the dataset.
  • ...and 6 more figures