MatMech: A Multimodal Dataset for Exploring Causal Mechanisms in Materials Science Literature

OverviewThe MatMech Dataset is a large-scale collection of 61,200+ materials science papers, each represented by a structured JSON file and associated figure images.Each paper is represented not just by its text, but by a detailed JSON structure that maps the scientific findings to the Materials Science Tetrahedron (Processing --> Structure --> Properties --> Performance).This dataset provides both the full raw collection (compressed) and sample cases for quick exploration.Data CompositionThe dataset contains:matmech.zip: Full dataset (\~19 GB) containing all papers, organized by DOI folders. Each folder includes:• paper.json: parsed paper metadata and text content• images/: figures extracted from the PDFcase/: A curated set of small example cases for quick testing and inspectiondataset_summary.json: Machine-readable dataset metadataREADME.md: Human-readable dataset description and usage instructionsThe directory structure is shown as follow:matshare_dataset/│├── README.md├── matshare_full.zip├── case/│ ├── doi_001/│ ├── doi_002/│ └── ...└── data_summary.jsonIntended Use CasesThis dataset supports a wide range of materials informatics and scientific text mining tasks, including:Materials mechanism extractionStructure–property reasoning and knowledge graph constructionMultimodal learning from scientific imagesLLM evaluation on scientific domain understandingAutomatic captioning and figure–text alignmentCausal chain prediction and reasoningResearch trend analysis based on materials categoriesFor more information, please refer to the README.md