Benchmarking Multimodal Large Language Models on Electronic Structure Analysis and Interpretation

Large language models (LLMs) are increasingly adopted in materials science, enabling automated literature mining, domain-specific scientific reasoning, and autonomous materials design. However, most existing systems are still limited to single-modality inputs, preventing the use of rich multimodal information inherent in the field. The electronic structure of materials is essential for predicting material properties, understanding their origins, and guiding new materials design, yet its integration into multimodal LLM (MLLM) frameworks has been rarely explored. Here, we present the first systematic benchmark of pre-trained MLLMs for density of states (DOS) interpretation. Using a high-fidelity dataset from first-principles calculations, we evaluate MLLMs on visual question answering and captioning tasks related to the interpretation of electronic structures, with captions scored by both human experts and MLLM-based evaluators. Our results reveal the capabilities and limitations of MLLMs in electronic structure analysis and provide a foundation for developing next-generation multimodal AI systems for materials design.