{
  "id": 15345112,
  "title": "Giving AI a “crystal clear” view of the materials world",
  "url": "https://urgent.news/2026/10/18/giving-ai-a-crystal-clear-view-of-the-materials-world",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-18T04:00:00.000Z",
  "source": {
    "name": "MIT News AI",
    "slug": "mit-news-ai",
    "url": "https://news.mit.edu/2026/giving-ai-crystal-clear-view-of-materials-world-1018"
  },
  "original_language": "en",
  "account": "Artificial intelligence could simplify the process of selecting the ideal material for a specific application, akin to ordering food from a menu. However, current language models struggle with this task due to the lack of appropriate data. Most materials are composed of crystalline structures, stored in specialized formats, rather than the text streams on which large language models are trained. Additionally, the existing literature is biased towards a few popular types of crystals, skewing AI outputs. MIT researchers have developed a multimodal AI model that integrates 3D atomic structures with natural language descriptions from experimental research papers and computational databases. This unified description includes chemical composition, structure, properties, processing considerations, and potential applications, enabling the AI to learn more about what makes materials impactful. The researchers demonstrated that their system can identify better material designs than models trained on text streams alone, as it \"sees\" the atomic arrangements and bonding patterns. When asked to identify suitable materials for solid-state lithium-ion battery electrolytes, the model's top results included materials previously recognized as promising candidates. The researchers also observed that their model learns geometric representations of material structures more effectively than text-only models. Ju Li, the lead researcher, emphasized the importance of structure in materials science, stating that understanding atomic structures enables predictions about properties and performance. The model was co-developed by Ju Li, Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, and Elsa Olivetti. To overcome the limitations of text-based models, which rely heavily on text from published sources and lack understanding of visual geometry, the researchers created a vision-language model for crystalline materials. They trained a language model using contrastive learning to generate textual descriptions of crystal structures, aligning geometric structures with text descriptions. Comparing the predictions of two versions of the model, one trained on AI-generated descriptions and the other on text-only datasets, revealed that the AI-generated dataset version learned to align geometric structures with text descriptions more effectively. The researchers aimed to build a broadly applicable AI model for materials science by synthesizing narratives rather than relying solely on existing papers. They found that the model trained on the narrative dataset could predict the function of materials more accurately. The researchers believe their approach can be applied to many crystalline materials and hope it will accelerate materials science research.",
  "summary": "New research can help AI models understand crystalline materials in 3D and reason about their properties.",
  "key_points": [],
  "editors_take": "Integrating 3D atomic structures with natural language allows AI to overcome the biases of text-only datasets, providing a more accurate way to predict material properties and accelerate scientific research.",
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "MIT News Research",
        "title": "Giving AI a “crystal clear” view of the materials world",
        "url": "https://urgent.news/2026/10/18/giving-ai-a-crystal-clear-view-of-the-materials-world-15345743",
        "published": "2026-10-18T04:00:00.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}