const blogs = [
    {
      id: "llm-scientific-tasks",
      title: "Evaluating LLMs for Scientific Tasks",
      date: "February 15, 2025",
      excerpt: "How effective are large language models at handling specialized scientific concepts?",
      content: `# Evaluating LLMs for Scientific Tasks
  
  Scientists and researchers are increasingly turning to Large Language Models (LLMs) to accelerate their work. But how good are these models at truly understanding scientific concepts?
  
  In our recent work, we developed a framework to evaluate LLMs on tasks specific to material science and chemistry. Here are some key findings:
  
  1. LLMs trained on general text corpora struggle with specialized scientific reasoning
  2. Domain-specific fine-tuning significantly improves performance
  3. There's still a gap between the best models and human experts
  
  ## Methodology
  
  We created a benchmark consisting of:
  - Multiple-choice questions from graduate-level textbooks
  - Reasoning tasks requiring chemical intuition
  - Structure prediction tasks
  - Literature-based synthesis tasks
  
  Read our full paper for detailed results!`
    },
    {
      id: "multimodal-materials",
      title: "The Promise of Multi-Modal Models in Materials Science",
      date: "January 10, 2025",
      excerpt: "Exploring how multi-modal AI can transform materials research and discovery.",
      content: `# The Promise of Multi-Modal Models in Materials Science
  
  Modern materials research generates diverse data types - text, images, spectra, crystal structures, and more. Multi-modal models that can process all these data types simultaneously represent an exciting frontier.
  
  ## Current Limitations
  
  While multi-modal models show promise, our recent work highlights several limitations:
  - Visual chemical reasoning often fails for complex reactions
  - Understanding of spatial relationships in crystal structures is limited
  - Integration of spectroscopic data remains challenging
  
  ## Future Directions
  
  We're exploring several promising approaches:
  - Structure-aware pretraining objectives
  - Specialized tokenization for materials data
  - Physics-informed neural architectures
  
  Stay tuned for our upcoming paper!`
    },
    {
      id: "gnn-materials",
      title: "Geometric Deep Learning for Materials Science",
      date: "December 5, 2024",
      excerpt: "How graph neural networks are revolutionizing computational materials discovery.",
      content: `# Geometric Deep Learning for Materials Science
  
  Graph Neural Networks (GNNs) have emerged as powerful tools for materials science, offering a natural way to represent atomic structures and predict properties.
  
  ## Advantages of GNNs for Materials
  
  - **Natural structural representation**: Atoms as nodes, bonds as edges
  - **Invariance to rotation and translation**: Critical for molecular properties
  - **Hierarchical information processing**: From atomic to global properties
  - **Computational efficiency**: Faster than traditional DFT methods
  
  ## Current Research
  
  Our lab is developing specialized GNN architectures that incorporate physical constraints and uncertainty quantification for high-throughput materials screening.
  
  
  Our current research focuses on incorporating physical constraints and uncertainty quantification into these architectures.`
    }
  ];
  
  export default blogs;