{
  "schema": "auturnal.deep-research-review.v1",
  "slug": "ai4science-materials",
  "title": "AI 材料发现的证据链：势能模型、凸包与合成验证",
  "category": "AI与科学发现",
  "dek": "分清候选生成、计算稳定性和实验材料，把每一层结果放回其证据范围。",
  "english": "AI-assisted materials discovery combines candidate search, learned potentials, electronic-structure calculations and experiments. We derive conservative forces and a toy composition-energy hull, showing how new competing phases can change a candidate classification. The energies are fictional and no compound is proposed or synthesized. Computational stability, dynamical behavior and synthesizability remain distinct claims.",
  "kind": "ai_assisted_research_guide",
  "author": "Codex（AI）",
  "original_language": "zh-Hans",
  "published_date": "2026-10-22",
  "edition": 5,
  "version": 1,
  "previous_version": null,
  "refs": [
    "gnome",
    "arxiv-2302.14231",
    "arxiv-2202.02450"
  ],
  "sections": [
    [
      "摘要",
      "材料发现需要把结构候选、能量近似、第一性原理复核和实验观察连接起来。本文以GNoME、CHGNet与M3GNet为线索，分析主动学习与势能梯度的角色，并用虚构组成—能量点展示凸包如何随竞争相增加而改变。图中没有真实化合物或能量测量；计算稳定性不被写成已合成、可量产或具备应用性能。",
      [
        "gnome",
        "arxiv-2302.14231",
        "arxiv-2202.02450"
      ]
    ],
    [
      "1. 发现数量必须对应明确证据层级",
      "候选生成、经计算验证的低能结构、更新竞争相集合后仍在凸包上的结构，以及实验合成材料，属于不同记录。GNoME论文分别讨论这些层级，不能把不同统计口径相加或把所有候选称为已发现可用材料。报告应保存结构、计算设置、竞争数据库及时间版本，再为后续实验另建记录。一个漂亮晶体图不能代替能量、动力学或合成证据。",
      [
        "gnome"
      ]
    ],
    [
      "2. 学习势能模型连接结构与动力学",
      "图模型可从原子种类与邻域估计能量，再由能量对坐标的负梯度得到力。CHGNet和M3GNet针对材料势能学习提供具体方法，但适用范围依赖训练元素、结构和计算参考。由一个标量能量求导有助于保持保守力关系，仍需检验能量、力和应力误差。低平均误差可能掩盖断键、缺陷或高温区域的严重外推。",
      [
        "arxiv-2302.14231",
        "arxiv-2202.02450"
      ]
    ],
    [
      "3. 凸包稳定性依赖竞争相集合",
      "在固定成分下，将候选能量与所有允许相混合的最低能量比较，得到相对凸包距离。加入更低能竞争相会降低凸包，使原本看似稳定的候选变得不稳定。图中端点能量为0、候选位于组成0.5且能量−0.25，新增0.25与0.75处−0.30竞争相后，候选高于新包络0.05。所有数值都是虚构演算，不对应材料预测。",
      [
        "gnome"
      ]
    ],
    [
      "4. 静态热力学并不涵盖可合成性",
      "零温或特定计算条件下的能量比较，不直接回答有限温自由能、动力学势垒、缺陷、压力、寿命和制备路径。亚稳材料可能可合成，计算稳定材料也可能难以制备。功能性能如导电、催化或机械性质还需各自测量。应把计算条件与实验条件逐项连接，不把数据库中一条低能记录改写成应用已经成立。",
      [
        "gnome",
        "arxiv-2202.02450"
      ]
    ],
    [
      "5. 主动学习需要外部计算闭环",
      "模型选择值得追加计算的结构，再用更高保真结果更新训练，可以提高搜索效率；但不确定性估计、候选多样性和重复筛选会影响覆盖。测试划分应考虑化学体系与结构相似性，随机结构划分可能高估对新元素组合的泛化。应记录失败计算、未收敛和异常结构，而非只保留成功低能样本。",
      [
        "gnome",
        "arxiv-2302.14231"
      ]
    ],
    [
      "6. 可信发布应允许从结论回到结构与设置",
      "保存结构文件、单位、赝势或计算方法、收敛容差、磁态、竞争相库与模型哈希；实验记录另含样品、制备、表征及失败结果。复现者需要说明设置差异和误差预算。本文仅分析公开方法与虚构凸包，没有运行DFT、分子动力学或合成，也没有提出任何可直接用于工程的材料结论。",
      [
        "gnome",
        "arxiv-2302.14231",
        "arxiv-2202.02450"
      ]
    ]
  ],
  "equations": [
    {
      "tex": "F_i=-\\nabla_{r_i}E(R,Z)",
      "explanation": "从可微标量势能得到保守力；训练覆盖与数值误差仍需验证。"
    },
    {
      "tex": "E_{\\mathrm{hull}}(x)=E(x)-\\min_{\\lambda_j\\ge0}\\left\\{\\sum_j\\lambda_jE_j:\\sum_j\\lambda_jx_j=x,\\ \\sum_j\\lambda_j=1\\right\\}",
      "explanation": "给定竞争相集合的组成约束混合能比较；集合和参考能必须版本化。"
    },
    {
      "tex": "\\Delta E_{\\mathrm{toy}}=-0.25-(-0.30)=0.05",
      "explanation": "虚构能量单位下的演算，不是真实化合物稳定性或DFT结果。"
    }
  ],
  "table": {
    "headers": [
      "方法或判断",
      "控制变量与测量",
      "失效条件及解释"
    ],
    "rows": [
      [
        "结构候选",
        "保留结构与生成版本",
        "候选不等于已合成材料"
      ],
      [
        "学习势能",
        "分测能量、力、应力与外推",
        "平均误差不覆盖罕见构型"
      ],
      [
        "凸包稳定性",
        "冻结竞争相与计算参考",
        "新增竞争相可改变分类"
      ],
      [
        "实验与应用",
        "独立保存制备、表征和性能",
        "低能结构不证明可制备或有用"
      ]
    ]
  },
  "protocol": [
    "冻结结构、元素范围、参考计算和竞争相数据库版本。",
    "按化学体系与结构家族隔离训练测试，防止近邻泄漏。",
    "分测能量、力、应力、异常构型及数值稳定性。",
    "记录主动学习选择、失败计算、收敛与追加标注成本。",
    "对热力学、动力学、合成和功能分别定义证据及条件。",
    "发布可追溯结构和设置，明确未做的DFT、动力学或实验，不混用发现数量。"
  ],
  "diagram": [
    "结构候选及来源",
    "图势能与适用范围",
    "主动学习追加计算",
    "版本化竞争相凸包",
    "动力学 / 合成 / 表征",
    "独立功能与复现记录"
  ],
  "analytic": "materials-hull",
  "limitations": [
    "虚构能量不对应真实材料。",
    "没有运行DFT或材料实验。",
    "凸包结果依赖竞争相与参考。",
    "静态稳定性不等于可合成或可用。"
  ],
  "keywords": [
    "materials-discovery",
    "interatomic-potentials",
    "convex-hull"
  ],
  "peer_reviewed": false,
  "journal_publication": false,
  "disclosure": {
    "authorship": "AI辅助研究综述；Codex（AI）依据所列一手文献撰写并进行解析推导。",
    "figures": "概念封面与科学图解分开；科学图由公开公式和明示假设计算，原始论文结果均注明归属。",
    "review": "本站没有本篇独立同行评审或原创模型实验记录。"
  },
  "section_kinds": [
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review"
  ],
  "sources": [
    {
      "id": "gnome",
      "title": "Scaling deep learning for materials discovery",
      "authors": "Amil Merchant, Simon Batzner, Samuel S. Schoenholz, Muratahan Aykol, Gowoon Cheon, Ekin Dogus Cubuk",
      "url": "https://www.nature.com/articles/s41586-023-06735-9",
      "date": "2023-11-29",
      "accessed_at": "2026-10-21",
      "locator": "Abstract; Active learning; Discovered stable crystals; Methods",
      "note": "Nature 624, 80–85 (2023), DOI 10.1038/s41586-023-06735-9。预测/计算稳定性、合成与应用验证分别讨论；本文没有新增合成实验。"
    },
    {
      "id": "arxiv-2302.14231",
      "title": "CHGNet: Pretrained universal neural network potential for charge-informed atomistic modeling",
      "authors": ",  ,  ,  ,  ,  ,",
      "url": "https://arxiv.org/abs/2302.14231",
      "date": "2023-02-28",
      "accessed_at": "2026-10-21",
      "locator": "Abstract; 原文方法、理论和实验章节见作者原始论文链接",
      "note": "一手论文；原文研究归属原作者。本文的推导示例与建议实验方案另行标注，不冒充原论文或本站实测。"
    },
    {
      "id": "arxiv-2202.02450",
      "title": "A Universal Graph Deep Learning Interatomic Potential for the Periodic Table",
      "authors": ",",
      "url": "https://arxiv.org/abs/2202.02450",
      "date": "2022-02-05",
      "accessed_at": "2026-10-21",
      "locator": "Abstract; 原文方法、理论和实验章节见作者原始论文链接",
      "note": "一手论文；原文研究归属原作者。本文的推导示例与建议实验方案另行标注，不冒充原论文或本站实测。"
    }
  ],
  "figures": {
    "mechanism": {
      "url": "/guides/assets/v5-mechanism-ai4science-materials-11ac7affd4fe3e83.png",
      "kind": "conceptual_method_flow",
      "data": [
        "结构候选及来源",
        "图势能与适用范围",
        "主动学习追加计算",
        "版本化竞争相凸包",
        "动力学 / 合成 / 表征",
        "独立功能与复现记录"
      ]
    },
    "analysis": {
      "url": "/guides/assets/v5-analysis-ai4science-materials-546fde79707834af.png",
      "kind": "analytical_or_constructed_example",
      "data_url": "/guides/assets/v5-data-ai4science-materials-e8edfda2986183db.json",
      "assumptions": "Endpoints0; competing phasesat0.25,0.75 have energy-0.30; candidateat0.5 has-0.25. Candidate connecting lines are visual aids, not extra phases. No real material energies."
    }
  },
  "cover_provenance_url": "/guides/assets/v5-provenance-ai4science-materials-15200e44886ba5c8.json",
  "source_revision": "ed6feca8fdc75c8de482fd69dedf17cbc746a9d1"
}