首页 > AI前沿 > Synthetic Crystal Data Generation for Diffusion Pretraining in Material Property Prediction

Synthetic Crystal Data Generation for Diffusion Pretraining in Material Property Prediction

arXiv机器学习 2026-11-17 11:27 10 阅读 查看原文
用户需要我作为专业的新闻排版工程师,把一段纯文本的新闻正文排版成语义化的HTML片段。首先得明确规则:只输出HTML片段,不用Markdown围栏,不用解释;每个自然段落用

包裹,长文用

/

划分小节;列表用
  • ,引用用
    ,代码用
    ,重要句子用;保留所有信息,不能增删改;不要文档级标签。
    
    先看原文内容,这是一篇关于扩散预训练在材料性质预测中应用的研究文本。首先,原文是一个长段落,需要拆分自然段落吗?不对,原文是一段,但内容有不同部分:先讲扩散预训练的优势和现有依赖问题,然后讲本工作的方法,包括合成数据生成的三个策略,接着是具体做法,然后是对比实验,最后是结果。不过规则说每个自然段落用

    包裹,所以可能需要按语义拆分自然段落?不对,原文本身是一个自然段落,但其实内容可以分几个部分,不过用户要求保留原文所有信息,不得增删改,所以应该按照原文的结构,还是直接用

    包裹整个原文?不对,再仔细看原文,原文是一段,但有逻辑层次:第一句讲扩散预训练的优势,然后讲其依赖的问题,接着讲本工作的思路,然后讲引入的合成数据生成流水线的三个策略,接着讲具体使用的方法,然后讲对比实验,最后讲结果。 不过规则里说长文用

    /

    划分小节,这里的文本是研究性内容,可能需要划分小节。比如,先有一个总起,然后“研究背景与问题”部分,然后“研究方法”部分,然后“实验与结果”部分?不过要注意不能增删改内容,所以小节标题应该是基于原文内容提炼吗?不对,用户说“保留原文所有信息,不得增删改内容”,所以小节标题不能额外添加,因为原文没有这些标题。哦,对,不能增删改,所以不能自己加小节标题,因为没有在原文中出现。那这样的话,原文是一个自然段落,所以用

    包裹整个内容?但里面有需要强调的部分,比如重要句子,还有策略列表。 等一下,原文里提到“combining three complementary strategies: perturb-and-relax, generative sampling via MatterGen, and protostructure enumeration.”,这里三个策略是列表内容,所以需要用

    • 来呈现,对吧?因为规则说列表内容用
      • 。那这部分需要处理成列表,而不是在段落里。那原文的结构是:“We introduce a synthetic data generation pipeline combining three complementary strategies: perturb-and-relax, generative sampling via MatterGen, and protostructure enumeration.” 所以这句话后面跟着的三个策略,应该做成列表。那需要把这句话拆出来,然后列表。 然后,原文里有一些重要句子,比如“Diffusion-based pretraining has emerged as a powerful strategy for learning transferable representations in graph neural networks for materials property prediction without requiring quantum-mechanical labels such as energies, forces, and stresses for every pretraining structure.” 这里“powerful strategy”可以加吗?或者整个关键部分?还有“this reliance constrains both the scale and the structural diversity of available pretraining data, leaving large regions of crystal structure space unexplored.” 可能也是重要句子。 现在,处理步骤:首先,原文的第一部分,从开头到“leaving large regions of crystal structure space unexplored.” 是一个部分,讲扩散预训练的优势和现有问题,这部分作为第一个

        ,里面的关键词句用强调。然后,“In this work, we connect diffusion-based pretraining with synthetic structure generation to reduce this bottleneck.” 作为第二个

        ,引出工作内容。然后,“We introduce a synthetic data generation pipeline combining three complementary strategies: ” 这里后面跟着三个策略,所以这句话作为

        ,然后下面是

        • 列表,每个策略一个
        • 。接下来,“Where structural relaxation is required, we use fast machine-learned interatomic potentials instead of performing a new DFT relaxation for each generated structure.” 作为下一个

          。然后,“We further compare a coordinate-denoising control with continuous-time variance-preserving (VP) and sub-VP diffusion objectives using a randomly initialized Orb-v2 backbone. The source comparison covers seven MatBench property-prediction tasks, and an eighth task, \texttt{mp\_gap}, is included in the protostructure scaling study.” 这里注意“\texttt{mp\_gap}”是代码,所以要用

          包裹,对吧?规则说代码用
          ,所以“mp\_gap”需要处理成
          mp\_gap
          。然后这部分作为

          。最后,“Every synthetic-pretraining condition numerically outperforms random initialization on all seven source-comparison tasks. Scaling the protostructure corpus from 20k to 165k under a fixed 400-epoch schedule numerically matches or exceeds MP-20k pretraining on most tasks; because this protocol increases both data diversity and optimization updates, the result reflects joint data-and-compute scaling.” 这部分作为最后一个

          ,里面的结果部分可以用强调。 现在检查是否保留所有信息,没有增删改。然后,注意代码的处理,“\texttt{mp\_gap}”在原文中是代码格式,所以用

          mp\_gap
          。然后,三个策略作为列表,用
          • 。重要句子用,比如“diffusion-based pretraining has emerged as a powerful strategy”、“this reliance constrains both the scale and the structural diversity of available pretraining data”、“Every synthetic-pretraining condition numerically outperforms random initialization”、“Scaling the protostructure corpus from 20k to 165k under a fixed 400-epoch schedule numerically matches or exceeds MP-20k pretraining”这些关键部分。 现在组合起来: 首先第一个

            ,里面包含优势和问题,重要部分加:

            Diffusion-based pretraining has emerged as a powerful strategy for learning transferable representations in graph neural networks for materials property prediction without requiring quantum-mechanical labels such as energies, forces, and stresses for every pretraining structure. However, the method retains a critical dependency: it requires structures residing near local minima of a potential energy surface, which have traditionally been obtained through costly DFT relaxations of known or physically synthesized compounds. This reliance constrains both the scale and the structural diversity of available pretraining data, leaving large regions of crystal structure space unexplored.

            然后第二个

            :“In this work, we connect diffusion-based pretraining with synthetic structure generation to reduce this bottleneck.” 可以加吗?比如In this work, we connect diffusion-based pretraining with synthetic structure generation to reduce this bottleneck. 因为是本工作的核心思路,属于重要句子。 然后第三个部分:“We introduce a synthetic data generation pipeline combining three complementary strategies: ” 作为

            ,然后

              列表:

              We introduce a synthetic data generation pipeline combining three complementary strategies:

              • perturb-and-relax
              • generative sampling via MatterGen
              • protostructure enumeration
              然后下一个

              :“Where structural relaxation is required, we use fast machine-learned interatomic potentials instead of performing a new DFT relaxation for each generated structure.” 然后下一个

              :“We further compare a coordinate-denoising control with continuous-time variance-preserving (VP) and sub-VP diffusion objectives using a randomly initialized Orb-v2 backbone. The source comparison covers seven MatBench property-prediction tasks, and an eighth task,

              mp\_gap
              , is included in the protostructure scaling study.” 这里注意代码部分的包裹。 然后最后一个

              ,强调结果:“Every synthetic-pretraining condition numerically outperforms random initialization on all seven source-comparison tasks. Scaling the protostructure corpus from 20k to 165k under a fixed 400-epoch schedule numerically matches or exceeds MP-20k pretraining on most tasks; because this protocol increases both data diversity and optimization updates, the result reflects joint data-and-compute scaling.” 现在检查是否有遗漏的信息,原文的所有内容都包含了吗?原文的每一句话都在:从扩散预训练的优势,到依赖问题,到本工作的思路,到合成数据流水线的三个策略,到结构松弛用机器学习势,到对比实验,到结果,都有了。没有增删改,只是排版。 然后,有没有