Elements: Crowdsourced Materials Data Engine for Unpublished XRD Results

虽然数据驱动分析已经被誉为基础材料科学（如x射线衍射（XRD）分析）的新范式，但高价值的材料数据集往往没有公开，也没有得到充分利用。该项目设计并开发了CRUX，这是一个众包数据基础设施和服务，用于策划、发现、分享和推荐未发表的XRD数据和分析结果。CRUX通过允许使用最先进的众包、知识收获和机器学习技术共享和探索未发表的数据，促进未充分利用的高质量材料科学数据。CRUX提供了一个众包知识库，允许科学家和公众共享和访问未发表的数据资源。它还提供(a)一个新颖的搜索引擎，支持简单的关键字搜索，可以在没有精确的关键字匹配时提供相关的数据资源，并自我进化以提高搜索质量；(b)一个“数据馈送”服务，允许用户轻松接收和跟踪感兴趣的特定数据资源的更新。开发的基础设施和工具实现了一个开放、协作和可持续的平台，可以促进未发表的XRD数据和发现的交换，解决新的研究问题（例如，使用多相数据对材料成分进行预测分析），并为数据驱动的材料科学激发机器学习管道（例如，深度神经网络）的新设计。CRUX将为包括材料科学家、数据分析师、软件开发人员和公众在内的广泛社区提供和共享材料数据资源，从而促进长期合作研究、软件开发和教育。所开发的CRUX系统能够(1)用三层知识图模型连贯地表示材料、数据、元数据和知识；(2)可扩展的XRD元数据管理和信息提取技术，为数据驱动的材料研究提供高价值的未发表的XRD数据源；(3)自适应、自我完善的搜索和推荐技术，根据用户的请求和反馈推荐相关数据集，并在项目结束后持续使用；(4)交互式和探索性搜索技术，以解释和推荐超出初始查询范围的相关数据集。CRUX将通过发现新的高温铁电体等基础材料研究，利用已建立的人在环知识库和主动机器学习算法进行评估。研究社区将能够通过“一键式”上传共享XRD数据资源（分析结果、机器学习模型、处理数据），搜索高质量的数据资源，并（重新）发现机器学习管道的新资源。CRUX支持多个组件推进数据驱动的材料研究，包括材料知识图谱模型、自动数据集成和探索性查询引擎，支持XRD分析的“Why”和“What-if”分析。开发的解决方案将有利于数据驱动的材料科学。例如，研究人员可以利用未发表的两相数据来预测新材料的组成，通过机器学习工具的参数化来确定溶解度限制，并使用更复杂的技术（如深度神经网络）来完善机器学习模型。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Although data-driven analysis has been heralded as a new paradigm in fundamental material science such as X-ray Diffraction (XRD) analysis, high-value material datasets are often not made public and are underutilized. This project designs and develops CRUX, a crowdsourced data infrastructure and services to curate, discover, share, and recommend unpublished XRD data and analytical results. CRUX promotes underutilized high-quality material science data by allowing the sharing and exploration of unpublished data with state-of-the-art crowdsourcing, knowledge harvesting, and machine learning techniques. CRUX provides a crowdsourced knowledge base to allow scientists and the general public to share and access unpublished data resources. It also provides (a) a novel search engine that supports simple keyword search, can provide relevant data resources when the exact keyword matching does not exist, and self-evolves to improve the search quality, and (b) a "data feed" service to allow users to easily receive and track updates of specific data resources of interest. The developed infrastructure and tools enable an open, collaborative, and sustainable platform that can facilitate exchanging of unpublished XRD data and discoveries, unlock new research problems (e.g., predictive analysis of materials compositions with multi-phase data), and inspire the novel design of machine learning pipelines (e.g., deep neural networks) for data-driven materials science. CRUX will make materials data resources available and shareable for a broad community including materials scientists, data analysts, software developers, and the general public, and thus promote long-term collaborative research, software development, and education. The developed CRUX system enables (1) coherent representation of materials data, metadata, and knowledge in terms of a three-tier knowledge graph model; (2) scalable XRD metadata curation and information extraction techniques to promote high-value unpublished XRD data sources for data-driven materials research; (3) adaptive, self-improving search and recommendation techniques to recommend relevant datasets upon user requests and feedback, with sustainability beyond the time of the project; and (4) interactive and exploratory search techniques to explain and recommend the relevant datasets beyond the scope of initial queries. CRUX will be evaluated with established human-in-the-loop knowledge bases and active machine learning algorithms by cornerstone materials research such as the discovery of new high-temperature ferroelectrics. The research community will be able to share XRD data resources (analytical results, machine learning models, processing data) via "one-click" upload, search for high-quality data resources, and (re)discover new resources for machine learning pipelines. CRUX enables several components to advance data-driven materials research, including a materials knowledge graph model, automatic data integration, and exploratory query engine that support "Why" and "What-if" analysis for XRD analysis. Developed solutions will benefit data-driven material science in general. For example, researchers can make use of unpublished two-phase data to predict new materials compositions, identify solubility limits through parameterization by machine learning tools, and refine machine learning models with more sophisticated techniques such as deep neural networks.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.