Collaborative Research: CDS&E: Systematic Multiscale Modeling using the Knowledgebase of Interatomic Models (KIM)

该奖项支持OPENKIM，该奖项支持研究人员社区使用基于牛顿定律的原子计算机模拟来攻击材料科学，化学，工程和物理问题，从而发现新材料，设计新设备，理解生化过程等等。原子模拟在现实的科学、工程和工业应用中起着关键作用。这些模拟越来越多地使用拟合原子间模型（IMs），描述原子相互作用时作用在原子上的力的数学公式，来预测材料的特性，它们对外部应力的反应方式，并设计创新的纳米结构，原子的微小结构比人的头发小10万倍。在过去，这种原子模拟的潜力受到几个因素的限制：(1)缺乏标准化的应用程序编程接口，使得IMs从一个模拟程序转移到另一个模拟程序变得困难；(2)缺乏用于存储和交换计算机实现的管理电子存储库，使得难以复制已发表的结果；(3)缺乏比较即时消息准确性的工具，使得难以在新应用中自信地使用即时消息。这些限制已经通过原子间模型开放知识库（OpenKIM）的创建得到了解决，这是一个协作的在线材料项目，用于合理化、标准化和表征im。该奖项支持OpenKIM在从电子电路到飞机制造等领域的科学和工程进步方面取得重要进展。它将对竞争对手的即时通讯和模拟方法进行尖锐的评估和比较，使计算研究人员能够快速探索其他已发表的即时通讯，或开发和验证新的即时通讯。它还将促进科学模拟结果的复制。该项目将扩展OpenKIM，以便将计算化学和分子生物学社区吸引到这一材料努力中，促进两个社区之间的交流，这些社区具有共同的目标和兴趣，但迄今为止被语言、单位和计算惯例所分割。该组的学生和博士后有机会与一个由知名科学家和工程师组成的国际跨学科小组合作，研究具有挑战性的科学问题的横截面，例如决定材料性质的缺陷的作用以及电子设备操作中不满足化学键的影响。通过降低进入计算材料科学的门槛，OpenKIM正在促进代表性不足的群体和发展中国家的群体进入这一技术和科学的核心领域。该奖项支持OpenKIM，这是一个协作在线材料项目，用于合理化、标准化和表征用于表示材料模拟中原子间能量和力的原子间模型（IMs）。该项目旨在支持、扩展和利用OpenKIM进行科学研究。首席研究人员将把材料界的智慧和经验与机器学习、数据挖掘和信息几何等先进方法结合起来，从根本上简化和严格材料的原子模拟领域。OpenKIM提供了一个难得的机会来回答基本的科学问题。通过全面开放的访问，pi预计许多研究人员将使用丰富的OpenKIM存储库来解决该领域的科学和方法问题。pi将通过整合新的im、参考数据和测试，通过扩展KIM标准以支持远程静电场、蒙特卡罗和生物分子键合力场，并通过继续提供关于KIM的文档、讲座、讲习班和教程来支持这些活动。为了进一步推进KIM任务，pi将解决对成功的顺序多尺度建模至关重要的两个广泛而有趣的问题：(1)为了在连续尺度上准确地模拟X现象，IM需要重现哪些关键特征？该项目将提供工具来回答这个问题，通过(a)开发各向异性材料特性的功能形式，以概括已知缺陷和界面的行为，这些特性已经被确定为对微观结构演化的连续模拟至关重要，以及(b)使用从信息几何理论中收集的流形学习方法，该方法将微分几何技术应用于概率论领域。寻找经验启发或规则，提供对一类im的更高尺度行为的洞察，以及对现实世界的洞察。(2)对于给定的应用程序X，给定的IM有多可靠？pi将解决不确定性量化的这一组成部分，也称为IM可转移性，通过(a)使用机器学习技术识别与重要连续尺度材料特性密切相关的关键原子间配置，并使用统计方法估计这些配置的IM不确定性；(b)利用IM拟合参数的大不确定性为IM预测中的系统误差提供贝叶斯信息几何估计。

NONTECHNICAL SUMMARYThis award supports OPENKIM which supports the community of researchers using computer simulations of atoms based on Newton's Laws to attack materials science, chemistry, engineering, and physics problems enabling the discovery of new materials, the design of new devices, the understanding of biochemical processes and much more. Atomistic simulations play a key role in realistic scientific, engineering, and industrial applications. These simulations increasingly use fitted interatomic models (IMs), mathematical prescriptions that describe the forces acting on atoms when they interact, to predict the properties of materials, the way they respond to external stresses, and to design innovative nanostructures, tiny structures of atoms some 100,000 times smaller than a human hair. In the past the potential of atomistic simulations of this kind has been limited by several factors: (1) the lack of a standardized application programing interface has made it difficult to transfer IMs from one simulation program to another; (2) the lack of a curated electronic repository for storing and exchanging computer implementations of IMs has made it difficult to reproduce published results; (3) the lack of tools for comparing the accuracy of IMs made it difficult to use IMs with confidence in new applications. These limitations have been addressed by the creation of the Open Knowledgebase of Interatomic Models (OpenKIM), a collaborative online materials project to rationalize, standardize, and characterize IMs. This award supports OpenKIM as it goes forward in important ways that will facilitate scientific and engineering progress in fields from growing electronic circuits to airplane manufacture. It will make sharp evaluations and comparisons between rival IMs and simulation methods, allowing computational researchers to rapidly explore alternative published IMs or develop and validate new ones for their use. It will also facilitate replication of results in scientific simulations. This project will extend OpenKIM in order to draw the computational chemistry and molecular biology communities into this materials endeavor, facilitating communication between two communities with common goals and interests but hitherto divided by language, units, and computational conventions. Students and post-docs in the group have the opportunity of collaborating with an international, interdisciplinary group of well-known scientists and engineers on a cross-section of challenging scientific problems, such as the role of defects in determining properties of materials and the effect of unsatisfied chemical bonds in electronic device operation. By lowering the barriers to entry into computational materials science, OpenKIM is facilitating the entry of underrepresented groups and those from developing nations into this technologically and scientifically central field.TECHNICAL SUMMARYThis award supports OpenKIM, a collaborative online materials project to rationalize, standardize, and characterize interatomic models (IMs) used to represent energies and forces between atoms in materials simulations. This project is aimed to support, extend, and leverage OpenKIM to do science. The Principal Investigators will blend the wisdom and experience of the materials community with advanced methods from machine learning, data mining, and information geometry to radically simplify and make more rigorous the field of atomistic simulations of materials. OpenKIM represents an unusual opportunity to answer fundamental scientific questions. With full and open access, the PIs anticipate many researchers will use the rich OpenKIM Repository to address scientific and methodological questions of the field. The PIs will support these activities by incorporating new IMs, reference data, and tests, by extending the KIM standard to support long-range electrostatic fields, Monte Carlo, and biomolecular bonded force fields, and by continuing to provide documentation, talks, workshops, and tutorials on KIM. To further the KIM mission, the PIs will address two broad and fascinating issues of critical importance to successful sequential multiscale modeling: (1) What key features does an IM need to reproduce in order to accurately model phenomenon X at a continuum scale? The project will provide tools to answer this question, by (a) developing functional forms for anisotropic materials properties to encapsulate the behavior of known defects and interfaces which are properties already identified as vital for continuum simulation of microstructure evolution, and (b) using manifold-learning methods gleaned from information geometry theory, which applies the techniques of differential geometry to the field of probability theory, to find empirical heuristics or rules that provide insight into the higher scale behavior of a class of IMs, and insight on the real world. (2) How reliable will a given IM be for a given application X? The PIs will address this component of uncertainty quantification, also called IM transferability, by (a) using machine-learning techniques to identify key interatomic configurations which strongly correlate with important continuum scale materials properties and using statistical methods to estimate IM uncertainties for these configurations, and (b) using the large uncertainties in IM fitted parameters to provide Bayesian information geometry estimates for the systematic errors in IM predictions.