EAGER: Collaborative Research: MATDAT18 Type-I: Development of a machine learning framework to optimize ReaxFF force field parameters

非技术摘要。该奖项支持材料研究人员与MATDAT18数据马拉松活动中点燃的数据科学家的持续合作。智能手机、电池和太阳能电池等技术设备的最新进步是新材料发现和应用的结果。原子系统的计算机模拟在预测和发现新材料方面可能具有深刻的见解。基于量子力学的模拟在计算上是昂贵的，除了几个原子的系统之外，对所有系统来说都是令人望而却步的。分子动力学利用原子间相互作用的模型，可以进行涉及大量原子的模拟。ReaxFF就是这样一个相互作用模型，它也可以描述化学键。目前，超过1000个学术团体和公司正在使用ReaxFF来模拟原子系统。完全指定ReaxFF模型需要许多参数。这些参数控制着原子之间的相互作用，必须针对不同类型的材料单独优化。由于可能的参数组合数量非常多，这种优化过程既耗时又复杂，因此限制了ReaxFF的适用性。一个可以在合理时间内产生最佳参数集的程序将通过加速对原子尺度上的潜在物理和化学的研究来促进新材料的研究。机器学习的最新发展在解决这种高维全局优化问题方面很有希望。本研究的目标是开发一种程序，该程序将使用机器学习模型实现快速和高质量的力场开发，并使所有当前和未来的ReaxFF用户都可以访问该程序。本项目的研究成果也可应用于其他大规模多目标优化问题，对涉及大数据、复杂数据的许多科学学科产生影响。开发的机器学习代码和优化程序将通过宾夕法尼亚州立大学材料计算中心和GitHub与研究人员共享。将开展一些外展项目，以教育下一代材料科学家、数据科学家和统计学家。研究小组将在他们的实验室中根据种族、性别和国籍创造多样化的环境。该研究还将提供一个极好的机会，从代表性不足的群体中招募学生，参与材料科学、数据科学和统计学之间的接口项目，并与社会需求高度相关。该奖项支持在MATDAT18数据马拉松活动中点燃的材料研究人员和数据科学家之间的持续合作。ReaxFF是一种常用的反应力场方法，能够模拟大型原子体系中的键形成和解离。为了通过ReaxFF模拟准确揭示这些系统背后的物理特性，必须针对每种不同的材料系统进行力场参数优化，并在优化过程中深入探索高维力场参数景观。然而，由于现有参数的大量存在，限制了力场发展的优化阶段，传统的优化方法非常耗时。这个挑战可以通过开发一个有效的优化框架来解决。在本项目中，将开发一个高效的顺序优化框架，包括“最小能量”顺序搜索和用于高效高斯过程建模的新型“分而治之”策略。这项研究将使ReaxFF力场的开发更加实用，这将使物理和化学快速进入广泛的材料系统，以加强新材料的设计。该项目是如何使用严格的统计/机器学习方法来解决材料科学和工程中的重要问题的一个例子。该项目可能具有变革性，因为它可以通过使用数据科学和机器学习中的新技术来增强对材料系统的原子尺度理解。开发的迭代优化程序将在Python编程语言下组合，以方便商业化分子动力学软件包的实现。从统计学的角度来看，采用分而治之和基于设计的子样本聚集的方法来降低高斯过程建模的计算复杂度是一种创新。它可以在统计学/数据科学的大数据设置中开辟一条新的道路，并可以导致机器学习和优化的进步。为高维问题构建的序贯优化框架为研究大量复杂输入结构问题开辟了新的途径，为统计学和机器学习领域的理论和应用研究注入了新的活力。该奖项由数学和物理科学理事会的材料研究部和数学科学部共同资助。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。

NONTECHNICAL SUMMARIES.This award supports continued collaboration of materials researchers with data scientists kindled at the MATDAT18 Datathon event. Recent advancements in technological devices, such as smart phones, batteries, and solar cells, are consequences of the discovery and application of novel materials. Computer simulations of systems of atoms could be insightful in predicting and discovering new materials. Simulations based on quantum mechanics are computationally expensive and prohibitive for all but for systems of a few atoms. Simulations involving a much larger number of atoms can be done using molecular dynamics which utilizes models for the interactions between atoms. ReaxFF is one such interaction model which can also describe chemical bonding. Currently, more than a thousand academic groups and companies are using ReaxFF to model systems of atoms. It takes many parameters to fully specify a ReaxFF model. These parameters control the interactions between atoms and must be individually optimized for different types of materials. Due to the prohibitively large number of possible combinations of parameters, this optimization process is time consuming and complex, and consequently limits the applicability of ReaxFF. A procedure that can produce optimum parameter sets within a reasonable time will facilitate novel material research by accelerating the investigation of underlying physics and chemistry on the scale of atoms. Recent developments in machine learning are promising in terms of solving such high dimensional global optimization problems. The goal of this study is to develop a procedure that will enable fast and high-quality force field development using machine learning models and make this procedure accessible to all current and future ReaxFF users.The results of this project can also be applied to other large-scale multi-objective optimization problems and can have impacts on many scientific disciplines that involve large and complex data. The developed machine learning code and optimization procedure will be shared with researchers through the Materials Computation Center at Penn State University and GitHub. Some outreach programs will be conducted for educating the next generation of materials scientists, data scientists and statisticians. The research teams will create diverse environments in their laboratories in terms of race, gender and national origin. The research will also provide an excellent opportunity to recruit students from underrepresented groups to participate in projects at the interface between materials science, data science, and statistics and is highly relevant to societal needs.TECHNICAL SUMMARYThis award supports continued collaboration between a materials researcher and a data scientist kindled at the MATDAT18 Datathon event. ReaxFF is a commonly used reactive force field method, capable of simulating bond formation and dissociation in large atomistic systems. In order to reveal the physics behind these systems accurately by using the ReaxFF simulations, the force field parameters must be optimized for each different materials system, and the high-dimensional force field parameter landscape should be explored thoroughly during optimization. However, the large number of existing parameters limit the optimization stage of the force field development, as the conventional optimization approaches become time-consuming. This challenge can be resolved by the development of an efficient optimization framework. In this project, an efficient sequential optimization framework will be developed, including a "minimum energy" sequential search and a novel "divide-and-conquer" strategy for efficient Gaussian process modeling. This study will make ReaxFF force field development more practical, which will enable fast access to physics and chemistry in a wide range of material systems to enhance novel material design. This project can serve is an example of how rigorous statistical/machine learning methods can be used to tackle important problems in materials science and engineering. The project may be transformative, as it can empower the atomistic-scale understanding of materials systems by using novel techniques in data science and machine learning. The developed iterative optimization procedure will be combined under Python programming language to facilitate implementation to commercial molecular dynamics packages. From a statistical point of view, the idea of divide-and-conquer and design-based subsample aggregation to reduce computational complexity of Gaussian process modeling is innovative. It can open a new path in statistics/data science with big data settings and can lead to advances in machine learning and optimization. The sequential optimization framework constructed for high-dimensional problems may open new avenues for studying problems with massive and complex input structure and energize both theoretical and applied research in statistics and machine learning.The award is jointly funded through the Division of Materials Research and the Division of Mathematical Sciences in the Mathematical and Physical Sciences Directorate.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.