LEAPS-MPS: Deep Learning Phonons using Graph Neural Networks

该项目由LEAPS-MPS奖支持，旨在更好地理解和预测原子如何在不同材料中振动。这些振动很重要，因为它们会影响材料的性能，特别是在高温下。然而，目前的材料数据库包含的关于这些振动的信息很少，这限制了预测给定材料在各种条件下将如何表现的能力。该项目旨在创建一个新的计算框架，使用先进的机器学习技术来模拟这些原子振动。应用这样的计算框架将促进包括热力学性质的大型材料数据库的开发，从而有助于为各种应用设计新材料，例如被称为高熵合金的某些合金和用于管理热量的材料。为了实现这一目标，PI将使用不同建模技术的组合，包括基于量子力学的高级模拟、高吞吐量计算和机器学习。这种多学科方法将使人们能够更深入地了解原子振动及其对材料性能的影响。该项目还将支持波特兰州立大学的计算材料科学本科生和硕士生的教育。它将有助于弥合传统机械材料工程和现代数据驱动工程方法之间的差距。PI将与波特兰州立大学的几个项目合作，与来自代表性不足的群体的学生和K-12水平的学生接触并招募他们参与该项目。技术概述所有材料都由振动原子组成，即使在绝对零度温度下，由于量子效应。在周期性固体中，这些振动被量子化为声子，声子反映在固体中的有限温度特性和能量转移中。然而，目前的材料数据库缺乏声子的动力学信息，限制了材料在环境温度或更高温度下的预测。该项目由LEAPS-MPS奖支持，进行计算研究和教育活动，旨在开发一个分层计算框架，利用图形神经网络来模拟周期表中固体中的原子振动。这将使高熵合金和热存储和管理材料的计算设计得以扩展。该项目的具体目标是：1）使用图形神经网络开发通用机器学习力场进行声子建模，2）通过不确定性主动学习扩展声子数据库，以及3）获得对原子振动的数据驱动理解，以塑造多组分合金的热力学稳定性和热管理材料的传热。这一总体目标将通过多学科方法实现，该方法具有多种建模技术的协同集成，包括先进的非谐晶格动力学模拟，高通量密度泛函理论计算，以及使用图神经网络的机器学习。该奖项反映了NSF的法定使命，并且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。

NON-TECHNICAL SUMMARY This project, supported by a LEAPS-MPS award, aims to better understand and predict how atoms vibrate in different materials. These vibrations are important because they affect the properties that materials exhibit, especially at high temperatures. However, current materials databases contain little information about these vibrations, which limits the ability to predict how a given material will behave under various conditions. This project aims to create a new computational framework that uses advanced machine learning techniques to model these atomic vibrations. Applying such a computational framework will facilitate the development of a large materials database that includes thermodynamic properties and will thus help to design new materials for various applications, such as certain alloys, known as high-entropy alloys, and materials for managing heat. To achieve the goal, the PI will use a combination of different modeling techniques, including advanced simulations based on quantum mechanics, high-throughput calculations, and machine learning. This multi-disciplinary approach will enable a deeper understanding of atomic vibrations and their effects on material properties. The project will also support the education of undergraduate and master's students in computational materials science at Portland State University. It will help bridge the gap between traditional mechanical & materials engineering and modern data-driven engineering approaches. The PI will collaborate with several programs at Portland State University to engage with and recruit students from underrepresented groups and students from K-12 levels to participate in the project.TECHNICAL SUMMARYAll materials consist of vibrating atoms, even at absolute zero temperature, due to quantum effects. In periodic solids, these vibrations are quantized as phonons, which reflect themselves in finite-temperature characteristics and energy transfer in solids. However, current material databases lack dynamical information on phonons, limiting the prediction of materials at ambient or higher temperatures. This project, supported by a LEAPS-MPS award, conducts computational research and education activities with the aim to develop a hierarchical computational framework that exploits graph neural networks to model atomic vibrations in solids across the periodic table. This will enable a broadened computational design of high-entropy alloys and of thermal storage and management materials. The specific objectives of this project are: 1) developing universal machine learning force fields using graph neural networks for phonon modeling, 2) expanding the phonon database through uncertainty-informed active learning, and 3) gaining a data-driven understanding of atomic vibrations in shaping thermodynamic stability in multi-component alloys and heat transfer in thermal management materials. This overarching goal will be achieved by a multi-disciplinary approach featuring a synergistic integration of multiple modeling techniques, including advanced anharmonic lattice dynamics simulation, high-throughput density functional theory calculations, and machine learning using graph 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.